Cluely’s CEO confesses: the $7M revenue was a lie

Cluely was last year’s hottest startup, blowing up TikTok and Twitter with its “Cheat Everything” concept. Bryan Kim from a16z invested $15M in a Series A, and all of Silicon Valley was talking about this rage-bait marketing success story.

On March 5th, Cluely’s founder Roy Lee went on X and personally confessed to his lie. The $7M ARR he announced to TechCrunch? It was fake. The real number was $5.2M. He inflated it by $1.8M. In his own words: “I got a call out of nowhere, rambled a bit to a reporter, and didn’t think it would actually become a story.”

This is clearly wrong. He admits it himself. And even after the confession, conflicting accounts between him and TechCrunch about the circumstances of the interview keep chipping away at his credibility.

And now, people are placing Cluely right next to Theranos and FTX.

My history with Cluely

I’ve met Roy in early 2025. The moment I heard that Big Tech had rescinded his job offer for cheating an interview, I tracked down Roy’s personal email and reached out. We met over Zoom while he was in his Columbia dorm room. It was before he’d started fundraising, and I remember him saying I was one of the first VCs he’d ever spoken with.

He introduced me to his co-founder over the laptop camera and walked me through their plans. My impression at the time: a brilliant, hungry college kid overflowing with ambition — and in that brief encounter, unmistakably an outlier. We bonded over both being Korean American, and he offered allocation for me to invest, but unfortunately, the deal fell apart because the whole 'cheating company' thesis wasn't well-received.

Can you really put this on the same line as Theranos and FTX?

Let me give you the conclusion first — not yet. And I don’t think it will be.

Theranos deceived investors and patients simultaneously with technology that didn’t exist. There was no working product. FTX siphoned billions in customer assets into a sister company. The core of both is the same — they lied to investors.

Cluely padded $1.8M on top of real revenue of $5.2M. They inflated the number. And here’s what matters most: telling an inflated number to a reporter and telling it to investors are two very different problems (even though both are wrong). There is exactly one threshold where this gets legally serious — whether he fed the same numbers to a16z and other investors. And that hasn’t been established yet.

I’m not a lawyer, but to my understanding, the anti-fraud provisions of federal securities law apply when a false statement is made “in connection with the purchase or sale of securities.” Inflating your ARR to a reporter, by itself, is unlikely to meet that threshold. But if that TechCrunch article was used as a reference in investor meetings, or if it influenced investment decisions, the story changes entirely.

If he lied to investors too — if he directly provided falsified materials — that’s a career-ending event. That’s the Theranos and FTX ending. But I believe this was a moment of competitive bravado in a press interview, not investor fraud. If he didn’t lie to his investors, this isn’t a crime — it’s an incident that lives squarely in the gray zone of 'fake it till you make it'. You can criticize it, and plenty of people despise this Silicon Valley and consulting-world mental model, but it’s nothing more and nothing less than that.

So why is he getting torn apart like this?

A $1.8M revenue exaggeration, even if it’s not a legal issue, is absolutely worthy of criticism. A company that built its brand on “cheating” interviews and life just cheated the public. But I think the reason Roy is catching this much heat goes deeper than that.

Cluely pushed the 'interview cheating' angle so hard that they built an audience primed to turn on them. Virality breeds fans and haters simultaneously. When the controversy hit, the haters swarmed all at once — 90-second response delays, admin passwords left exposed in public directories — every single thing came under a microscope. The perception: this team spends all its time chasing clout and almost none building the product. With that track record stacked up, the suspicion compounds, and one mistake detonates everything at once.

This is the double-edged sword of media. Use it well and it’s a gift. Slip up and the exposure comes back at you tenfold. If Roy had quietly focused on substance after using virality to raise his funding, these things would have slid by. But because of all the heat and notoriety he'd accumulated, it all came back at once. (Which is why I believe not every founder needs to play the media game. There are certain people who can pull it off, and the rest shouldn’t try.)

Retention, retention, retention — not virality

Virality is a means to an end. But at Cluely, virality became the end itself. They stopped being a team that built product and became a team that sold attention.

I’ve written about this before — the common thread among startups dying on Instagram is that they couldn’t nail retention. And to nail retention, your product has to be ready.

I’m not dismissing viral strategy outright. Creating buzz for fundraising, then pouring time into the product, then generating noise again when needed — that’s a solid cycle. But when you fixate on the buzz alone, it becomes a problem.

At the end of the day, boring as it may be, consistently building the best product and driving retention is what matters most. Stay quiet during development, then create buzz once the product is ready. That’s the right order. Cluely flipped it.

VCs aren’t free from responsibility either

One more thing — this isn’t Roy’s failure alone.

Youth isn’t an excuse, but the role of an investor backing someone who’s brilliant but still inexperienced shouldn’t be limited to cheerleading. Especially with a college-age founder — I’m not saying micromanage, but share other case studies, offer experience-based mentoring. While the excessive virality was spiraling, while the rage-bait was outrunning the product, someone should have stepped in.

Bryan Kim, who led this deal at a16z, said at the time of the investment: “What matters is building the plane while falling off a cliff.” In the hyper-growth VC mindset, maybe that checks out. But when that plane is nosediving, it’s also the investor’s job to grab the controls.

a16z is one of the most experienced VCs in Silicon Valley. If they invested $15M and failed to warn that virality and fame were consuming the founder, that’s not investing — that’s abandonment. And since this whole thing blew up, a16z’s official statement? There isn’t one. That silence is already a message.

Fame Escape Velocity

Between this incident and the quiet suffering I see among semi-famous people, a concept crystallized for me — Fame Escape Velocity.

In physics, escape velocity is the minimum speed needed to break free of a celestial body’s gravity. SpaceX’s Starship needs to hit escape velocity to leave Earth’s atmosphere.

I realized the exact same dynamics apply to fame.

At the center sits what I call the Normal Person core. Comfortably recognized within their company, enjoying work-life balance and everyday life, living a normal existence. Surrounding that core is a gravity field of surveillance, envy, and jealousy. The source of that gravity? People hiding among the normal ones — the ones who, the moment someone pulls ahead, think: “I thought we were on the same level, and now suddenly you’re hot stuff?”

Launch a rocket — get even a tiny bit famous — and the first thing you enter is the Influencer Orbit. This is the danger zone. You’re awkwardly famous enough to be watched, but not enough to monetize. When someone you thought was your equal suddenly blows up, people apply absurdly strict standards. If they can’t tear down that person’s success, they feel like they’re falling behind or failing. So a small mistake gets repackaged as “fraud,” and a $1.8M exaggeration gets placed on the same shelf as Theranos.

The outermost layer is the Celebrity Orbit. MrBeast, Elon Musk — I’d put them here. The scrutiny still exists, but overwhelming fandom acts as a shield, and the revenue alone offsets the gravitational pull. Reach this orbit and fame itself becomes an asset.

The other key element of this concept is fuel. Virality is solid fuel — ignite it and it’s explosive, but it burns out fast and once you light it, you can’t turn it off. When it’s spent, you need a new booster. The moment you depend on virality, you must constantly manufacture the next viral moment, and when the fuel runs out, gravity drags you down. Cluely and Roy have been riding exactly this trajectory.

Product and retention are liquid fuel. Far harder to handle. Collecting user feedback daily, fixing bugs, improving quietly — no big bang, and it’s a pain to manage. But liquid fuel is controllable. You can throttle it, meter it precisely as needed.

Roy launched the Cluely rocket on solid fuel called virality. It climbed fast, but the uncontrollable fuel kept burning away, demanding endless controversy and marketing. Instead of switching to the liquid fuel engine — spending time building the product — he kept chasing the next solid booster. Building a product is far harder than generating buzz. But the harder path is ultimately the right one. A rocket out of fuel cannot beat gravity.

It ain’t over until it’s over

The Roy Lee I see is a young, smart, ambitious founder. I felt it when I first met him, and the sheer execution behind $5.2M in real revenue proves it.

Personally, I view this incident as a mistake born from engaging with media while still lacking experience. He lied — that’s a fact. But if it’s not a legal issue — if it was competitive bravado aimed at the press, not investors — then this is a chapter of learning and reflection, not the end. Honestly, the fact that he voluntarily confessed something he didn’t have to say shows a degree of integrity.

If it’s not financial fraud, his career doesn’t need to end here. Stewart Butterfield ran two game companies into the ground and then built Slack. Silicon Valley isn’t a place that buries people forever for honest failure — it’s a place that watches what you learned from it. He’s proven he’s great at finding solid boosters and burning them hard. Now it’s time to fire up the liquid fuel engine. Whether it’s with Cluely or the next company.

I’ll probably catch heat for this, but I’ll be completely honest: even if I could turn back time, I think I still would have wanted to invest. VC is, by nature, a bet on people with dreams different from everyone else’s. In a world of venture capital where 95% fail, I believe it’s the obsessive, execution-driven, polarizing outliers who win. Of course, if it turns out Roy deceived his investors too, then I was wrong about the person. But if that’s not the case, my desire to bet on what this person builds next hasn't changed.

Thank you for reading, as always.

Ian

TLDR

Apple announced the MacBook Neo at $599 (education: $499). It runs an iPhone chip (A18 Pro) instead of M-series — a deliberate compromise. But the play isn't about specs. It's about turning the Mac into the new gateway to Apple's ecosystem. The $499 education price matches Chromebooks. Apple isn't selling a laptop. It's selling the next 20 years of a customer — and their family.

Apple Just Did Something Weird

Apple just made the Mac cheap. $599 cheap. Education discount: $499.

Here’s the thing — a luxury brand dropping its entry price by 40% is usually a sign of desperation. Hermès releasing a $100 handbag would stop being Hermès the moment it hit shelves. But Apple isn’t a luxury brand anymore. Apple is an ecosystem company. And ecosystem companies don’t sell products. They sell lock-in.

My first reaction: "Has Tim Cook lost his mind on margins?" But the deeper you dig, the clearer it gets — this isn't a margin play. It's a trap.

The $999 Wall Just Fell

For the past decade, the cheapest Mac cost $999. Apple held that line deliberately — premium segment only, no exceptions. The result? Mac market share: 9.4%. Lenovo: 27.2%. Classrooms: dominated by Chromebooks. Hundreds of millions of iPhone users worldwide who wanted a Mac but couldn’t justify the price.

MacBook Neo cuts that wall in half. And the $499 education price puts it squarely in Chromebook territory.

It ships March 11. It runs an A18 Pro chip (that’s an iPhone chip, not an M-series), 8GB RAM, two USB-C ports, no MagSafe. On paper, it’s a compromise machine.

But judging the Neo by its specs is like judging a Disneyland ticket by its paper quality. The point isn’t the ticket. It’s what happens after you walk through the gate.

The Entry Point Just Flipped

The old Apple funnel worked like this: buy an iPhone → get curious about Mac. The iPhone was the gateway drug.

Problem: the iPhone market is saturated. I looked at the Galaxy S26 this week and thought aliens built it. Camera, display, weight, design — it matches or beats the iPhone 17 Pro on nearly every spec. And yet I can’t switch.

One reason. The ecosystem.

From where I sit — and I say this as a recovering excel-slave and ThinkPad loyalist — the Mac opens instantly. Text copied on my iPhone pastes on my MacBook. My iPhone hotspot connects before I finish turning off airplane mode. Universal Control lets my cursor glide from MacBook to Mac Mini like they’re one machine. AirPods follow me between devices. Apple Watch unlocks my Mac.

None of this shows up on a spec sheet. All of it shows up in daily experience. And that experience is what keeps people from leaving.

MacBook Neo flips the funnel from “Buy iPhone → get curious about Mac” to Buy $599 Mac → never have a reason to leave the ecosystem

The entry point reversed. The Mac is now Apple’s new front door.

The 20-Year Customer, for $499

Follow the causal chain on the education play:

A high school student gets a MacBook Neo for $499. Files accumulate in iCloud. They use Mac through college. After graduation, they buy an iPhone — because the ecosystem is already home. AirPods follow. Apple Watch follows. A decade later, they set up Family Sharing. Spouse and kids enter the ecosystem.

Apple isn’t selling a $499 laptop. Apple is selling the next 20 years of a customer’s spending. And their family’s.

AI Makes the Moat Deeper

There’s a reason OpenAI shipped the ChatGPT desktop app on macOS first — most of their engineers use Macs. In Silicon Valley, “a company that issues Windows laptops” is a quiet red flag in offer negotiations.

The 2025 Stack Overflow developer survey put macOS usage at 31% — over 3x Mac’s actual PC market share. GitHub Copilot, Cursor, Claude Code — in AI-native startups, finding a developer who doesn’t use a Mac is the hard part.

This creates a flywheel:

Developers prefer Mac (31%) → apps ship Apple-first, and better → user experience improves → more users choose Apple → developers prefer Mac even more → vibe coding brings non-developers into the loop → app quantity and quality rise simultaneously

AI is spinning this wheel faster. Non-developers are building apps on Apple Silicon via vibe coding. MacBook Neo is the device that puts more people on this flywheel. More users → more data → better experience → more users.

The virtuous loop that every AI model company wants to build, Apple has been running for a decade — with hardware, software, and services integrated under one roof.

“Samsung Can Do This Too”

The strongest version of the counter-argument: Samsung builds world-class hardware. Galaxy-Windows integration has genuinely improved. Microsoft’s Phone Link, Samsung DeX, Quick Share — the gap is narrowing.

Fair point. But here’s the structural difference:

Apple controls chip, OS, hardware, App Store, cloud, and payments. One company. One incentive structure.

Samsung makes the hardware. Google makes the OS. Microsoft makes the desktop OS. Google runs the app store. Microsoft runs the cloud. Four companies. Four sets of incentives.

This isn’t an effort or tech problem. It’s a governance problem. Does Google have an incentive to optimize Android specifically for Samsung? Does Microsoft have an incentive to tailor Windows for Galaxy Book? Each company is optimizing for its own flywheel, not Samsung’s.

Apple can sell a low-cost device without diluting the brand because the brand isn’t the hardware. The brand is the ecosystem. Samsung can match the price. Samsung can beat the specs. Samsung cannot replicate the integration — because the integration requires controlling the entire stack.

So What?

MacBook Neo’s real spec isn’t the A18 Pro or 8GB RAM. It’s this: ecosystem entry ticket, $599.

What makes this play fascinating is what Apple didn’t do.

Most companies scramble to fix their weaknesses. “AI is behind” → panic-invest in AI. “Too expensive” → slash prices and damage the brand.

Apple went the other direction. They acknowledged the AI gap, plugged it with external resources (Google’s Gemini) just enough to neutralize the liability, and then doubled down on what they already do better than anyone: the ecosystem.

They didn’t fix the weakness. They sharpened the strength.

In a world where every software product gets replicated and model performance converges, the durable moats are network effects, ecosystem lock-in, physical integration, and community. These are the moats that matter more in the AI era, not less.¹

And that raises a question worth sitting with — for startups, for VCs, for anyone building something:

In an era where AI can patch your weaknesses, what’s the strength you should be sharpening?

We’ll keep watching.

Ian

¹ No, I don’t mean humanoid robots. Put your hand down. 😄

The Best Marketing Campaign Nobody Planned

Two of the most powerful men in the world — one running the country, one running the company synonymous with AI — managed to turn Dario Amodei into a martyr. A saint. The last honest man in Silicon Valley.

This martyrdom became the most effective marketing campaign in AI history. By Monday, Claude was #1 on the App Store. Then the servers crashed from demand. Anthropic didn't spend a dollar on it (well, besides the Super Bowl ad that started the first wave).

Here’s the thing — he’s not a saint. But to understand why that matters, you first have to understand how the martyrdom was manufactured.

Act 1: The Face Slap from DOW

The timeline is now well-documented, so I’ll keep this brief.

1. Anthropic told the Pentagon “No” — refused to allow Claude for mass surveillance and autonomous weapons targeting

2. Sam Altman publicly backed Anthropic — told employees and press he shared “the same red lines”

3. Trump kicked Anthropic out of the entire federal government — not just DoW, but Education, HHS, NASA, everything. Slapped them with a “supply chain risk” designation previously reserved for Chinese and Russian companies.

4. Hours later, Altman signed the Pentagon deal himself — on those “same red lines” he just publicly supported

5. OpenAI announced $110B in funding at $730B valuation — Amazon $50B, Nvidia $30B, SoftBank $30B

Now, the “supply chain risk” designation is worth pausing on. That’s a tool the US government has historically used against Huawei, Kaspersky, and companies with alleged ties to hostile foreign intelligence services. Trump used it against an American AI company — one whose product was already running on classified military networks — because its CEO said no to a change of contract.

Act 2: The Back Stab from OpenAI

If Trump slapped Amodei in the face, Altman stabbed him in the back. And the back stab did more damage to Altman himself than to Amodei.

Altman publicly told employees he shared Anthropic’s red lines. Then he signed the Pentagon deal. Then he claimed the terms were equivalent. But if the terms were truly identical, why would the government switch? The Pentagon already had Claude on its classified network. CENTCOM validated it in live combat. You don’t replace a battle-tested system with an equivalent for no reason.

Platformer and The Verge confirmed this. OpenAI’s Pentagon contract boils down to three words: “any lawful use.” If it’s legal, the military can do whatever it wants with OpenAI’s technology. The laws Altman cited — EO 12333, FISA, the National Security Act — are the same legal framework that enabled the mass surveillance programs Snowden exposed in 2013. All “lawful.”

Miles Brundage, OpenAI’s former head of policy research, said it on X: “OpenAI caved + framed it as not caving, and screwed Anthropic while framing it as helping them.” (He later softened the “caved” language — but the core point stands.)

This didn’t happen in a vacuum. Altman’s contradiction fits a pattern well-documented elsewhere — from board testimony to contract disputes to licensing controversies. Each incident follows the same structure: support in public, undercut in private.

In a TechCrunch interview, Altman admitted the deal was “definitely rushed” and “the optics don’t look good (spoiler alert: it’s not just optics)”. He then rushed to host an unprompted AMA on X — a sign he felt the narrative slipping.

The net effect: Trump through punitive retaliation and Altman through his own behavioral pattern manufactured Dario Amodei’s martyrdom without Dario having to do anything at all. He just had to stand still and say no.

Act 3: The Martyrdom (a.k.a. The Best Marketing Campaign in AI History)

And the martyrdom worked. Not because of PR — because it created a cascading transfer of legitimacy across every dimension that matters.

The Tesla problem: For OpenAI

This is the structural point most people miss. Tesla’s core customers are affluent metro buyers; when the company leaned into politics that alienated that base, the brand eroded with its actual buyers.

Now look at AI: who are the biggest spenders and heaviest users? Tech companies, startups, developers, and researchers concentrated in major tech hubs. By absorbing the Pentagon’s “any lawful use” mandate, OpenAI introduced a fundamental friction with its core customer and employee base.

Consumer: From Banned to #1

ChatGPT’s traffic was already eroding — web share from 86.7% to 64.5%, mobile from 69.1% to 45.3%. The Super Bowl ad accelerated it. The Pentagon incident finished it.

A “Cancel and Delete ChatGPT!!!” post on Reddit hit 30K upvotes. CancelChatGPT.com launched. As of March 1, Claude hit #1 on the US Apple App Store — also #1 in Germany and Canada.

And then Claude crashed. Monday morning, all consumer apps went down for five hours. Anthropic’s explanation: “unprecedented demand.”

When a government ban and your competition’s betrayal generates so much demand that it literally takes down your servers, you’re not being punished. You’re being marketed.

Talent: Signatures and Solidarity

In the current state of AI war, losing talent is scarier than losing users.

Sutskever, Murati, Schulman — the exodus from OpenAI was already well underway. Sam Altman is the only active co-founder left at the company he co-founded. Then 97 current OpenAI employees — alongside 772 from Google — signed “We Will Not Be Divided” in solidarity with Anthropic.

In 2023, OpenAI employees rallied to bring Altman back. In 2026, they’re rallying against him. A 180-degree reversal in three years.

There’s a saying in the industry: fewer than 100 people are truly elite AI engineers. For them, moving from OpenAI to Anthropic isn’t a risk anymore — it’s pre-IPO equity, mission alignment, and distance from a company that sold out its users to the government. A rational career decision.

Every vector, flowing in one direction. A martyrdom so perfect it almost looks scripted.

Almost.

Act 4: But He’s Not a Martyr

Here’s where I have to be honest — because the narrative is so clean right now that everyone’s forgetting the messy parts. And in my experience, when a narrative is this clean, that’s exactly when you should start pulling threads.

Dario Amodei is not a saint. He just looks like one because the contrast is so stark

Thread 1: The Safety Pledge He Dropped — The Day Before

On February 25th, one day before Anthropic told the Pentagon “No,” Anthropic quietly dropped its Responsible Scaling Policy commitment — the voluntary pledge not to deploy frontier AI models unless external safety reviews confirmed they were safe.

The timing is remarkable. Anthropic co-founder Jared Kaplan justified it with logic that could have come directly from OpenAI: “It doesn’t make sense to stay on the sidelines when competitors are pushing ahead.” The same “competitive pressure” argument that Anthropic was founded to reject.

On Day 1, Amodei dropped his company’s signature safety commitment because the competition was too fierce. On Day 2, he told the Pentagon his conscience wouldn’t allow it. Both things can be true simultaneously — but together, they complicate the narrative considerably.

Thread 2: Claude Was Already in the Kill Chain

The same day Amodei was saying “No” to the Pentagon, Claude was already being used by CENTCOM — through Palantir’s Maven platform — for active military operations, including the Iran strike planning. According to Axios and NBC News, this included kinetic targeting. An Anthropic executive called Palantir to express concern; Palantir reported the call to the Pentagon.

The “principled refusal” wasn’t about keeping AI out of warfare. Claude was already there for other operations including Venezuela. The refusal was about specific expansions — mass surveillance, bulk data collection, autonomous targeting. Credit to Amodei for drawing the line. But the popular narrative — “Anthropic refused to let its AI be used by the military” — is factually wrong.

Thread 3: The Safety Blowback

Dario has publicly claimed a “25% probability of catastrophic AI events.” At Davos, he compared selling AI chips to China to “selling nuclear weapons to North Korea.” Fortune described his long essay on AI’s existential risk as “as much marketing as prophecy.”

Here’s where it gets structural. That fear wasn’t just rhetoric — it was strategy. Anthropic lobbied hard behind Biden’s AI executive order. They backed California’s SB 1047 after it was weakened to favor incumbents. Meta’s Yann LeCun accused them of lobbying to restrict open-source models — which directly serves Anthropic’s competitive interest. David Sacks, the administration’s AI czar, has been calling this “regulatory capture” for months. He has a point.

And here’s the irony: when you spend years convincing the world that AI is nuclear-weapon-grade dangerous, you can’t be surprised when the government says “great, we’ll regulate it like nuclear weapons.” This “Safety Blowback” — apocalyptic safety rhetoric, deployed for fundraising and regulatory advantage, eventually creates the political conditions for the very government overreach that threatens your business. Anthropic built the rhetorical scaffolding. Trump just moved in.

Palmer Luckey and Ben Thompson’s question — “Do you believe in democracy?” — cuts both ways: when a company lobbies for safety regulations that also happen to kneecap competitors, is that principled advocacy or business strategy wearing a lab coat?

TechCrunch called it “The Trap Anthropic Built for Itself.” By branding so heavily on safety, any compromise — no matter how pragmatic — reads as hypocrisy. The RSP rollback, the Palantir situation, Dario’s own admission that autonomous weapons “may prove critical for national defense” — each one widens the gap between brand and reality. The “safety at all costs” positioning is already being quietly retired.

So What?

Two men converged to create a narrative moment: Trump through punitive retaliation, Altman through contradiction. Together, they manufactured the most effective martyrdom and marketing campaign in AI history. Dario Amodei didn’t have to campaign for sainthood. He just had to say “no” while everyone else said “yes.”

But being made a martyr doesn’t make you a saint.

Dario dropped his safety pledge the day before the Pentagon “No.” Claude was already in the kill chain when he drew his red line. His company lobbied for regulations that kneecapped competitors. And the fear-mongering that built Anthropic’s brand also built the political conditions for the very ban.

So the question that actually matters isn’t moral — it’s structural: does this martyrdom convert into durable market position?

Let’s put the damage in perspective. The federal ban costs Anthropic roughly $200M in government contracts. Against $4-5B+ in ARR by most estimates, that’s less than 5% — not small, but worth the brand equity gained. But also, brand equity without retention is just a sugar high. Claude is #1 on the App Store today, but protest downloads don’t equal product loyalty. We will need to keep track of this.

Those 97 OpenAI employee signatures on “We Will Not Be Divided” are meaningful. In an industry where fewer than 100 people are truly elite AI researchers, and many of them are already OpenAI alumni, this moment is structurally significant. The question worth watching: do they actually leave? If, over the next 90 days, even five or six top-tier researchers make the move, the talent flywheel becomes self-reinforcing — and that’s when the martyrdom converts into something structural.

Two villains, one accidental hero, and zero saints. Welcome to AI in 2026.

We’ll keep watching.

Ian

[Next Michigan: Part 1] The End of Work Isn’t Coming. But Your Job Might Be.

Technology Has Never Set Humans Free

Last week, I summoned the 20 greatest VCs in history in 30 minutes.

I built something called AI VC Reports — an investment committee simulation. Imagine the most legendary venture capitalists sitting around one table, tearing apart your pitch deck. I recreated that with AI. Thirty minutes. Done. (If you haven’t tried it yet, check the link in my last issue → [AIVC Reports link])

But the thing that hit me wasn’t “wow, this is amazing.”

It was: I just built this in 30 minutes. The next person will do it in 20. The one after that, 10. And at some point, everyone can build it — so nobody pays for it.

And then I realized: this isn’t just a VC problem. Coders, analysts, writers, consultants — anyone working in a domain where data exists is watching the same movie play out in real time.

These days, everything reminds me of the Industrial Revolution and what happened to Detroit, Michigan. So I’m turning those thoughts into a three-part series. Today is Part 1: The End of Work Isn’t Coming. Part 2 will cover how to survive in this storm. Part 3 will tackle the structural fractures coming for society.

Musk’s Weekend Vegetable Garden

In November 2025, Elon Musk told a US-Saudi investment forum:

“My prediction is that work will be optional. It’ll be like playing sports or a video game. You can go to the store and just buy some vegetables, or you can grow vegetables in your backyard. It’s much harder to grow vegetables in your backyard, but some people still do it because they like growing vegetables.”

Within 10 to 20 years: robots performing surgery, the concept of money fading away, humans working as a hobby. Work becomes optional — like growing your own vegetables when the grocery store is right there.

Sam Altman goes even further. “Universal extreme wealth” — AI-generated abundance shared by all of humanity. He’s even floating the idea of distributing AI tokens to every person on Earth.

Look, I’d love for this vision to come true. Who wouldn’t?

But history has never once worked that way.

Remember When Flights Were Off-Limits?

Ten years ago, being on a plane was the ultimate excuse for not answering emails. “Sorry, I was on a flight.” That one sentence forgave everything. Then in-flight WiFi showed up. At first, everyone thought it was wonderful. “I can get online at 35,000 feet!”

Now? If you’re on a 10-hour flight and not answering emails, people know you’re not actually in a meeting — and they’re wondering why you’re ignoring them. A 10-hour flight is now a 10-hour work session. In-flight WiFi wasn’t liberation. It was an expansion of the leash. Time you couldn’t work became time you should work.

This is the pattern of every technological revolution. New technology doesn’t reduce work. It raises expectations.

When Excel launched, did accountants work less? No. Analyses that nobody would have dreamed of doing by hand became “obviously you should be running this.” Quarterly financial reviews became weekly. Five scenarios became twenty. Better tools, higher expectations.

When email arrived, did communication get more efficient? No. Daily messages went from 5 to 150. In the era of physical letters, a one-week reply was normal. With email, 24 hours without a response gets you a “why are you ghosting me?” With Slack? That 24 hours shrank to 24 minutes.

When smartphones showed up, did we gain more free time? We became 24/7 connected humans. Slack notifications at the beach. “Urgent” emails during dinner. The smartphone didn’t give us freedom — it demolished the walls of the office. Now everywhere is the office.

When technology boosts productivity, the world demands exactly that much more output.

In the 200 years since the Industrial Revolution, human productivity has increased by orders of magnitude. Dozens of times over. And working hours? Still 40 hours a week. In some countries, they’ve actually increased. If productivity went up by a factor of 50, shouldn’t we be working 4 hours a week by now? We’re not. And we won’t be.

Technology has never liberated humans. Not once.

It just makes us do harder work, and more of it.

And AI Is No Different

This isn’t a prediction. It’s already happening.

While I was writing this piece (February 9th), a Harvard Business Review study from UC Berkeley dropped. They tracked 200 tech workers at US companies over eight months after AI tool adoption. Did work decrease?

The exact opposite.

Three things happened. First, the scope of work expanded. PMs started writing code. Researchers started doing engineering tasks. AI gave everyone the feeling of “I could probably do that myself.” Work that used to go to another team or an outside vendor? Now it’s on your plate.

Second, the line between work and not-work vanished. Plugging in prompts over lunch. “Just one more prompt” before logging off. Checking results before bed. Prompting doesn’t feel like work — it feels like chatting. But when you look back, your breaks weren’t breaks.

Third, multitasking exploded. Writing code manually while simultaneously running AI on the side, spinning up multiple agents in the background, pulling out tasks you’d been procrastinating on. “I have a partner now, so I can handle more.”

One engineer in the study put it perfectly:

“I thought AI would boost productivity and I’d save time and work less. But in reality? You don’t work less. You work the same amount, or more.”

It’s the in-flight WiFi story all over again. Better tools, higher expectations. AI writes a report in 10 minutes — your boss says “great, write 5 more by end of day.” AI auto-generates code — “cool, add 10 more features by next week.” AI handles design — “perfect, give me 30 mockups.”

The total volume of work doesn’t shrink. It grows. Musk’s weekend vegetable garden isn’t coming.

But here’s where it gets strange. If work is increasing, why are people losing their jobs? If there’s more work to do, why are we seeing layoff headlines every single day?

Work Grows. Your Seat Disappears.

This is the real point. The total amount of work stays the same or increases. But the same people don’t keep doing it. We’ve seen this play out in history, over and over.

The 1870s: Telegraph Operators.

Telegraph operators were skilled specialists — the software engineers of the 19th century. They memorized Morse code from youth, practiced thousands of hours, and were only recognized as “experts” once they could tap out 40+ words per minute. Tens of thousands across America. Good pay. High social status.

Then the telephone arrived.

Twenty years of telegraph expertise became worthless overnight. No matter how fast you could tap Morse code, it meant nothing in front of a telephone. Instead, an entirely new profession — the telephone switchboard operator — exploded into existence. Mostly young women. People who could adapt to new technology quickly, with none of the “but we’ve always done it this way” inertia.

Work didn’t disappear. The entire communications industry grew explosively. But the people were replaced.

How many 45-year-old telegraph operators successfully transitioned to switchboard operation? Almost none. They were told to “retrain.” In theory, it was possible. In practice, erasing 20 years of muscle memory and starting over? It almost never happened.

The 1910s: The End of the Horse.

When Ford introduced the assembly line, building a car went from 12 hours to 2.5 hours. Productivity exploded. But here’s what’s interesting — as productivity rose, car prices fell. As prices fell, more people could afford cars. As more people bought cars, more cars needed to be built. The total amount of work didn’t decrease. It exploded.

But who did that work?

The carriage makers, harness craftsmen, and coachmen — their 30 years of experience went to zero. The skill of carving a carriage wheel was worthless in an auto factory. Instead, hundreds of thousands of unskilled assembly line workers were hired. The assembly line didn’t require craftsmanship. New people, new skills, new rules.

Work increased. But that work belonged to new people.

2007: The iPhone.

The day Steve Jobs announced the iPhone, the job title “app developer” didn’t exist on planet Earth. There was no concept of an App Store.

Ten years later: roughly 12 million people employed in the global app economy. A job that didn’t even have a name a decade earlier created 12 million positions. Uber drivers, Instagram marketers, YouTube creators — careers unimaginable in 2006 now supported millions of livelihoods.

But who got displaced? Dedicated GPS device companies. Kodak and the camera film industry. Brick-and-mortar travel agencies. The music industry. Professionals with 20 or 30 years of expertise in these fields got pushed out.

Try telling a senior engineer at a GPS company, “12 million app developer jobs were created, so it’s all fine.” That number means absolutely nothing to them.

See the pattern?

Every time technology shifts — total work is maintained or increases. That’s true. But displacement happens. People optimized for the old technology get pushed out and replaced by younger workers suited for the new one. Telegraph operators replaced by switchboard operators. Carriage artisans replaced by factory workers. Kodak engineers replaced by app developers. The displaced experts get swapped out. The people.

And the displaced hear the same thing every single time: “Just retrain.”

In reality, 45-year-old telegraph operators almost never became switchboard operators. 50-year-old carriage craftsmen almost never became Ford factory workers. 55-year-old Kodak engineers almost never became app developers.

A discarded class emerges. Every time. Without exception.

This is exactly what happened in the place this series is named after: Michigan.

The First Michigan

1950. Detroit. Population: 1.85 million. Fifth-largest city in America.

The rules were simple. Graduate high school, get a job at Ford or GM. Work hard, buy a house, send your kids to school, collect a pension. Your father did it. Your grandfather did it. Work hard and you’ll be fine. That was the social contract.

When factories moved overseas and automation began, car production didn’t decline. It increased. Global auto production kept climbing. Toyota, Hyundai, Volkswagen — all building more cars than ever.

Work didn’t disappear. Detroit’s workers were replaced. By younger workers in Mexico, by robots, and eventually by non-union plants in southern states. The exact same pattern as telegraph operators being replaced by switchboard operators. More work. Different people.

Detroit’s population today: 630,000. A 65% decline. Poverty rate: 35%. Child poverty: 50.2%. In 2013, the largest municipal bankruptcy in American history.

Numbers alone can’t capture it.

When a factory closed, the factory wasn’t the only thing that shut down. Think about it. The neighborhood plant closes. Two thousand workers lose their jobs. The diners where those 2,000 people ate lunch every day lose their customers. The diners close. The diner owners can’t pay taxes. City revenue shrinks. School budgets get slashed. Schools deteriorate, so young couples with kids leave. Young people leave, and the only ones left are the ones who can’t leave. No jobs, no mobility, no safety net. Drugs spread. Families disintegrate.

This wasn’t an economic problem. An entire community collapsed like dominoes.

And what did the elites say?

“Learn software.” “Retrain.” “Adapt to new industries.”

To people whose entire foundation of life had been obliterated, that wasn’t a solution. It was an insult. For 40 years, they followed the rules faithfully — and then they were told, “It’s your fault you couldn’t adapt.” We didn’t make the rules. We didn’t change them. And now we’re the problem?

This isn’t just an American story. Britain’s Rust Belt. Korea’s Gumi and Geoje. Japan’s regional cities. The same pattern repeated worldwide. When industry leaves, people leave. When people leave, communities die.

And economic displacement doesn’t stay economic.

It becomes ideology. It becomes politics. It changes history. The displacement of the Industrial Revolution gave birth to communism and world wars. The 30-year accumulation of betrayal in Detroit split America in two. Michigan, Wisconsin, Pennsylvania — all three were ground zero for manufacturing hollowing-out, and all three decided the presidential election.

I’m not trying to make a political argument. I’m stating facts. Economic displacement inevitably flows into social fracture. (I’ll go deeper on this in Part 3.)

“Next Michigan”

The first Michigan targeted blue-collar manufacturing workers.

The second Michigan targets the entire college-educated white-collar workforce.

“This time is different”? That’s what they said every time. “This technology is different — it’ll actually reduce work.” They said it about Excel. About email. About smartphones. It never happened once. The HBR study I cited above already proves it.

By US Bureau of Labor Statistics data, roughly 70 million Americans work in “Management, Professional, and Related Occupations.” That’s 44% of the total workforce. Not a few times larger than Michigan’s auto workers — dozens of times larger.

Same pattern. Total work won’t decrease. It’ll increase — because AI raises expectations. AI writes a report in 10 minutes, you’ll be told to write 5 more. AI generates code, you’ll be told to add 10 more features. More work than ever.

But the people doing that work will be replaced. AI-native generations will enter, and those optimized for the old way will be pushed out. Exactly like telegraph operators. Exactly like carriage craftsmen. Only this time, three things are different.

Scale. Detroit’s displacement affected millions. This time it’s tens of millions. 70 million in the US alone. Globally? A number you don’t want to think about.

Speed. Detroit took 30 years to collapse. Slowly enough that people could tell themselves “it’s still fine.” AI needs just a few years. I built AI VC Reports in 30 minutes. Not 30 years — 30 minutes.

The intensity of the rage. This is the scariest part. The betrayal felt by a high school grad who went straight to the factory is categorically different from someone who invested four years of college, two years of grad school, unpaid internships, and endless job applications. Massive student loans, hundreds of interviews, all while believing “just follow these rules and you’ll be okay.” The more time and money invested, the sharper the rage of “I followed every single rule.” It’s no coincidence that politicians keep trying to cancel student debt.

And this time, the elites telling people to “just adapt” are telling it to themselves.

Everyone's Watching the Wrong Fire

In 2025, 245,953 tech workers were laid off across 783 companies. Since ChatGPT launched in late 2022, the cumulative total exceeds 500,000. Layoffs explicitly citing AI as the reason: 54,694.

2026? In January alone: 35,000. That’s 856 people per day. Amazon alone cut 30,000 — 10% of their workforce. Meta: 1,500. Oracle is reportedly considering an additional 30,000 reductions. It’s still February.

Layoffs are loud. They make the news, blow up on Twitter, get questioned in Congress. The numbers are visible. The names are visible. There’s a target for the anger.

But everyone is looking in the wrong direction.

“AI is killing junior roles.”

That’s the line you hear most in tech right now. The Stanford-ADP paper came out. UK graduate hiring down 67%. US down 43%. The numbers look terrifying. I wrote about this topic myself last September in a piece titled “It’s the Economy, Stupid.”

My conclusion then: this isn’t AI — it’s a recession. When downturns hit, companies freeze new hiring before cutting existing staff. That’s a textbook pattern in labor economics. The weakest link — entry-level and junior hires — is the first door to close. AI didn’t replace them. The door slammed shut before they could walk through it.

I still believe that. In the short term, junior workers are suffering because of the economic cycle and outsized fear.

But let me push one step further.

In the medium to long term, juniors actually have the advantage.

Forrester Research found — unsurprisingly — that Gen Z has the highest AI readiness score (AIQ) at 22% across all generations. Boomers: 6%. Gen Z has zero resistance to learning new tools. No “but we’ve always done it this way.” Someone who’s invested 20 years into the current way of doing things can’t abandon it. But someone who hasn’t invested in anything yet can start fresh on day one.

When new factories were built after Detroit’s decline, they weren’t staffed by 20-year assembly line veterans. They were staffed by new people with new skills.

The real target of the quiet restructuring isn’t juniors. It’s the existing workforce.

And companies know this. According to Forrester, 55% of employers who used AI as a reason for layoffs already regret it. Half will quietly hire people back. But at lower salaries, overseas, or at the junior level. They’re not recalling the displaced seniors. They’re finding people who are cheaper, more flexible, and better at using AI.

This time, it’s not the factory that closes. It’s the office. And you’re the one being displaced.

So What?

Musk’s vegetable garden isn’t happening. Total work isn’t shrinking. It’s growing.

But your seat might vanish. Not because work is disappearing — because the person doing the work is being replaced. Just like telegraph operators. Just like carriage artisans. Just like Detroit autoworkers.

And history tells us one more thing simultaneously. Every technology transition has followed a pattern. Every time, they promised “this time, humans will be free.” Every time, it was a lie. Every time, they told the displaced to “just retrain.” Every time, it was an insult. The real answer was always somewhere else entirely.

If you’re college-educated, white-collar, a developer, someone who followed the rules — if you’ve been doing what you’re told, staying comfortable in the direction handed to you...

What exactly makes you different from a Michigan factory worker?

In the next installment, I’ll talk about how to survive inside that storm. One hint: the answer isn’t where you think it is.

🔴 Next: [Next Michigan, Part 2] — “Surviving the Storm: Look at the North Star, Not the Wave in Front of You”

If this series resonated with you, please share it with someone who needs to hear it.

And honestly, wouldn’t this be the perfect setup to pivot to: “Want to learn how to survive in this world? Just sign up for my course! For the low price of $3,000, you too can learn to build an AI automation empire! Solo founder! One-person unicorn! If you could change your financial future, isn’t that a bargain?!”

Thanks for reading, as always.

— Ian

Plus: A free prompt so your pitch deck can face the world’s best investors

It took 30 minutes. I asked Claude to build a system where 20 real VC personas — each with their own distinct investment philosophy — independently evaluate a pitch deck, then the top 3 sit down for a 5-round investment committee debate. And it just... worked. Polishing the design and generating five sample reports took a bit longer, but the core build? Thirty minutes.

It started simply enough. An anonymous associate in our KakaoTalk group built a web app that uses AI to analyze pitch decks and wanted feedback. They were planning to sell it to startups as a paid service, which I thought was a solid idea. But my mind went somewhere different: instead of charging startups, what if I used something like this as a personal tool — a way to go deeper on the decks that land in my inbox?

So I sent Claude the screenshots, told it to replicate the functionality, and what came back was more than sufficient for what I needed.

For anyone who wants to try it themselves, I’ve published the full prompt on the website. Run it on Claude, ChatGPT, whatever you prefer. I’ve also added a direct link to pitch me your deck — we’re about to start sourcing pre-seed deals more aggressively (official announcement next week).

→ [AIVC Reports, Prompts, and Pitch Your Deck]

The results were more surprising — and more fun — than I expected.

I fed it thefacebook’s 2004 pitch deck. 9 out of 10 voted Invest. Zero passes. Near-unanimous. And yes, AI-Peter Thiel made the same call real Peter Thiel did.

Then I ran WeWork’s 2012 deck. 3 out of 10 voted Pass. The AI raised red flags. In reality? Masa Son committed $4.4 billion after a 12-minute meeting. (Maybe I should add Masa to the persona list and re-run it... )

I also tested Airbnb, Uber, and YouTube. The AI predicted outcomes reasonably well and surfaced genuinely interesting analysis. But here’s the thing — in every case, the actual investment was made by a human.

So can AI replace VCs?

The Question Itself Is Wrong

“Will AI replace VCs?” is the wrong question. It’s not a binary. The right question is: “Where exactly is the boundary between what AI can replace in venture investing and what it can’t?”

And that boundary isn’t defined by AI’s capabilities or some fancy system architecture.

It’s defined by the presence — or absence — of data.

From Public Markets to Venture: The Data Disappears

My entire career has been a journey from data-rich to data-poor environments. Math and economics undergrad, an economics PhD attempt (spectacular failure), an applied CS master’s — I wrestled with numbers for years. Then consulting → PE → VC → LP (reviewing VCs) → back to VC. Each step took me further from the world of abundant data into territory where numbers barely exist.

What drove each transition was a growing conviction: no matter how meticulously you analyze the numbers, it always comes down to people. I distinctly remember thinking, “An AI is going to automate this Excel work any day now.” That anxiety pushed me toward the parts of investing where humans still matter most.

And now that exact thing I worried about is actually happening.

So let’s map the investment world by data density:

Public equities — Real-time prices, financial statements, trading volumes, analyst reports. Everything is a number. Quants and algorithms already beat most humans here. (Though almost no quant fund has outperformed index funds over 15+ years, but that’s another story.)

Private equity — Satellite imagery counts cars in Walmart parking lots. AI analyzes ships docked at ports. Firms buy Whole Foods receipt data through loyalty point schemes. Quantification is advancing on all fronts, and now that LBO modeling is getting automated through Claude, AI’s territory expands daily.

Late-stage venture — Revenue, MRR, churn rates, CAC/LTV — real metrics exist. AI handles first-pass filtering and benchmark comparisons just fine.

Early-stage venture — No revenue. No users. The market itself might not exist yet. All you have is one human being called “a founder.” This is the domain of insight, reasoning, imagination — not spreadsheets.

The further you move from public markets toward early-stage venture, the more data evaporates. And AI without data is powerless.

AI doesn’t fail to replace VCs because AI isn’t good enough. It fails because the things that matter most in early-stage investing haven’t become data yet.

The Things That Never Make It Into a Database

I’ll share one story about why people matter most — and I generally avoid writing publicly about specific founders (unless they’re celebrities like Sam Altman), because I think it’s poor form. But this one stuck with me.

There was a founder who gave me a bad feeling. Not a smoking gun — just a collection of circumstantial signals that didn’t add up. My exact words at the time: “There’s smoke coming from everywhere, but I can’t find the fire.” (I should have remembered the old saying: where there’s smoke, there’s fire.)

The company’s growth numbers were strong, and the lead partner pushed hard. Most of the IC went along. But one partner dissented with a line I’ve never forgotten:

“Life is too short to work with people like this guy.”

We overruled that partner. The founder eventually imploded. The fund took a significant loss.

This wasn’t an isolated incident. I’ve watched company after company with picture-perfect metrics collapse — from bad luck, geopolitics, macroeconomic shifts, new technology, or founder mistakes. Each time, my conviction grew stronger: what matters isn’t today’s numbers, but whether a founder can survive whatever comes next.

Meeting people face-to-face. Extracting the information hidden behind the numbers that only you can access. That, I’ve come to believe, is the essence of venture capital.

I’ve said it many times: in early-stage investing, the founder is everything. Are they more obsessed with the problem than I am? Smarter about the solution than I am? More persuasive about their vision than I am? Would I want to work for this person? If they fail this time, would I back them again? None of this shows up in a spreadsheet. It’s not in any database. The only way to know is to sit across the table and talk.

To be clear — AI is excellent as a first filter. It structures the intuitions partners already have, catches blind spots, asks sharp questions. For pre-meeting research, it’s outstanding.

But the final call, and the accountability that comes with it, still belongs to humans.

Three Things That Will Change

None of this means VC stays the same. AI won’t replace VCs, but it will absolutely reshape the landscape.

First, the junior VC crisis. The traditional research-analyst junior role is dying. Big-picture research has been leveled up across the board. The tool I built in 30 minutes is proof. Market sizing, competitive analysis, benchmark comparisons — work that used to take a junior two days, AI now handles (imperfectly) in half an hour. What juniors need going forward isn’t basic research skills. It’s the ability to verify research quality and accuracy, the ability to build direct relationships with founders, operator experience in specific verticals, generational instincts that partners lack. Without an edge beyond research, survival gets hard.

Second, the rise of non-traditional VCs. As the research barrier to entry collapses, the intellectual capital that traditional VCs hoarded becomes commoditized. Knowing “how big is this market” used to be a competitive advantage. Now anyone can find out. Paradoxically, this creates opportunity. VCs with communities sense market needs first. Domain-expert VCs always have their research done. VCs with media platforms have founders coming to them. You don’t need a finance or founder background anymore — if you have your own unique leverage, this might be the best time ever to start a fund. Conversely, finance-background VCs who relied purely on market research to evaluate deals will be the first to struggle.

Third, it’s no longer about what you know — it’s about how you see. The VCs who get stronger from here are those who compete on perspective and insight, not volume of information. Insight isn’t something you Google. It lives inside people who’ve internalized it over years. As AI levels the research playing field, the axis of differentiation shifts from “what do you know” to “how do you see it.” Same with networks. It’s not “who do you know” — it’s “who have you made money with?” The strongest networks are forged between people who’ve succeeded and failed together. Not LinkedIn connection counts.

So What?

The democratization of research will actually make insight more visible, not less.

Howard Marks compared investment insight to height in basketball — you’re either born with it or you’re not. Annoying, but there’s some truth to it. Some people genuinely see what others miss.

But I think insight is closer to muscle than height. Height is genetic. Muscle is built.

In basketball, being tall helps. But height alone doesn’t get you to the NBA. Shooting form, court vision, defensive positioning — all built through thousands of hours of repetition. Steph Curry isn’t the tallest player in the league. But tens of thousands of practice shots created the muscle memory that made him the greatest shooter ever.

Investment insight works the same way. Constantly studying, agonizing over your own thesis, sitting through hundreds of pitch meetings, surviving dozens of failed investments, living through market reversals in your bones — this is how “this will work” and “this won’t” becomes instinct that sticks to you like muscle. Not in one or two years. Five years. Ten years. And it requires innate curiosity, the personality to go deep, and an environment that sustains it. You can’t just grind your way there.

Until now, this difference was hard to see. Do enough research, delegate to interns, pile up materials, and you could produce a convincing-enough investment memo. Time investment papered over the gap.

Now AI handles that part.

With basic research and information gathering democratized, the real capability gap — previously hidden behind sheer hours of work — becomes nakedly obvious. AI gives everyone height. More data, wider vision, faster analysis. But it can’t build the muscle for you.

Will IC partners trust a VC who shows up with a generic, AI-generated investment memo and no original thesis? Will LPs allocate capital to that person?

The gap between VCs who genuinely love this work and have spent years sharpening their own perspective, and those who’ve been investing by consensus and following the crowd — that gap becomes more visible than ever.

And that is exactly why VCs that can’t be replicated in 30 minutes will continue to exist.

Could AI Eventually Replace VCs?

Will there come a day when conversational nuance, a founder’s aura, the chemistry between people — all of it becomes data? I think it’s either very far off or permanently impossible. But if that day comes, then yes, AI could replace VCs too. But that would mean we’ve quantified human beings.

An AI that quantifies and judges people. Is that even possible?

And if it is — is that a world we actually want?

Thanks for reading, as always.

Ian

The Founder Influencer Trap: Startups Are Dying on LinkedIn and Instagram

You can’t growth-hack your way out of a retention problem.

By now you’ve probably seen the Cluely playbook. Rage-baiting, controversy-farming, pure social media chaos — and it worked. They went viral, raised $15M from a16z, and captured early users and revenue. (Full disclosure: I was one of the first VCs to meet with them. The founder has genuine raw intelligence and enormous energy, and I liked that - could pass the IC though)

His thesis? Get people in the door first. Build the product second. Pivot based on customer reactions.

I appreciate founders who can sell. Sales and marketing are among the most important skills in startups. But I can’t co-sign this strategy.

And yet, way too many founders are trying to copy the Cluely playbook.

Here’s the problem: most of them are doing it before their product is ready.

Even Cluely wasn’t immune to this. They got the customers, but the product was thin. Their “cheat on everything” software got blocked. Churn was brutal. Eventually, they pivoted to an AI notetaking app. (Apparently they’ve been too busy building to even post on Twitter lately — which, honestly, is probably correct.)

Retention is 100x more important than early growth.

Without retention, it doesn’t matter how many users you acquire. You can’t build sustainable revenue on a leaky bucket.

B2B SaaS became the darling of Silicon Valley for one reason: stickiness. High retention makes revenue predictable. Predictable revenue enables steady growth. Steady growth gets you to M&A or IPO.

But if your product isn’t ready and you suddenly go viral? Those users will be massively disappointed. They’ll leave. And they won’t just leave quietly — they’ll tell everyone how bad it was.

Disappointed customers don’t come back. Worse, they poison your future customer acquisition. Bad reviews spread faster than good ones.

What follows is a death spiral: low retention → stalling growth → harder fundraising → repeat until you’re dead.

So why do founders keep doing this?

Because improving product and retention is painful, slow, and invisible.

But growing followers, landing speaking gigs, accumulating downloads, and bumping up valuations? That’s easy to measure, easy to brag about, and feels like progress.

There’s also a darker logic: even if the company dies, your followers stay. In a twisted way, building a personal brand while your startup burns is like moving assets from the company to yourself.

I call this “Founder Influencer Syndrome” — The Founder Ego Trap. Founders who spend more time polishing their LinkedIn than their product. Who do the podcast circuit before they’ve found PMF. Who treat personal brand as the main event when it should be a side quest (if anything).

Let me be clear: if your product sucks and you’re constantly posting on Instagram, Threads, and X, doing speaking circuits like you’ve already made it — you’re not building a company. You’re building a content creator career with startup cosplay.

Investors notice. “This guy is always on social media” isn’t a compliment. Neither is “wow, interesting takes” when those takes reveal you’re not focused on the right things.

So how do you actually find PMF?

Old school, but it works: find 100 true fans who love your product. Talk to them obsessively. Iterate until they’re not just satisfied — they’re loyal. Build your moat from there.

Great products don’t need influencer founders. They grow on their own. Product-Led Growth is real. And if you’ve nailed retention but need more traffic, then you can consider becoming an influencer. Not before.

The startup ecosystem needs a values reset:

• DPI over unicorn status. Paper valuations are fantasy. Distributed cash is real.

• Retention over revenue. One-time sales are fantasy. Recurring revenue is real.

• Exits over fundraising. Round sizes are fantasy. Actual exits are real.

Clubhouse had explosive early growth. WeWork raised astronomical sums. Quibi had everything on paper. Where are they now?

Startups are a long game. You can’t day-trade your way to an exit.

To get to that exit — the only thing that actually matters — you don’t need an “influencer founder.” You need rock-solid recurring revenue.

So if things are going well right now, don’t get cocky. And if things are hard, don’t despair.

You haven’t failed until you’ve quit. It’s not over until it’s over.

Win the war, not just the battles.

As many of you know, I’ve been a die-hard SpaceX fan. When people doubted and mocked, I believed this would become one of the greatest companies in human history. I’ve toured the headquarters. I got invited to watch a launch (weather delay, sadly). I consider myself one of its biggest advocates.

So when Elon Musk announced that SpaceX is acquiring xAI at a combined valuation of $1.2 trillion—potentially the largest IPO in history—my first reaction wasn’t excitement.

It was: “Ah. So this is how it’s going to go.”

Having sat on both the GP and LP side of this industry, I’ve seen countless “restructurings.” I know what hides behind the elegant language of mergers. I’ve watched successful companies absorb failing ones to cover investor losses. I’ve seen what gets sacrificed in the process.

This SpaceX-xAI merger is the largest, most audacious, and most cleverly disguised piece of financial engineering I’ve ever witnessed.

So today, I’m writing not as a fan, but as an investor.

Chapter 1: The Dumpster Fire Called Twitter

Let’s rewind to 2022.

When Musk acquired Twitter for $44 billion, the world split into two camps: “genius move” and “this is insane.” I was firmly in the latter.

Why? Because the numbers told the story.

Twitter was already unprofitable at acquisition. It ran on advertising revenue, and advertisers started fleeing Musk’s chaotic management style almost immediately. Post-acquisition, ad revenue collapsed. The company’s value began melting off the balance sheet.

The real problem? The investors who participated in this deal.

Sequoia. a16z. Valor. Oracle. The most influential names in Silicon Valley bet billions on Musk’s Twitter LBO.

Why? Because it’s Musk. Everything he touches turns to gold. Tesla did. SpaceX did.

But this time was different.

Twitter didn’t turn to gold. It became a cash incinerator. Reports circulated that the company’s value had fallen to less than half of the acquisition price. Sequoia and other mega-funds had to book massive losses.

This is where Musk’s dilemma begins.

He’s not just a CEO. He runs an “empire of loyalty.” The investors who believed in him since early Tesla, who funded SpaceX, who joined the Twitter acquisition—if they take losses, Musk’s greatest asset crumbles: the myth that “if you’re on Elon’s side, you never lose.”

For Musk, this isn’t about reputation. It’s about survival.

Because all of his companies constantly need capital, and that capital comes from “people who believe in Musk.”

So he started making moves.

Chapter 2: The Washing Machine Called xAI

In 2024, Musk announced the merger of xAI and X (formerly Twitter).

The official reasoning sounded plausible. “AI and social media synergy.” “Grok learns from X’s real-time data.” “A new platform for the AI era.”

But the real reason was buried in the numbers.

Through this merger, Twitter investors received xAI equity. At valuations equivalent to their original Twitter investment ($44 billion). Their paper losses magically vanished.

“Made whole.”

Wait. Doesn’t that seem off?

How does a company with collapsing ad revenue, increasing user churn, and damaged brand value still get credited at $44 billion?

Simple answer: It can’t. Not by free market logic.

But in Musk’s world, it can. Because this isn’t a market transaction. It’s a “family deal.”

xAI, 100% owned by Musk, acquires X, where Musk is the largest shareholder, while distributing xAI shares to Musk’s friends. Value gets “determined” internally, with no external market validation.

You know what this is called?

Self-dealing. Conflict of interest.

It’s not illegal. These are private companies. But is it healthy? I can’t nod along to that.

Anyway, Phase 1 complete. Twitter investors are now xAI shareholders.

But there’s one remaining problem.

xAI doesn’t make money either. It’s another cash-burning beast.

Chapter 3: The Ark Called SpaceX

xAI builds “Grok,” an AI model competing with OpenAI, Anthropic, and Google.

You know what this business fundamentally is? A cash incinerator.

GPU clusters cost billions. Power bills run into hundreds of millions. Talent acquisition, hundreds of millions more. Why does OpenAI need trillions of won in investment every year? Why does Anthropic keep extending its hand to Amazon and Google? AI model development doesn’t make money—it burns money. At least for now.

xAI is no different. Maybe worse. OpenAI generates real revenue through ChatGPT. Grok is limited to the X platform.

So xAI investors—who are now also former Twitter investors—start getting nervous again. “Is this stock actually worth anything?”

Right on cue, Musk plays his card.

SpaceX.

The most successful private space company in human history. A company generating real cash through Starlink. A company leading humanity into a new era. A “guaranteed” trillion-dollar valuation the moment it IPOs.

February 2, 2026: SpaceX acquires xAI. Combined valuation: $1.2 trillion. xAI’s implied value: over $200 billion.

Now the picture is complete.

Twitter investors ($44B bet, facing losses)
    ↓ X-xAI merger → xAI shares
xAI shareholders (burning cash, exit unclear)
    ↓ xAI-SpaceX merger → SpaceX shares
SpaceX shareholders (holding the most valuable private stock on Earth)

People who were about to lose money on Twitter, after two mergers, now hold SpaceX stock.

You know what the press(Axios) is calling this?

“Lifeline.”

That’s the actual word from the coverage: “xAI shareholders effectively get a lifeline through the deal.”

The journalists know exactly what this is.

Chapter 4: The Gift Wrap Called “Sentient Sun”

Of course, Musk doesn’t call it a “bailout.”

He’s a poet. A dreamer. A prophet of humanity’s future.

“SpaceX has acquired xAI to form the most ambitious, vertically-integrated innovation engine on (and off) Earth. This marks not just the next chapter, but the next book in SpaceX and xAI’s mission: scaling to make a sentient sun to understand the Universe and extend the light of consciousness to the stars!”

“A sentient sun to extend the light of consciousness to the stars.”

Beautiful words. Genuinely. Reading them, your heart races and you can almost see humanity’s future unfolding before you.

But reading that sentence, I had a different thought:

“This guy is an absolute marketing genius.”

Because that one sentence buries every uncomfortable question.

“Why acquire an unprofitable AI company?” → “Space data centers need AI.”

“Why are Twitter investors getting SpaceX stock?” → “Vertically integrated innovation engine.”

“Why do this complex merger right before IPO?” → “We must extend the light of consciousness to the stars.”

He’s wrapping financial engineering in sci-fi vision.

And remarkably, people believe it. Or rather, they want to believe it.

Because it’s Musk.

Just like people believe Tesla isn’t “just a car company” but will become an “AI robotics company,” they want to believe SpaceX isn’t “just a rocket company” but will become a “space AI infrastructure company.”

I don’t want to dismiss this entirely. Maybe space data centers will exist someday. Maybe Grok will power Mars rover brains.

But “maybe someday” and “worth $1.2 trillion right now” are completely different propositions.

And right now, here’s what’s certain:

People who were about to take losses on Twitter and xAI have boarded the ark called SpaceX.

Chapter 5: The Real Game Hasn’t Even Started

If you’ve read this far, you might be thinking:

“Okay, he rescued Twitter and xAI investors. Feels a bit off, but SpaceX is such a great company—isn’t it fine?”

I thought so too. At first.

Then I thought about one more thing, and it sent chills down my spine.

Tesla.

Do you know Tesla’s current situation?

The EV market is no longer Tesla’s alone. China’s BYD, European legacy manufacturers, countless startups—everyone makes EVs now. Price competition is brutal, margins are shrinking, growth is slowing.

In January, Tesla announced it’s discontinuing Model S and X production. Official reason: “Converting production lines to Optimus (humanoid robot) manufacturing.”

I interpret this differently.

As always, Tesla can no longer justify its stock price as a “car company.” Car company P/E ratios typically hover around 10x. Tesla’s? Tens to hundreds of times that.

What bridges that gap?

“Expectations about the future.” Self-driving. AI. Humanoids.

So Tesla must keep saying “we’re not a car company, we’re an AI robotics company.” The Model S/X discontinuation—driven by profitability issues as a car company—becomes part of that narrative.

“See, we’re abandoning old-era cars and moving to future robots.”

The question is how long this narrative can survive its collision with reality.

Space data centers? Thermodynamics experts point out that “vacuum isn’t a freezer, it’s a thermos”—GPU heat has nowhere to go.

Optimus? Interestingly, a16z—who participated in the Twitter bailout—has itself warned about “the uncrossable gap between demos and reality.”

And the bigger issue is Musk’s personal financial situation.

Musk has borrowed heavily against his Tesla shares. A significant portion of Twitter acquisition funding came from these margin loans.

If Tesla stock crashes? Collateral value drops, margin calls come, he must either add collateral or sell shares. Selling shares pushes the price down further. A death spiral begins.

In other words, defending Tesla’s stock price is Musk’s personal survival condition.

Now let’s connect the puzzle pieces:

• Tesla is losing its appeal as a car company

• Maintaining the stock price requires the “AI robotics company” narrative

• But Optimus is far from commercialization, and FSD keeps saying “next year”

• Meanwhile, SpaceX is acquiring xAI and transforming into a “space AI infrastructure company”

• All of Musk’s companies are beginning to merge into one narrative

See where this is heading?

Chapter 6: X Holdings, or Tesla’s Funeral

I believe Musk ultimately won’t leave Tesla as a “car company.”

The most likely scenario: the birth of “X Holdings.”

A holding company structure placing SpaceX, Tesla, xAI, and X under one roof. Like Alphabet (Google’s parent) or Meta.

What happens then?

Tesla’s weak automotive performance gets diluted within SpaceX’s growth story. Tesla shareholders become not “car company shareholders” but “space+AI+robots+social media conglomerate shareholders.” Even if individual business units underperform, the group’s “vision” can defend the stock price.

Looking at it optimistically, could it become a second Berkshire?

But if this happens, Musk’s influence won’t be what it was. Once everything becomes public companies, there’s less room for the current maneuvers.

Another scenario: “Cherry-picking.”

Merging all of Tesla is too big and complex, so just extract the core. The Optimus division. Self-driving technology. The AI team.

“Tesla’s robotics technology is essential for SpaceX’s Mars colonization. It must combine with xAI’s brain.”

Sounds plausible, right?

The SpaceX+xAI alliance acquires Tesla’s robotics division for a massive sum. Tesla gets huge cash infusion, balance sheet improves. Future businesses transfer to the unlisted mega-entity.

What’s left at Tesla?

The shell of car manufacturing. Gradually becoming “SpaceX Group’s robot production subcontractor.”

This won’t be easy, of course. Tesla is public. The SEC watches. Countless retail and institutional investors own shares. Shareholder lawsuits could rain down like with SolarCity. Antitrust issues exist.

Nevertheless, Musk will absolutely deploy financial engineering to protect his empire.

And the allies who’ve made money through him will support him.

So I think this is a plausible scenario.

Chapter 7: Why Are SpaceX Shareholders Staying Quiet?

A question arises here.

SpaceX’s existing shareholders aren’t stupid. Why do they agree to this merger? Why accept mixing “cash-burning beasts” into their “golden goose”?

Several reasons.

First, short-term gain.

SpaceX’s valuation jumped from $800 billion to $1.2 trillion. Book value of assets up 50%. Looks good for now.

Second, faith in Musk.

SpaceX’s early investors have been with Musk for a long time. They’ve watched him accomplish things that seemed impossible multiple times. “There must be something to this too.”

Third, no alternatives.

SpaceX is private. You can’t sell shares whenever you want. Without Musk-approved secondary transactions, there’s no liquidity. Even if you’re unhappy, you can only “grin and wait for IPO.”

Fourth, they’re all in the same boat.

Musk’s smart move: he already gave Twitter and xAI investors opportunities to invest in SpaceX too. Almost no one invested only in obviously-doomed Twitter or xAI. Most wanted to get in on “humanity’s future” companies like SpaceX or Neuralink.

Fifth, the carrot of IPO.

“The largest IPO in history.” “Over $1.5 trillion listing.” Against this dream, a little risk seems acceptable. Better to ride the flagship than get stuck with random small-caps.

But I see this as “forced boarding on a gamble to inflate size and list at higher prices while bearing the risk.”

SpaceX was originally a clean company. Build rockets, launch satellites, make money with Starlink. Simple and clear. “King of space transport and future of internet.”

Now? It’s become a “space+AI+social media monster.” Valuation methodology got complicated. Analysts now have to wonder “how do we even value this thing?” Major variables added to IPO momentum and timing.

Plus, xAI and X’s risks are now mixed into SpaceX’s financials. Cash burn, regulatory risk, brand risk. All of it muddying the clean image of the “space flagship.”

As I mentioned on my English podcast last week, I’m upset as a SpaceX shareholder. (Kevin, who only holds xAI, is thrilled lol)

But SpaceX shareholders can’t say anything.

Because without Musk, there is no SpaceX.

Chapter 8: The Essence of What We’re Witnessing

Let’s return to the beginning.

SpaceX has acquired xAI. Combined valuation: $1.2 trillion. Probably the largest IPO in history.

Is this really “the marriage of space and AI”? “The next chapter of human civilization”?

I see it differently.

This is the process of Elon Musk using one of humanity’s greatest private companies as collateral to bail out his failed bets and escape his financial predicament.

The Twitter acquisition failed. Ad revenue collapsed, brand value was destroyed. But Twitter investors received xAI shares and covered their losses.

xAI isn’t profitable yet. It’s burning massive cash. But xAI investors received SpaceX shares and secured their exit.

Tesla is losing its growth engine. Its appeal as a car company is fading. But Musk defends the stock with a “AI robotics company” narrative about the distant future, and will ultimately try to integrate it into the SpaceX group to protect his empire.

At the center of all this sits SpaceX.

The company that actually makes money. The company with a certain future. The company that will be valued astronomically the moment it goes public.

SpaceX is the breadwinner of Musk’s empire. The lifeboat. The ark.

And right now, that ark is being loaded with all kinds of cargo.

A ship that once carried only rockets and satellites is now loading AI models, social media, and perhaps soon, cars and robots.

The ship is getting heavier.

Epilogue: The Shadow of “Sentient Sun”

I used to respect the old Elon Musk. Genuinely.

Without him, humanity would still be using disposable rockets. EVs would still be treated as “golf cart cousins.” He accomplished things that seemed impossible and inspired countless people.

But watching his post-COVID trajectory and this latest move, I’m confronting an uncomfortable truth.

A great visionary and a cunning financial engineer can be the same person.

“Sentient Sun” is a beautiful vision. Extending the light of consciousness to the universe—what a magnificent story. But beneath that vision’s shadow, failed investments are being laundered, losses transferred, and risks diluted.

I don’t want to call this “fraud.”

It’s not illegal. All transactions were consensual. Investors participated voluntarily.

But I also can’t call it “just.”

SpaceX was built by thousands of engineers pouring their blood and sweat. A company that made humanity’s dream of space exploration real. That company is becoming a dumping ground for one person’s distressed assets.

Of course, SpaceX will probably remain a great company. Rockets will keep flying, Starlink will keep making money, and we’ll soon send humans to Mars.

But in the process, we’ve lost our SpaceX.

The “rocket company” SpaceX that was making humanity a multiplanetary species and bringing internet to the world. The SpaceX with clean valuation. The SpaceX where we never had to think about financial engineering or shareholder bailouts.

The era of genuine innovators is over.

Whoever Controls Voice Will Control AI

And Apple just dropped $2 billion to prove it

Another OpenAI hardware rumor dropped recently - this time, earbuds that would compete directly with AirPods. And then this morning, Apple announces its second-largest acquisition ever: Q.ai, an Israeli startup that can read your facial micro-movements to understand whispered and silent speech.

Coincidence? I don’t believe in those anymore.

Let me break down why this matters - and why I’ve been pounding the table on this thesis for two years.

The Real Value of AI Isn’t Intelligence - It’s Interface

I know you’re tired of hearing this from me, but I’ll say it again: the most underrated value of this generation of AI isn’t the “intelligence” part. It’s the interface revolution.

Throughout history, the humans who win are the ones who use tools best. Computers are humanity’s greatest tool. Therefore, the winners of this era will be the people who use AI to maximize their computer productivity.

Current-gen AI can’t truly reason or imagine. What it can do is run calculations that humans don’t have time for, and organize the internet’s entire knowledge base faster than we ever could. But here’s the thing: computers have always been better than humans at certain tasks. What’s different now is that the gap is widening exponentially at this inflection point.

That means the real question isn’t “how smart is the AI?” It’s “how quickly and accurately can humans interface with it?”

The Input Revolution: Voice and Context Dumps

Voice is dramatically faster than keyboards or mice as an input method. AI transformed what used to be useless data (messy human speech) into the most powerful input mechanism we have.

The era of staring at your iPhone, hunting for apps, and pecking at a tiny keyboard is ending. Once Gemini fully integrates with Apple’s foundation models and Siri becomes genuinely intelligent, most simple tasks will be voice-first. At minimum, the whole “find the app, open it, navigate the interface” workflow will disappear.

Context Dumps

Here’s how I actually use voice now: I do what I call “context dumps” - just rambling stream-of-consciousness into a recorder, no filter, capturing all the threads and tangents of whatever I’m thinking about. Then I let AI clean it up.

This newsletter? I spent 15 minutes talking through ideas from a conversation I had with SNU entrepreneurship students yesterday. That voice memo became the skeleton for everything you’re reading.

The advantage: AI understands my intent much better when I give it more context in the same amount of time. And I can store those insights for future use. Sure, I occasionally ask Gemini to tell me everything it knows about me so I can correct it, but more context generally means better understanding, which means better output.

The Dream: 24-Hour Context

This workflow made me realize something: I want an AI that maintains context on my thoughts and conversations 24/7. Understanding things about myself that I don’t even consciously know. Never missing details about work or life. Always maintaining that thread.

Beyond being useful, there’s something delightful about learning things about yourself you didn’t know. And this continuous context-gathering is ultimately how you build the perfect “AI assistant” - one that knows your preferences and intentions at all times.

Privacy and On-Device AI (It’s Already Here)

The obvious concern with 24-hour voice capture is privacy. I share it. I periodically delete my cloud AI conversations and avoid sensitive topics entirely.

That’s why I’ve been advocating for on-device AI for two years now - models that run locally, keeping your data on your hardware.

Here’s what convinced me we’re closer than I thought: I recently found an app called Whisper Note on the App Store. Five dollars. 700MB. A high-quality AI model that runs entirely on your iPhone. Lifetime access, no subscription, local processing.

I’d tried similar apps before and they were garbage. This one? It performs as well as the cloud-based subscription service I was paying for monthly. It handles my English-Korean code-switching perfectly. There’s a Mac version too. I’ve completely replaced my paid subscription.

From Scaling to Optimization

This is what AI’s future looks like. We’ve hit diminishing returns on the “make the model bigger, make it think longer” approach. Now everyone’s focused on making models cheaper, smaller, and more efficient.

On-device AI is coming faster than most people realize. The opportunity for VCs and founders is figuring out the new applications and business models that this environment enables.

The Output Dilemma: Maybe Language Itself Needs to Evolve

I love voice as input. But for output? Reading is still faster than listening.

That’s why companies like Meta are trying to put screens directly in front of your eyes with glasses. But I think that hardware solution is further away than people expect. (More on this below.)

So here’s a weird thought: maybe instead of waiting for hardware to catch up, human language itself will evolve.

When we read, we don’t process every letter or even every word sequentially. We step back and pattern-match on keywords, absorbing paragraphs at a glance. But voice communication has always been strictly linear - one word after another.

What if AI could deliver voice output differently? Faster speech with variable pacing. Keywords emphasized. Overview-first structure. Non-linear audio that mimics how we actually process text?

If AI’s speed and repeatability could help us experiment with new forms of verbal communication, we might discover a language structure optimized for voice output. And then we’d truly enter the screenless voice era.

How AI Is Changing Human Communication

Language has always evolved - “vibe,” “rizz,” “67.” That’s not new. What’s new is that I now spend more time talking to AI than to humans. My personal language patterns are shifting. Future generations will shift even more.

Think about it: an AI that already knows my context better than any human, remembers everything, and doesn’t require social niceties or emotional management? I can be radically more efficient in that conversation. Shorter. More direct. And that efficiency might evolve into an entirely new mode of AI communication.

I suspect this is happening to many of you without you realizing it.

The concerning part: what happens to human-to-human communication in a world where efficient AI-speak becomes the norm? Traditional complete sentences might become as rare as Latin. And we’re already living in an era where human communication is painfully difficult.

My guess? Just like we now write emails that AI writes for the recipient’s AI to summarize, we’ll eventually have AI conducting conversations on our behalf.

People keep saying “being human” and “imperfection” are how we’ll survive the AI era. But when I watch GPT mimicking “um” and “ah” filler words, I think: AI can perfectly imitate imperfection too. I don’t have an answer here yet.

AI Hardware’s Future: Open-Ear and Device Mesh

Apple’s $2 Billion Bet on Q.ai

As if on cue, Apple just announced its acquisition of Q.ai this morning - reportedly its second-largest acquisition ever at roughly $2 billion.

Q.ai’s technology reads “facial skin micro-movements” to understand whispered and silent speech. Their patents show applications for headphones and glasses. The founder, Aviad Maizels, previously sold PrimeSense to Apple in 2013 - the company that enabled Face ID.

This is exactly the direction I’ve been predicting.

The Open-Ear Thesis

For any device capturing 24-hour context, you need something comfortable enough to wear all day. It needs to let you hear your environment while also letting only you hear the AI clearly. That points to open-ear form factors - possibly bone conduction - rather than earbuds that block outside sound.

Q.ai’s technology adds another dimension: you could potentially communicate with AI silently, using only facial micro-movements. No speaking required.

Device Mesh, Not Single Device

I don’t think one device will do everything. Instead, I see a mesh of redundant devices working together: open-ear audio as the foundation, supplemented by rings (cover your mouth for private communication), glasses, pins, necklaces, pens. Each device catches what the others miss. Together, they maintain continuous context.

I’m skeptical about glasses as the primary device. For people who don’t already wear glasses, they’re heavy, uncomfortable, hot, short battery life. Forcing adoption is hard. And getting a usable display into glasses is further out than the hype suggests.

Meta’s neural wristband that captures micro-muscle movements? That might actually be a more promising input mechanism than glasses.

Full disclosure: I’m an angel investor in an early-stage team working on exactly this thesis. Still in stealth mode, so I can’t share names, but I’m putting my money where my mouth is.

Quick Sidebar: Gen Alpha and Visual Storytelling

One more output evolution worth noting: the next generation after Gen Z - Gen Alpha - are complete video natives. They don’t read long text. For them, the ability to instantly convert context into visual storytelling matters enormously.

That’s why I’m bullish on tools like StoryTribe that help creators work with AI on visual content. Think “Canva for AI.”

I tell founders this constantly: future communication will be built on “minimum words” plus “visualization.” VCs with ADHD and 5-second attention spans are basically Gen Alpha in adult bodies. Pitching us requires deep understanding distilled into the fewest possible words and clearest possible visuals.

Going the other direction - writing long explanations (like this newsletter, yes, I see the irony) - shows you spent time, but it also dumps the comprehension burden on the reader. The people who can compress complex ideas into single images? They demonstrate mastery.

To sell to VCs and Gen Alpha alike, voice and visuals will matter more than text. I want to keep thinking through this evolution with all of you.

So What? Consumer Hardware’s Great Reset

Historically, every interface shift triggers massive wealth transfer.

The web-to-mobile transition gave us Uber and Instagram. The AI-and-voice transition is a “Great Reset” where every piece of software and every service starts from zero on a level playing field.

Whoever understands and rides this wave will dominate the next era.

I want to meet founders with this ambition. Let’s think through it together.

(TMI: I’m in “closed-door cultivation” mode right now, so April is the earliest I can meet. But send me an email at ian@ianpark.vc and I’ll follow up then!)

Thanks for reading.

Ian

Watching Microsoft’s stock crater today reminded me of something I talked about with my VC friends a few days ago. Let me be clear: I have zero talent for predicting stock movements. (If I could, I’d be trading, not writing newsletters.) But having built my entire career in finance—where Excel is basically scripture—I’ve been thinking about what happens when AI finally conquers the spreadsheet.

The Headlines vs. The Real Story

The official narrative for today’s drop:

Astronomical AI spending with nothing to show for it.

Margin collapse—profit margins fell from 72% to 67% over the last five quarters.

Azure deceleration—cloud growth disappointing.

All valid concerns. But I think the real existential threat is something else entirely.

The collapse of Excel’s moat.

The Liberation of the “Shortcut Slaves”

Here’s something non-finance people don’t understand: The reason finance stayed married to Excel—and by extension, Windows—was never about performance. It was about muscle memory.

Decades of keyboard shortcuts, burned into the fingers of every banker and analyst. Windows Excel shortcuts are completely different from Mac Excel. Google Sheets, being browser-based, could never replicate the feel because of shortcut conflicts. This “familiarity” was Microsoft’s most powerful weapon for 30 years.

Then Claude in Excel happened.

The moment AI can write formulas and understand data flows, that “no-mouse, keyboard-only” flex that finance bros have bragged about for decades becomes a relic. Thirty years of accumulated shortcut mastery? Worthless overnight.

Google’s Counterstrike: When Skill Barriers Disappear

When tool proficiency stops being a competitive advantage, the entire game changes.

If the keyboard shortcut barrier vanishes, why use heavy, clunky Excel at all? A mass migration to Google Sheets—lighter, collaboration-native—becomes inevitable.

But here’s what’s scarier: distribution. Google can plug any AI model into its Workspace ecosystem instantly. What Claude can do today, Gemini will do tomorrow. As AI performance converges, the winner is whoever has better UX, easier access, and stronger alliances. Google has all three—plus that Apple partnership.

This isn’t just an Excel problem. It threatens the entire Windows OS and PC ecosystem.

The New Order: Google + Apple vs. Microsoft

Microsoft falls behind

Let’s be honest: Companies have stuck with Microsoft because of security concerns and bundling—not because the tools are actually good. (Hot take: None of Microsoft’s tools work particularly well.) If Excel—the crown jewel of that bundle—crumbles, the reason to stay disappears.

Security is more complicated. Many enterprises still distrust the cloud. But we’re approaching the tipping point where AI productivity gains outweigh security convenience.

Sure, companies with seasoned (read: change-resistant) CTOs will move slowly. But eventually everyone faces the “adopt or die” moment.

Using Microsoft won’t kill your company tomorrow. But when everyone else is doing 100 km/h with AI and you’re stuck at 10 km/h on legacy tools, getting left behind isn’t a possibility—it’s a certainty.

Google and Apple pull ahead

Hardware already tilted. Apple’s M-series chips deliver exactly what the AI era demands—performance, boot speed, battery life, longevity. Outside of gaming, the PC advantage keeps shrinking. And gaming? Steam Deck and the upcoming Steam Machine 2026 are chipping away at Windows’ last stronghold. Even in AR/VR, Microsoft is losing to Meta and Apple.

Google’s data and ecosystem plus Apple’s hardware dominance—if they start competing on price with their cash reserves? The AI era gets divided between Google (software) and Apple (hardware).

So What Does This Actually Mean?

The change won’t come from Microsoft shops suddenly switching. (Organizations that flexible probably already left.) It’ll come from Google-native companies dramatically outproducing Microsoft-native ones, growing faster, and eventually reshaping the market itself. Microsoft doesn’t collapse—it just gets relatively smaller as the Google + Apple camp grows disproportionately larger.

Could AI trigger not just the decline of B2B SaaS, but the decline of Windows OS itself?

I built my career on Excel. I was a devoted ThinkPad Carbon user, a true believer in the red TrackPoint nub. I only recently defected to the Apple ecosystem. So maybe I’m overreacting to this shift because it’s personal.

But even accounting for that bias—the fall of Microsoft Excel could shake the global financial system. It would be a historical event none of us forget.

Just something to think about.

Gemini’s rise might just hand Google a perfectly legal monopoly.

I’ve been using Gemini lately and noticed something interesting: ask it about me or my work, and it disproportionately surfaces YouTube videos—especially ones I’ve made or appeared in.

I don’t think this is Gemini deliberately promoting YouTube. It’s structural. Google understands exactly how Gemini learns and retrieves data, and they’ve obviously built the pipeline so YouTube content integrates most seamlessly into AI search results.

Flip it around: YouTube has a massive advantage in “Gemini Search Optimization” (call it AEO, GEO, whatever). Text, video, transcripts—everything lives inside Google’s ecosystem where Gemini can read it most efficiently.

Here’s the uncomfortable part: as Gemini’s market share grows, Google’s entire ecosystem—Shopping, Maps, local data—becomes the default source of truth for AI answers.

Looking for a restaurant? Google Maps advertisers surface first. Protein powder? Google Shopping ads. Flights? Google Flights inventory. Any topic? YouTube videos beat TikTok and Instagram every time.

And here’s the genius: Google announced Gemini won’t have ads. Sounds noble until you realize they don’t need them. OpenAI has to figure out how to monetize ChatGPT. Google just routes Gemini queries to platforms already printing money. The ads are already there—baked into Maps, Shopping, YouTube, Flights. Gemini just becomes the world’s most sophisticated traffic funnel to Google’s existing cash machines.

In the old world, this would be textbook self-preferencing—instant antitrust bait. But AI gives Google an incredible defense: “It’s a black box. We don’t control the outputs.”

Their argument: “We’re not manipulating results. Our Shopping and YouTube data just happens to be most ‘optimized’ for general artificial intelligence to understand.”

If that flies, Google completes a structure where they legally capture ad revenue, commerce transactions, AND AI subscription fees. All of it, all at once.

That’s the future we should be watching for—or pricing in.

But there’s a problem: the AI industry’s definition of “ARR” would make any SaaS veteran cringe.

If you’ve spent time in enterprise software, you know ARR is sacred. It’s the metric that built the last decade of tech investing. Predictable, contracted, sticky revenue. The kind that lets you sleep at night.

Here’s the thing: the AI industry didn’t lie about ARR. They just redefined it—and hoped nobody would notice.

The Bait and Switch

Traditional ARR (Annual Recurring Revenue): Revenue from 1-3 year contracts. Customers are locked in. Churn is low. The best companies actually grow revenue from existing customers over time. This is the metric that justified SaaS’s premium valuations.

AI’s “ARR” (Annualized Run Rate): Last month’s revenue × 12. That’s it.

Every curious user who tried the product once. Every developer who ran a quick API test. Every company that kicked the tires and moved on. Multiply by 12, call it ARR, and watch the headlines roll in.

The term “run rate” has existed forever. But it was always clearly labeled. Calling it “ARR” in the AI era is a deliberate choice—borrowing SaaS’s credibility to dress up fundamentally different economics.

Tenants vs. Tourists

The distinction matters because of what happens next.

SaaS customers are tenants. They sign multi-year leases. Annual churn at top B2B companies runs around 5%. Many actually increase spending over time—net dollar retention of 100-120% is common. The revenue compounds.

AI customers are tourists. No long-term commitments. Why would there be? A better model drops next month, and switching costs are near zero. AI applications are seeing 3-8% monthly churn. Annualize that: you’re losing 40-60% of your customer base every year.

You’re filling a bucket with a hole in the bottom while getting valued as if it’s overflowing.

This isn’t theoretical. Even the market leaders feel it. Every time a competitor releases a better model, customers migrate overnight. When switching costs are zero, loyalty is a fantasy.

Software Margins, Manufacturing Economics

There’s another reason SaaS commanded premium valuations: margins. Software costs virtually nothing to replicate. Copy, paste, sell. Gross margins of 80%+ were standard.

AI flips this on its head.

Every query burns compute. Every response costs electricity. Revenue scales, but so do costs. It’s manufacturing economics wearing a software costume.

Most AI companies are deeply unprofitable, burning billions annually while racing to capture market share. The bull case assumes costs will eventually decline and margins will expand.

Maybe. But here’s the catch: when GPU costs drop for you, they drop for everyone. Competitors get the same benefit. Prices race to the bottom. We’ve seen this movie before—it’s called commoditization.

I actually believe AI will eventually reach SaaS-like economics. Scaling laws are plateauing. We’re entering the optimization era. On-device inference is coming.

But the question isn’t whether we get there. It’s who survives long enough to see it—while burning billions along the way.

The Valuation Paradox

Here’s what doesn’t make sense:

Traditional SaaS companies—profitable, with genuine recurring revenue and decades of customer lock-in—trade at 6-10x revenue. That’s the historical norm.

AI companies? Median private valuations run 20-30x revenue. The hottest names command 80x, 100x, even 200x.

The math doesn’t work. AI revenue is:

• Less predictable (monthly vs. contracted)

• Less sticky (high churn vs. locked-in)

• Lower margin (GPU costs vs. near-zero marginal cost)

By every measure, AI revenue is lower quality than SaaS revenue. Yet it commands multiples 3-10x higher.

The market is pricing in certainty that doesn’t exist. Remember when everyone declared Google dead? They came roaring back. This game is far from over.

How I’m Evaluating AI Deals Now

I’m not saying AI isn’t transformative. It obviously is. And some companies will justify their valuations and then some.

But when I look at AI investments now, I ask three questions:

1. What’s the actual contract structure? Monthly subscribers ≠ enterprise contracts. A company with $10M in 3-year enterprise deals is worth more than one with $20M in monthly API revenue. The distinction is everything.

2. What’s the churn profile? If they won’t share cohort data, assume the worst. The best companies are transparent because they have nothing to hide.

3. What’s the path to positive unit economics? Not “eventually”—what’s the concrete plan? What has to be true for margins to expand? What’s the timeline?

The best companies answer these clearly. The ones playing valuation games dodge.

The Bottom Line

The AI industry has pulled off a neat trick: repackaging volatile, low-margin, high-churn revenue as “ARR” and commanding the highest multiples in tech history.

It’s the oldest game in bubbles—invent new metrics to justify new valuations. “This time is different” has been the rallying cry of every mania in history.

Next time you see a headline about AI revenue milestones, ask the only question that matters:

Is this tenant revenue or tourist revenue?

The answer will tell you whether you’re looking at a business—or a bet.

Thanks for reading. Hit reply if you have thoughts.

This piece was adapted from my Korean newsletter 주간실리콘밸리 — if you read Korean, subscribe here.

— Ian

If you’ve been reading my newsletter, you know I’ve been pretty bearish on vibe coding and GPT wrappers. I’ve written about how vibe coding is a security nightmare, why solo/small teams will inevitably hit a wall, and why “speed is our only moat” is a fundamentally flawed thesis.

The pushback I always get? “But vibe coding companies are doing great!” or “GPT wrappers like Perplexity are killing it!” My answer has always been the same: “Companies take a while to die. Be patient.”

And now, the dying patterns are starting to show.

The Retention Data Doesn’t Lie

Chamath — who admittedly gets it right sometimes — shared this data on X showing 12-week cohort retention for vibe coding platforms. The picture isn’t pretty. The deeper you go into the cohorts, the more red you see, which means one thing: users are churning hard. People sign up, try it once, go “meh,” and bounce.

These platforms are essentially Claude wrappers. And as someone who values retention roughly 100x more than top-line revenue growth, this data does not signal an industry heading in a healthy direction.

The Obvious Limits of Vibe Coding: A House Built on Sand

I’ll give credit where it’s due. Vibe coding genuinely helps skilled developers boost productivity. It’s useful for quick MVP testing and building simple internal tools. Think of it like the Excel macros (VBA) that companies already build to automate tedious tasks — vibe coding handles that reasonably well.

But here’s the thing: you can now just use Claude Code directly from Anthropic to build those simple tools. You don’t need a dedicated vibe coding platform for that.

So can vibe coding build a real product that changes the world? Not a chance.

There’s no competitive advantage. If anyone can build it, there’s inherently no barrier to entry. The moment someone makes money with a vibe-coded product, someone else clones it overnight. Market defense is impossible.

Maintenance is a disaster. AI that can’t think independently — that just pattern-matches and stitches code together — produces something that looks functional on the surface. But it’s a ticking time bomb. Try adding one feature or fixing one bug, and the whole thing collapses.

You can never be the best. Great products demand meticulous architecture and craftsmanship. Vibe-coded products have neither depth nor durability. A clever idea might generate a brief spark, but without a moat, retention drops, and Google or a more polished team eventually eats your lunch.

Companies might use vibe coding for experiments or internal tools. But no serious business will entrust its core software to it. That’s why I’m confident that vibe coding tools — which are ultimately just Claude wrappers — won’t become mainstream. The “one-person unicorn” era isn’t coming. The “teams don’t need engineers” future isn’t happening.

The Dotcom Parallel: Namo Web Editor and Cursor

A funny analogy hit me recently. What we’re living through right now feels exactly like the dotcom era, when Namo Web Editor came out and everyone declared that web development would be fully automated.

For those unfamiliar, Namo Web Editor was a Korean WYSIWYG (what-you-see-is-what-you-get) HTML editor that was huge in Korea during the late ‘90s and early 2000s. Back then, simply being able to make a website was extraordinary — that’s the pre-vibe-coding era equivalent. Namo Web Editor was essentially the Cursor of the dotcom bubble.

So let’s all answer together: Did web development get fully automated?

Sure, it got easier for personal sites and nonprofits. But the websites that actually compete on quality? They got more beautiful, more complex, more sophisticated. As user expectations rose, web development evolved into a domain requiring deep expertise. Demand for great developers kept increasing. All Namo Web Editor solved was the simplest layer of the problem — which freed real developers to focus on higher-order challenges.

The same logic applies here. AI-powered coding tools aren’t opening the door for regular people to build world-class apps and get rich. They’re giving professional programmers the bandwidth to tackle harder, more complex problems than ever before.

So What Happens Next?

AI coding tools won’t just make existing things faster. They’ll make previously impossible things possible. We’re talking about an entirely different dimension of products — paradigm-shifting stuff.

Claiming vibe coding will replace programmers is like standing at the dawn of the internet and having no concept of what an iPhone app or VR/AR experience would look like. We’re going to encounter entirely new forms of human-computer interaction (HCI) that aren’t websites, aren’t mobile apps — something we can’t even imagine yet.

Technology has always created complexity and functionality beyond our imagination. This drives fiercer competition, which demands better talent, more capital, and longer timelines.

The AI era won’t be an exception. The nature of the work will change, but demand for elite developers will only increase.

First Principles Thinking Has Never Mattered More

I’m bearish on vibe coding platforms, but let me be clear: AI genuinely boosting skilled programmers’ productivity is a great thing. It frees top-tier engineers to work on higher-order problems. That’s unambiguously positive.

But this means we need to completely rethink the frameworks we inherited from the internet and mobile eras. We have to break out of old mental models. Building existing web apps and mobile apps faster is not how you win in this era. Truly great companies will build something beyond that.

What does that look like? Honestly, I don’t know yet. But it will likely involve hardware evolution alongside software — most people seem to agree on that much.

Getting there requires going back to the most fundamental question: “What problem are current websites and mobile apps actually solving for consumers?” Start from zero. No assumptions, no inherited frameworks. Pure first principles thinking.

Take Gamma AI, for example. They’re doing well, but at the end of the day, it’s a wrapper that makes PowerPoints faster. When Google Slides or a better tool shows up, it gets replaced instantly.

For a company like that to truly succeed, the question can’t be “How do we make PPTs better and faster?” It needs to be “How can AI help people communicate business ideas more effectively?” — and then build the answer from scratch.

Don’t “automate PowerPoints with AI.” Make AI do what PowerPoints were trying to do, but better.

Don’t “automate web browsing with AI.” Make AI accomplish what users were trying to do in the browser, but easier and faster. (This is exactly why ChatGPT’s Atlas felt so uninspiring to me.)

We’re not in an era of speeding up old frameworks. We’re in an era of replacing the frameworks entirely. The companies that go back to a blank sheet of paper, identify the fundamental problem these tools are solving, and rebuild the solution from the ground up using the best technology humanity has ever had — those will be the great companies. And I genuinely hope one of those founders is a Jushibal subscriber (who I’ve invested in, haha).

One Last Thought: The Addiction to Quick Wins

I’ll close with something I plan to write about more in a future newsletter. Since 2021, excess liquidity has made “fast exits” and absurd valuation jumps feel like the new normal. The market got addicted to quick money, and a lot of people are chasing that high blindly. (On the bright side, social media and viral marketing make it very easy for founders with flawed thinking to broadcast those flaws publicly — which actually makes VC due diligence easier.)

Here’s what I believe: early success does not guarantee final success. Hollow viral growth is closer to a death sentence than a victory lap. Startups are a brutally long game — 10 years or more of enduring doubt, ridicule, and loneliness while pushing forward. (I feel this more now that I’ve left my own comfortable nest.)

So to the founders reading this: don’t get arrogant from early wins, and don’t get crushed by early failures. Building a startup is a challenge and a process, not a destination. This was never supposed to be easy. What matters is keeping a long-term perspective, staying steady, and moving forward together. I know this road is lonelier and harder than anyone admits. I’m rooting for all of you, and I hope I can be of help along the way.

Thanks for reading, as always.

— Ian

On October 13th, OpenAI announced a 10GW custom chip deal with Broadcom. Broadcom’s stock surged 10% immediately, adding $150B in market cap in a single day.

The last four weeks have been even wilder. September 22nd: $100B investment from Nvidia plus a 10GW system deployment contract. October 6th: 6GW deal with AMD plus a 10% equity warrant (160M shares) — AMD jumped 24% in a day, and OpenAI made serious money just from the deal structure. September 10th: a 5-year, $300B cloud contract with Oracle.

The scale is staggering. Back-of-the-napkin math puts it at $1.5–2 trillion in commitments.

But here’s the thing. OpenAI’s 2025 revenue target is roughly $13B. They lost $5B in 2024 and expect losses to nearly double in 2025. Sam Altman told investors to expect $44B in cumulative losses through 2029.

To restate: a company losing $10B a year, projecting $44B in losses through 2029, signed deals worth nearly 150x its annual revenue in a single month.

The Money Goes Round and Round

Bloomberg called these deals “circular,” and they’re right. Hundreds of billions of dollars follow the arrows in a loop: Nvidia invests in OpenAI → that money flows back to Nvidia through chip purchases → Nvidia’s revenue justifies more investment in OpenAI. OpenAI pays Oracle → Oracle buys Nvidia chips → Nvidia reinvests in OpenAI. OpenAI’s deal with AMD sends money to AMD → AMD’s stock rises → OpenAI profits from the warrant → that money flows back to AMD. Broadcom, CoreWeave, and Lambda show the same pattern.

As Fortune pointed out, this is a circular investment structure where the money just spins inside the ecosystem, continuously inflating everyone’s valuations. It has the distinct odor of a Ponzi scheme.

“So what, good companies invest in each other all the time” — and I partially agree. But beyond Moody’s concern about whether OpenAI can actually pay Oracle, the market has deeper worries:

Dot-com déjà vu: In the early 2000s, Cisco and Nortel provided vendor financing to their own customers. When those customers went bankrupt, only bad debt remained. Networking equipment companies lost 90% of their value.

Accounting games: If Nvidia leases chips instead of selling them, OpenAI avoids depreciation on its P&L — making losses look better — while Nvidia absorbs all the inventory risk and depreciation.

Transparency gap: It’s impossible to know exactly how much revenue is generated by these circular structures.

Why Is He Doing This? The WeWork Playbook.

So why is Sam Altman signing $2T in contracts with a company losing $10B a year? Isn’t he worried about defaulting and going bankrupt?

Watching Sam’s clever — some might say cunning — playbook reminded me of WeWork’s Adam Neumann. And of oil tycoon Paul Getty’s famous line:

“If you owe the bank $100, that’s your problem. If you owe the bank $100 million, that’s the bank’s problem.”

What I see is Sam Altman trying to make OpenAI “too big to fail.” By embedding this deeply into the ecosystem — spinning a web of complex commitments and pulling future obligations into the present — OpenAI’s failure would no longer be OpenAI’s problem. It would be the entire AI market’s problem.

Just like WeWork’s failure would have meant admitting SoftBank’s failure, which is why Masa couldn’t cut his losses. Except OpenAI isn’t tied to just one house — it’s linked to multiple mega-corporations, and it looks deliberately designed so that the cascade effect of OpenAI going down would be catastrophic.

Personally, I think OpenAI hasn’t delivered the results everyone expected, and if the for-profit conversion fails, even bigger problems could emerge. Sam’s playbook is one that only works while OpenAI’s halo is still bright, while market liquidity is overflowing, and while the AI bubble is at peak inflation. The window is narrow. But credit where it’s due — he’s executing it pretty well.

OpenAI Is Underdelivering? How Bad Is It?

Let me be clear: I believe OpenAI is a great company that changed the course of human history. Whether the ending will be equally great remains uncertain — but they’ll do whatever it takes to get there. I covered this in depth a year ago, so today I’ll focus on recent developments.

Technical Leadership Is Eroding

OpenAI’s once-dominant model performance — the primary justification for its valuation — no longer holds. The technical moat has collapsed. As of the October 15, 2025 LMSys Chatbot Arena leaderboard, Google’s Gemini-2.5-Pro and Anthropic’s Claude Sonnet/Opus frequently rank equal to or higher than OpenAI’s GPT-4o and GPT-5. On benchmarks like MMLU, Anthropic’s Opus and Sonnet hold the #1 and #2 spots, beating GPT-5 (high).

This is what convergence looks like — the trend I’ve been writing about since February 2023. Competitors’ rapid catch-up means OpenAI’s once-unique architectural advantages are now widely understood and replicated. The AI frontier isn’t a solo race anymore; it’s a crowded battlefield.

Losing the Enterprise API Market

This technical erosion translates directly into financial consequences. According to Menlo Ventures’ market report from several months ago, Anthropic has overtaken OpenAI in enterprise LLM usage with 32% market share versus OpenAI’s 25%. Two years ago, OpenAI held 50%. That’s a dramatic reversal, driven by developers’ strong preference for Anthropic.

The Sidegrade Era: New Toys Over Better Performance

OpenAI’s rapid-fire consumer product launches aren’t just business expansion — they’re a forced, defensive pivot driven by competitive pressure.

The Sora app (a standalone TikTok-style platform for AI-generated video) signals a major strategic shift from capital-intensive infrastructure competition into consumer social media. Features like “adult chatbots” and “Clinician Mode” similarly target B2C markets.

Apps in GPT follows the same logic. GPTs and Plugins already tried and failed with a similar concept — this time they’re throwing a real punch. Whether they can create enough lock-in to make businesses feel they’ll lose traffic by staying out remains to be seen. I actually think it’s a good approach from the “AI is an interface, not intelligence” perspective.

This consumer pivot looks like a desperate move from a company boxed in on all sides. OpenAI can no longer justify its valuation on enterprise API sales alone — especially after losing the lead to Anthropic. Without a clear technical moat, competing on API access becomes a price war, which is unsustainable given OpenAI’s cash burn. They need a new kind of moat — network effects built on a massive consumer user base — and that’s what’s driving the push into social media, adult content, and app ecosystems.

“So what, companies grow and expand” — sure. But the reason this doesn’t look innocent is Sam’s track record. In August, when Cleo Abram asked him to name a decision where he prioritized the world over winning the AI race, Sam answered: “Well, we haven’t put sexbot avatars in ChatGPT yet.” In a separate Verge interview, he said: “Some companies will go build anime sexbots. We won’t. We’ll keep building useful apps.”

Two months later, he reversed course with “we’re not the morality police” and greenlit adult content to drive revenue. That tells you something about how the company is really doing — and reinforces why the “too big to fail” strategy is so necessary.

But Data Centers (Power) Will Still Be Massively Needed, Right?

Enough OpenAI doom. (Though I’ll note: companies and markets don’t collapse overnight. It takes time. But you need to keep reassessing against new information. Everyone denied model convergence too — it took 2.5 years for consensus to form. End of sermon.) Let’s talk about data centers and infrastructure.

We’ve established that model performance is converging. So what competition ignites next? Exactly: we’ve entered the optimization and cost reduction era.

The Inference Shift

A fundamental shift is underway from training-centric to inference-centric workloads, which could structurally reshape infrastructure demand. Industry consensus, including Gartner and McKinsey forecasts, projects that inference will become the dominant AI workload, with some estimates suggesting 80%+ of AI infrastructure spending will support inference by 2028.

Training requires massive clusters of high-end GPUs like Nvidia’s H100. But inference prioritizes low latency and cost-per-query, and can run on cheaper, more diverse hardware: CPUs, cloud providers’ custom ASICs (Google TPU, AWS Inferentia), and new architectures like Groq’s LPU. This makes cost reduction possible, and the inference market — high demand, lower costs, no single dominant player — will be far more competitive than training.

OpenAI’s strategy of building massive general-purpose GPU data centers is rooted in the 2022–2023 training-centric paradigm. But the market is rapidly shifting to cost-sensitive, specialized-hardware inference. OpenAI risks ending up with bloated, inefficient, and overpriced infrastructure for the dominant future workload. And when new GPUs with dramatically better performance arrive, it all needs replacing — making it an extraordinarily expensive liability.

Competitors building smaller, inference-optimized infrastructure from the start (like Anthropic) could secure significant cost advantages, undercut OpenAI on price, and capture the bulk of the commercial market. OpenAI could end up owning the world’s most expensive white elephant.

The Democratization of Model Training

The plummeting cost of training capable AI models directly refutes the assumption that AI belongs only to giants.

Andrej Karpathy’s “nanochat” project demonstrated that a full ChatGPT-style pipeline can be trained from scratch on a single 8xH100 node in about 4 hours for roughly $100. Scaled up: a model surpassing GPT-2 performance costs about $300; one with basic reasoning and coding abilities (approaching early GPT-3) costs about $1,000.

GPT-3’s training cost was estimated at $4.6M in 2020. Five years later, you can build GPT-3 for 0.2% of that cost — 1/500th. GPT-4’s training cost was estimated at $100M two years ago; current estimates are under $20M — about 1/5th. The deflation in model training costs is accelerating.

7B? No, 7M Parameter Models Performing Like This

Samsung’s 7M parameter (not 7B) Tiny Recursive Model (TRM) is shifting the paradigm from “bigness” to “smart design.” Despite being tens of thousands of times smaller than standard LLMs, TRM uses recursive thinking to outperform LLMs on complex reasoning and puzzle-solving tasks — with total training costs around $500.

As models shrink, they can run on consumer GPUs or low-power on-device chips instead of expensive data center hardware. This accelerates the on-device AI era and maximizes accessibility and cost efficiency.

The point isn’t that small models do everything large LLMs do. It’s that they’re catching up on specific tasks. What happens if you create 1,000 specialized 7M-parameter models and run them locally on a laptop or iPhone?

On-device and model miniaturization — another topic I’ve been insisting on since 2023. You saw it. I called it.

Token Prices Are Collapsing

Even excluding smaller, lower-performance models, the cost of using OpenAI’s flagship models has dropped precipitously. GPT-5 mini received strong reviews for its performance-to-cost ratio (the one thing I praised in my GPT-5 analysis).

Token prices — the cost of using AI — are falling rapidly. This happened while performance was still improving. Now that performance is converging and everyone’s focusing on efficiency, costs could drop even further. This means cost reduction on the provider side too. Combined with the trends above, we’re watching the raw material costs (GPUs) become more efficient in real time.

Dark Fiber Existed, But Dark GPU Doesn’t? “Not Yet.”

The current AI infrastructure buildout has striking parallels to the fiber optic boom that created the “Dark Fiber” crisis and triggered the 2000 dot-com collapse. Silicon Valley has started using the term “Dark GPU” — though bulls dismiss the AI bubble by saying “unlike dot-com, there won’t be dark GPUs.”

But the parallels are more alarming than people want to admit:

1. False premise of infinite growth: Dot-com: Justified unlimited network capacity by assuming internet traffic would double every few months indefinitely. AI: Justifying $2T in data center construction by assuming AI compute demand will grow exponentially forever.

2. Investment mania: Dot-com: WorldCom and Global Crossing spent $500B+ (in 1990s dollars) installing redundant fiber networks worldwide, with vendor financing as common practice. AI: OpenAI and partners have committed to comparable or larger infrastructure spending, with circular investment patterns emerging.

3. Unexpected efficiency shocks: Dot-com: Wavelength division multiplexing (WDM) suddenly increased single fiber data capacity by 100x, making most newly installed fiber unnecessary overnight. AI: Algorithmic innovation, optimization techniques, and the shift to smaller efficient models are delivering massive gains in performance-per-watt and performance-per-dollar, reducing the need for brute-force compute.

4. Inevitable oversupply and price collapse: Dot-com: Market flooded with excess capacity, wholesale data transmission prices collapsed, trillions in dark fiber buried underground. AI: Massive infrastructure buildout combined with radical efficiency gains is setting the stage for GPU oversupply — dark GPU.

5. Financial collapse that follows: Dot-com: Companies that financed fiber buildouts with massive debt (WorldCom, Global Crossing) went bankrupt, wiped out investors, and triggered a broader market crash. AI: Entities funding the AI infrastructure boom (Nvidia, Oracle, Microsoft, and their investors) face the risk of catastrophic write-downs if projected demand doesn’t materialize — potentially triggering a systemic correction across the tech sector.

So What?

I believe AI is a genuinely transformative technology — comparable to the internet, possibly bigger — with the potential to reshape the global economy. The value AI creates over the coming decades will far exceed what’s being invested today.

But I also believe the path to realizing that value will be anything but smooth. The current market structure, centered on OpenAI’s “too big to fail” strategy, displays every classic symptom of a speculative bubble: fantastical promises disconnected from financial reality, circular funding schemes, and massive infrastructure buildouts based on flawed assumptions that ignore destructive efficiency trends.

AI will create more value than all current investments combined — but that value will be concentrated in a few winning bets. The vast majority of today’s investment capital will evaporate into the bubble.

A market correction in AI is, in my view, a scenario with significant probability. Most of today’s investments will fail. These astronomical valuations will crash.

From that perspective, the key to AI investing right now is navigating this nuance: be optimistic about the technology’s long-term impact, but deeply cautious and critical about the current market’s unsustainable financial structure. I keep this in mind constantly and work at it every day. Our goal isn’t to make a quick buck riding the speculative froth of the AI bubble — it’s to invest in the lasting value of the AI revolution and play a part in changing the world.

Thanks for reading, as always.

— Ian

Breaking: The CEO Wrote an Emergency Letter on a Sunday Night?!

Investment bank Jefferies saw its stock plunge roughly 18% over five days after a $715M exposure to bankrupt auto parts supplier First Brands was revealed. CEO Rich Handler and President Brian Friedman issued an emergency public letter on Sunday night: “The losses are absorbable. The market reaction is excessive.”

Unwritten Wall Street rule: When a CEO writes an emergency letter on a weekend, it’s a signal that the crisis is real.

History proves it:

Bear Stearns (2008): “We’re fine” → Bankrupt 2 weeks later.

Lehman Brothers (2008): “Liquidity is sufficient” → Bankrupt 1 month later.

Credit Suisse (2023): “Greensill is a small issue” → Fire-sale acquisition 2 years later.

Jefferies stock is down ~18%. Whether this is an “overreaction” or not — we’ll find out soon enough.

What Happened?

The $715M Black Hole

Jefferies’ subsidiary trade finance hedge fund Point Bonita Capital was exposed to $715M in bankrupt auto parts supplier First Brands. That’s nearly 25% of Point Bonita’s $3B portfolio.

Here’s the structure: Point Bonita purchased invoices that First Brands had issued to retailers like Walmart, AutoZone, and NAPA for products like wipers and oil filters (invoice factoring).

But there was a fatal design flaw:

In theory: Walmart → pays Point Bonita directly. In reality: Walmart → pays First Brands (as “Servicer”) → First Brands forwards to Point Bonita.

First Brands acted as “Servicer,” receiving Walmart’s payments and supposedly passing them through to Point Bonita. In other words, Point Bonita gave First Brands money, and First Brands was supposed to give it back. The fund may never have received a single dollar directly from Walmart. Its lifeline was entirely in the hands of the very borrower it was supposed to be protected from.

On September 15th, that lifeline snapped. First Brands suddenly stopped forwarding retailer payments to Point Bonita. Money that had flowed without a single missed payment for six years — gone overnight.

Double Factoring Allegations: $2.3B Vanished Into Thin Air

A First Brands special investigation team is currently examining the possibility that the same invoices were sold to multiple parties simultaneously (double factoring).

An emergency filing by Raistone’s attorneys is even more shocking:

“According to sworn statements by the debtor’s representative and counsel, up to $2.3B in third-party factoring financing has simply vanished.”

When Raistone’s law firm Orrick asked First Brands’ law firm Weil “where did that enormous amount of money go, and how much is currently recoverable?” — the response was:

“We don’t know... $0.”

The same Walmart invoices were allegedly sold to Jefferies, then to another fund, then to yet another. This is rehypothecation fraud. And on Wall Street, 70% of companies that say “we’ve never done that”... well, you know how that story ends.

DOJ Investigation

Per FT reporting, the U.S. Department of Justice has launched an investigation. This isn’t just a civil bankruptcy — they’re looking at potential criminal fraud.

Who Is Jefferies?

Jefferies Financial Group — not a bulge bracket like Goldman Sachs or Morgan Stanley, but not small either. A mid-tier investment bank with $10.5B in total assets, $8.5B in tangible equity, and $11.5B in cash on the books.

Jefferies had been First Brands’ primary bank for over a decade, present throughout its growth from a small parts company to a major auto components player. The criticism now: the bank got too close to the client and let risk management slip.

Recently, Jefferies expanded its strategic alliance with SMBC (Sumitomo Mitsui), which had planned to increase its stake from 14.5% to 20%. That increase is almost certainly on hold now. The key question is how much damage SMBC absorbs.

A Wall Street Bank Run Has Begun

Institutions Heading for the Exits

As of October 11th, confirmed redemption requests:

BlackRock (world’s largest asset manager) — first redemption request filed.

Texas Treasury Safekeeping Trust — early redemption.

Morgan Stanley — formally initiated redemption procedures on October 11th.

This is a textbook Wall Street bank run.

Damage Spreading: Who Else Got Hit?

Direct exposure: UBS funds: First Brands-related risk represents 30% of assets. Cantor Fitzgerald: renegotiating its UBS O’Connor acquisition deal because of this. Western Alliance: passively entangled via leverage financing provided to Jefferies. Raistone: $631M exposed, half of staff laid off.

Across the Pacific: Norinchukin Bank + Mitsui & Co. joint venture: potential losses up to $1.75B. Allianz and other insurers: preparing for major claims.

How does one auto parts company’s bankruptcy entangle global financial institutions? This is the truth that trade finance’s “it’s safe” mythology has been hiding.

Why Now: “It Smells Like Subprime”

Jim Chanos’s Warning

Jim Chanos — the legendary short seller who called Enron in 2001 — recently warned:

“The current booming private credit market is operating in ways similar to the subprime mortgages that triggered the 2008 global financial crisis.”

What Chanos sees in common between 2008 Subprime and 2025 Private Credit:

Black box structures: Multi-layered structures that hide risk.

Excessive yields: First Brands collateral inventory debt was expected to yield over 50%.

Opacity: Enron at least had public disclosure obligations as a listed company. First Brands is a private company — only hundreds of loan managers with confidentiality agreements could access its financials.

“We rarely get to see how the sausage is made.” — Chanos

This isn’t a simple corporate bankruptcy. It’s the private credit market’s mask coming off — the market everyone thought was easy money, safe money, money-on-money. And the market is in an uproar because this echoes the Greensill Capital and Archegos events that brought down Credit Suisse.

Greensill Capital (2021): $10B Evaporated

Structure: UK fintech Greensill grew rapidly in trade finance post-2008. Companies sell products to big retailers, get invoices, and Greensill buys those invoices at a discount for instant cash. “Safe” — because Walmart is guaranteeing payment, right?

Problem: Same invoices repackaged and sold multiple times. They started selling future invoices that didn’t even exist yet. Credit Suisse had $10B+ invested in Greensill. 2021: Greensill goes bankrupt → Credit Suisse takes $10B loss.

Archegos Capital (2021): $5.5B in Losses

Structure: Bill Hwang’s family office took massive leverage from multiple banks — none of which knew how much the others had lent. Total return swaps hid actual positions. Reality: $200B+ in positions on $10B in capital (20x leverage).

Problem: March 2021, several stocks crash, margin calls hit, Hwang can’t pay → banks panic-sell. Credit Suisse: $5.5B loss. Nomura: $2.9B loss.

Credit Suisse endgame: Greensill ($10B) + Archegos ($5.5B) = total trust collapse → 2023 fire-sale to UBS.

Why First Brands Is Greensill + Archegos 2.0

The terrifying commonalities: multi-layered opaque structures, duplicate collateral sales, hidden leverage, excessive yields masking risk, and cascading counterparty exposure.

The difference? First Brands carries the DNA of both events:

Like Greensill: trade finance + selling the same collateral to multiple buyers.

Like Archegos: opaque structure + hidden debt.

Worst-Case Scenario: Financial Market Seizure

Double factoring confirmed → same invoices sold to 5 funds.

Jefferies criminal indictment + side letter lawsuit explosion.

Stock -60% → Point Bonita redemption cascade.

Forced liquidation → larger losses → trust collapse spreads across Jefferies’ entire $20B+ in managed funds.

UBS, Millennium found to have similar black boxes.

Entire $1.7T private credit market freezes.

LPs demand full redemptions → fund fire sales.

Leveraged loans crash -30% → Raistone, Onset, and other supply chain financiers fail in sequence.

Small and mid-size businesses lose access to working capital → supply chain paralysis.

Jefferies attempts a Credit Suisse-style sale but finds no buyer → bankruptcy → $10B in equity evaporates.

2008 Subprime = one mortgage repackaged into multiple CDOs. 2025 First Brands = one invoice sold to multiple funds.

Same shit, different wrapper.

So What?

Private Credit’s Winter Is Coming. The “Safe Investment” Fantasy Is Collapsing.

Everyone thought: “It’s an invoice guaranteed by Walmart. Default risk is basically zero, right?”

But it turns out the same invoice may have been sold to Fund A, Fund B, and Fund C simultaneously. Walmart pays $1M, but three funds each claim $1M. Who gets paid? Whichever one races to bankruptcy court first.

Point Bonita was registered in the bankruptcy filing as an unsecured creditor with a “contingent, unliquidated, disputed claim.” Translation: “We have no idea if we’re getting our money back.”

This is the real problem with the private credit market. What everyone believed was “safe” was actually a collection of hidden leverage and duplicate collateral.

In 2008, everyone said “subprime is a small problem.” Until Lehman blew up.

If trade finance — supposedly the safest corner of private credit — can blow up like this, what about the rest?

Inside this massive black box that attracted everyone with high yields — how many more time bombs are hiding?

If First Brands is the trigger, we might be witnessing another moment that goes down in financial history.

Thanks for reading, as always.

— Ian

For years, my answer has been: “It’s the ecosystem.” The supply chain. The ODM/OEM infrastructure. The speed to market. (I wrote about that here; K-Beauty Is Not Gangnam Style. It’s BTS)

And that’s true. Korea has the most efficient beauty supply chain on Earth.

But here’s the deeper question I kept avoiding: Why does Korea have that ecosystem in the first place?

Because ecosystems don’t just appear. They’re built by people. And people are shaped by culture.

Last week at Seoul Investor Forum’s SIGNITE Lifestyle Forum, I was on a panel about scaling K-consumer brands globally with Jay Brown — co-founder of Roc Nation and Marcy Venture Partners with Jay-Z.

Someone in the audience asked: “Why is K-Beauty succeeding?”

I started giving my usual answer about the ecosystem. Supply chains, ODM/OEM infrastructure, speed to market.

Then I stopped. And told them what I should’ve been saying all along.

The real answer to “Why did K-Beauty win?” isn’t about factories or supply chains.

It’s about a society that can look you in the eye and say, “You look fat.”

In the U.S., that’s an HR violation. In Korea, it’s just another Tuesday.

The Blunt Truth Behind K-Beauty’s Rise

That blunt, appearance-conscious culture? It’s not just rude. It’s what built the most efficient beauty ecosystem on Earth.

When people constantly give and receive feedback about how they look, awareness goes up. Self-care becomes a habit. Experimentation becomes normal. And innovation follows obsession.

That feedback loop — constant awareness → obsession → iteration — is why Seoul became the world’s plastic surgery capital. It’s why Korean skincare routines feel like rituals to outsiders. It’s why the best beauty trends start in Korea.

This isn’t marketing. This is cultural firmware that’s been running for decades.

In America, my golden rule is: “Don’t comment on anything someone can’t change in 30 seconds.” Hair in your teeth? Fine. Weight gain? Absolutely not.

But in Korea, those conversations are normal. “You look tired.” “Your skin’s breaking out.” “Did you gain weight?” Not malicious — just honest.

And that honesty? That’s the root cause of Korea’s beauty dominance.

Why Korea Can Say “You Look FAT” (And Why It Matters)

I used to just answer “ecosystem” when people asked about K-Beauty’s success. But last week, jogging through Seoul, I went one level deeper: Why does Korea even have this ecosystem?

Five reasons:

1. Concern, not criticism.
In Korea, “You’ve gained weight” often translates to “I care about your health” or “I’m paying attention to you.” It’s relationship maintenance, not an attack.

2. Homogeneity breeds comparison.
Korea is (relatively) ethnically homogeneous. Similar beauty standards mean changes are more visible and more discussable. Everyone’s measuring against the same baseline.

3. Extreme competition.
Small country, limited resources. Everything becomes competition — school, jobs, marriage, even appearance. Your weight isn’t just personal; it’s a signal of self-discipline and competitiveness.

4. Appearance as social capital.
Looking good isn’t vanity in Korea — it’s strategy. Better opportunities come from better impressions. So appearance feedback isn’t invasive; it’s socially relevant information.

5. Collectivism over individualism.
Group harmony matters more than individual privacy. Your appearance affects the group’s image. So commenting on it isn’t rude — it’s communal care.

This isn’t about Korean people being “mean” or “shallow.” It’s about a society where self-improvement is a collective sport, and feedback is the scoreboard.

But that pressure? That constant awareness? It’s toxic. And it’s also what makes Korea the world’s R&D lab for beauty innovation.

The Beauty — and the Burden — of Perfection

K-Beauty’s rise didn’t happen in a vacuum. It came from a country that also ranks among the highest in the OECD for suicide rates, especially among young people.

That statistic isn’t a footnote. It’s part of the same story.

When a society rewards achievement and appearance more than rest or failure, you get two outcomes: unbelievable excellence — and unbearable pressure.

K-Beauty isn’t a feel-good story. It’s what happens when a society turns anxiety into a competitive advantage.

That’s impressive. That’s also deeply fucked up.

The same drive that makes Korea out-innovate everyone in skincare also burns people out faster than anywhere else. The relentless “try harder, look better, never stop” mentality powers an ecosystem that can beat anyone — but at what cost?

If you admire the speed and quality of Korean brands, also recognize the human cost behind them. The next generation’s mission isn’t just to sell better products — it’s to build systems that don’t destroy the people creating them.

So What’s the Thesis?

Here’s what most people miss when they ask “Is K-Beauty sustainable?”: It’s not about whether people keep buying sheet masks. It’s about whether the cultural infrastructure that produces this behavior is stable.

And that infrastructure — the feedback culture, the competition, the collectivism, the 50-year supply chain — has been getting stronger, not weaker.

K-Beauty isn’t a product line. It’s the export of a national operating system for self-improvement. K-Pop and TikTok were just catalysts. The ecosystem was already there, rowing hard. This time, the tide finally came in.

So the question isn’t whether K-Beauty is a fad. It’s: What else can this system produce? Supplements? Functional foods? Longevity?

And here’s where it gets interesting: systems built on anxiety eventually eat themselves.

K-Beauty 1.0 was about perfection at any cost. Now Koreans are asking: What was the cost? Burnout. Mental health crises. A generation saying “I’m tired.”

That’s why wellness is exploding in Korea right now — not surface-level “self-care Sunday” wellness, but the kind that asks whether optimization is worth destroying yourself.

Look at what’s trending: SMCC (Seoul Morning Coffee Club) — morning coffee raves at 6 AM instead of Itaewon parties until 4 AM. Positive Hotel — spaces designed to let you rest, not optimize.

If Korea can export the pressure, maybe it can also export the antidote.

The infrastructure is there. The cultural literacy is there. The innovation engine is there. Now it just needs to point in a different direction.

That’s the K-Beauty I’m excited about — not the one that perfected anxiety as a business model, but the one that’s learning to live without it.

Thanks for reading,
Ian

(Rumor has it this whole essay started because my mom told me I looked fat during Korean Thanksgiving. So… thanks, Mom.)

Founders building in K-Wellness, K-Mental Health, K-Longevity: ian@ianpark.vc Especially if you’re working on the antidote.

“AI killed junior programmers.”

The narrative goes like this: since AI launched, junior developer hiring has collapsed, which proves AI is already replacing entry-level workers.

The recent Stanford Digital Economy Lab paper using ADP payroll data fits this narrative like a glove. The graphs show AI-exposed occupations seeing a sharp decline in hiring for 22–25 year olds, while mid-career and above remained stable.

It reads like a clean, satisfying confirmation of everyone’s worst fears. The timing is almost too perfect. And that’s exactly what made me suspicious — this felt like a paper designed to tell people what they already wanted to hear, optimized for attention rather than rigor. Predictably, the media ran with the working paper without any verification, chasing clicks. Self-proclaimed “AI experts” flooded feeds with grave-faced posts about the future of humanity.

So I read the whole thing carefully. My conclusion: it’s an interesting but deeply flawed working paper that hasn’t even been peer-reviewed yet, with significant gaps in data and logic. Here’s my breakdown.

Full disclosure — I wrote a similar prediction in a previous newsletter, so I could easily just say “You saw it. I called it. Stanford agrees.” But wrong is wrong, and my personality won’t let me coast on a convenient narrative. So: “Long-term, yes, this is coming. But not yet. Can everyone please calm down?”

It’s the Economy, Stupid

Before we dissect the Stanford paper, let’s look at the actual U.S. labor market numbers.

According to the Bureau of Labor Statistics:

2023 non-farm payrolls: +251,000/month average.

2024 initially reported: +147,000/month average.

2024 actual (after benchmark revision, April 2024–March 2025): +71,000/month — less than half of what was reported.

August 2025: just +22,000.

June was revised to -13,000.

The U.S. labor market has been far worse than anyone realized. This isn’t an “AI replacing juniors” pattern — it’s a textbook economic slowdown.

And here’s the thing everyone in economics already knows: downturns always hit the weakest link first — new graduates and junior workers. What we’re seeing in junior hiring isn’t AI suddenly devouring human jobs. It’s the historically recurring pattern of economic cycles, where junior workers get squeezed out first. AI is certainly an interesting variable, but juniors getting hit hardest in a downturn is literally in the econ textbook.

The Stanford-ADP Paper: Interesting, but Wildly Overinterpreted

Let me be specific about what this working paper shows:

AI-exposed occupations saw –6% decline in 22–25 year old employment, with some roles down –20%.

The same occupations saw +6–9% increases for workers 30 and above.

On the surface, it perfectly supports the “AI replaced juniors” narrative. But several critical questions emerge.

ADP Data Representativeness

About 1 in 6 U.S. workers get paid through ADP — a decent sample, but not overwhelming. The bigger issue: ADP’s client base skews heavily toward small and mid-size businesses (at least 65%). This means the employment trends captured in ADP data are structurally more sensitive to SMB economic cycles.

During periods of rising interest rates and tightening capital, mid-size companies are the first to freeze new hiring and cut junior/software/customer-facing roles. Meanwhile, big tech and large enterprises — the ones actually adopting AI most aggressively — typically use their own HR systems and don’t show up in ADP data at all.

See the problem? The companies most likely to replace workers with AI are barely represented in this dataset. The companies in this dataset are the ones most likely to cut juniors for plain old economic reasons.

Oversimplified Categories

Lumping software developers and customer service reps into the same “AI-exposed” category is a serious oversimplification. Even Klarna — the poster child for AI replacing customer service — had to rehire human agents after customers revolted over plummeting service quality.

Timing Doesn’t Add Up

The paper claims junior hiring started declining from late 2022, attributing it to AI. But think about the timeline: are we really supposed to believe that mid-size companies adopted ChatGPT the moment it launched and immediately started cutting junior headcount? Let’s all collectively remember what we were actually doing with AI in 2023. It was a novelty, not an enterprise deployment.

Other Gaps

ADP data shows aggregate employment trends but can’t distinguish between hiring freezes and layoffs. Labor economics has long established that companies in downturns freeze hiring before they fire — meaning the junior decline likely represents closed doors, not replaced workers.

Most critically: wages haven’t dropped significantly. If AI were genuinely replacing workers at scale, the first observable effect would be wage compression in affected roles — supply of replaceable labor exceeds demand, pushing wages down. That hasn’t happened. This alone makes the “economic slowdown” explanation far more persuasive than the “AI takeover” explanation.

The working paper found correlation. It did not prove causation.

Companies Haven’t Even Figured Out AI Adoption Yet

Let’s look at where enterprises actually stand:

BCG (2024): 74% of companies globally have failed to generate real value from AI or scale it successfully.

MIT NANDA (2025): 95% of GenAI pilots showed no ROI. (Though I’ll admit this report also feels somewhat exaggerated.)

US Census (2025): Large enterprise AI adoption actually dropped from 14% to 12%.

AI isn’t in the replacement phase. It’s still stumbling through the pilot phase, burning money along the way.

And mid-size companies — the ones dominating ADP’s data — are even further behind:

RSM Middle Market AI Survey (2025): 91% of mid-size firms say they use generative AI, but only 25% have fully integrated it into operations. Knowing how companies love to overstate adoption in surveys, if this is where things stand in 2025, we’re nowhere near junior replacement levels.

TXI Report (2025): Mid-size companies have big AI ambitions but remain stuck in exploration/pilot phases due to data infrastructure gaps and talent shortages.

OECD Report (2025): Large enterprises adopt multiple AI technologies simultaneously; mid-size firms are limited to single-technology, narrow implementations.

Andersen Institute (2025): AI adoption is most active at startups and large enterprises. Mid-size firms lag due to change management and cost barriers.

The conclusion: mid-size companies are active in AI pilots but structurally delayed in enterprise-wide execution. This further supports the interpretation that junior hiring declines in ADP data reflect economic slowdown and post-pandemic hiring corrections — not AI replacement.

So What?

“AI killed junior programmers” is a provocative headline and excellent engagement bait — I’ll give it that. But what’s actually happening is the impact of economic slowdown, the hangover from COVID-era over-hiring, and companies using AI as a convenient excuse for margin management.

The data shows a classic labor market weakening pattern consistent with recession.

The software industry massively over-hired during the pandemic, and we’re watching the correction.

Mid-size companies — let alone large enterprises — haven’t successfully deployed AI yet, and ROI is essentially zero.

ADP data is structurally skewed toward SMBs, making it dangerous to generalize across the entire tech/industry landscape.

The Stanford-ADP study found correlation but failed to prove causation.

This working paper is getting massive media coverage, but in terms of academic rigor, it’s still at the “preliminary interpretation” stage. It’s not from a Stanford core department, and it hasn’t been peer-reviewed. The data may have captured an interesting signal, but jumping straight to “AI killed juniors” is dangerous overinterpretation.

That said — AI has genuinely boosted productivity. I feel it every day in my own work. But it hasn’t reached the level of replacing juniors yet. Companies are struggling to even implement AI properly, hemorrhaging money in pilot phases. The current trend is companies using AI as a scapegoat to cover for recession-driven hiring freezes and failed hiring strategies.

Crisis Is Opportunity

None of this means juniors aren’t facing a crisis. They are — and I believe the full impact of AI on junior roles is coming eventually. But the important thing to recognize is that this crisis could also become the soil and fuel for building the next generation of great companies.

As I wrote before, integrating AI into existing systems structurally favors senior experts — they’re the ones with the experience and knowledge to verify and leverage AI’s outputs. But simultaneously, this is the best era in history to create something entirely new.

Rather than fitting AI into existing systems, the bigger opportunity for juniors — who lack deep experience in legacy frameworks — is building AI-native businesses and services from scratch. Their lack of institutional baggage is actually an advantage: through first principles thinking, they can generate fresher, bolder ideas that overturn existing paradigms and create entirely new markets. AI has opened that door.

AI is simultaneously a massive threat and an unprecedented opportunity for juniors. The key is not fearing the change but leaning into it aggressively, exploring new possibilities. It’s cliché, but “crisis is opportunity” fits perfectly right now.

To the talented juniors reading this: don’t let this moment pass you by. If you’re ready, or even just curious, reach out anytime at ian@ianpark.vc. (Fair warning — I’m building from the ground up myself these days, so replies might be slow. Bear with me.)

Thanks for reading, as always.

— Ian

This newsletter has been translated from Korean to English.

K-Pop Demon Hunters: California Rolls, Orange Chicken, and the Future of Korean Culture

Hello, it’s Ian! This week I want to share some thoughts on K-Pop Demon Hunters (KDH) — a film you’ve probably at least heard of, if not already watched on Netflix.

1. What exactly is KDH?

K-Pop Demon Hunters follows Huntr/x, a global K-pop girl group who also happen to be demon hunters. They guard the barrier (“Honmun”) between the human and demon worlds with the power of music, all while battling their rivals, the secretly demonic boy group “Lion Boyz.”

Directed by Korean-American Maggie Kang and infused with Korean folklore (grim reapers, talchum masks, etc.), it was produced by Sony Pictures Animation and released globally on Netflix. The result? Record-breaking success: the most-watched Netflix movie of all time, with over 230M views in just two months. The OST “Golden” hit #1 on the Billboard Hot 100, and four songs from the soundtrack charted simultaneously.

2. Who actually made it happen?

Here’s the twist: this mega-hit wasn’t led by “traditional” Korean creators. The director was Maggie Kang, the voice cast included Arden Cho and Ji-young Yoo — largely Korean-American talent. Sony produced, Netflix financed with $100M, and Sony walked away with $25M upfront. In other words, Netflix made the real money here.

The takeaway: this wasn’t “purely Korean.” It was Korean DNA reassembled by diaspora talent and global capital — and precisely because of that, it exploded.

3. Orange Chicken & California Roll Moments

This reminded me of two icons of Asian-American food culture:

• Orange Chicken (Panda Express): The #1 Asian food chain in America built its empire not on traditional Chinese dishes, but on orange chicken — a sweet, crispy, Americanized invention.

• California Roll: Sushi’s global breakout star didn’t have raw fish. It had avocado, crab stick, and mayo. A hybrid at best, blasphemy at worst — but it’s what carried sushi to global popularity.

KDH is standing right here. On the surface, global fans say they want “authentic Korea.” In reality, they’re consuming the California Roll version: palatable, remixed, localized. Authenticity isn’t the point anymore. If it tastes good, it wins.

So what?

If you want to win globally, you need to understand both Korea and the U.S. You can’t just say “we’re killing it in Korea.” You need to prove you can remix Korea’s strengths into a form that conquers the global stage.

There are two paths:

• Stay pure and differentiate with tradition (the 청국장 and chicken feet route).

• Go full localization and build your own Orange Chicken.

Both have merit. But you can’t do both half-heartedly. Running Korea and U.S. businesses simultaneously often means you’re really running two businesses. From my experience and from what I hear from other Korean VCs in the U.S., it’s far more effective to go Day 1 in America if you want scale.

The risk: others will beat us at our own game

We’re already seeing non-Korean founders succeed in K-beauty by borrowing Korean playbooks. The era of “only Koreans can use Korean culture” is ending. If we don’t master and weaponize our own cultural capital, someone else will — and maybe do it better.

The opportunity: a generation raised on KDH

Think about it: kids who grow up on KDH won’t see Korea as exotic. Just normal. Like how Disney shaped American childhoods, or Pokémon shaped global pop culture. For them, Korea isn’t “special” anymore — it’s just part of life.

That’s massive. It means the next 10–20 years aren’t just about selling Korean culture. They’re about building global brands and industries on top of Korean culture, while the foundation is already normalized.

Closing

We’re living through a once-in-a-lifetime alignment: AI on one side, Korea going global on the other. That’s rocket fuel for founders and investors who can think big.

So I’ll end with this:
Will you build California Rolls and Orange Chicken, or will you insist on 청국장 and chicken feet?

If you’re making (or want to make) the California Roll version of your product — and you want to take it global — email me. Let’s conquer the world together.

Thanks for reading!

— Ian

This newsletter has been translated from Korean to English.

AI: The Sharpened Spear. Vibe Coding: The Cracked Shield.

You've probably heard it ad nauseam, but the startup ecosystem is currently being swept up by vibe coding—a development method where AI handles everything from writing to deploying code. This allows a small team to launch a service at the speed of a large corporation, which is why the once-unrealistic idea of a "one-person unicorn" continues to be a topic of conversation.

However, while this trend is a tremendous innovation, if you think a little deeper and look beneath the surface, you'll see a time bomb that is growing and will soon explode.

The Shield Gets Thinner, The Spear Gets Sharper.

The Shield Gets Thinner Because of AI

As mentioned, the currently popular vibe coding has a fatal flaw: many who engage in it don't even know the basics of coding and security. By relying solely on vibe coding, they are bound to create products with poor security.

• In 2021, a research team from MIT and NYU discovered security vulnerabilities in over 40% of the code automatically generated by GitHub Copilot. Even with GitHub's new security tools, an analysis presented at the 2025 Silicon Valley Cybersecurity Conference (SVCC) revealed that over 30% of the code still had at least one vulnerability.

• In March 2025, GitGuardian noted that the rate of secret exposure in repositories using Copilot was 6.4%, much higher than the 4.6% in typical open-source projects.

• According to the Veracode 2025 GenAI Code Security Report in August 2025, an analysis of over 100 code generation tools that use large language models showed that 45% of all tests included code with security vulnerabilities. The ratio of vulnerabilities was particularly high in Java code, with serious security weaknesses like XSS (13.5%) and log injection (12%) appearing frequently.

• An Apiiro analysis in September 2025 pointed out that since the introduction of AI code, the number of new security flaws has skyrocketed by over tenfold from the end of 2024 to mid-2025, with more than 10,000 new vulnerabilities reported every month. Privilege escalation paths increased by 322% and architectural design flaws by 153%, confirming that structural security issues are rapidly proliferating within the code produced by AI.

In short, AI code assistants are increasing productivity while at the same time exponentially magnifying corporate security risks. The flashy facade of vibe coding, which "AI founders" promote as a way to enhance productivity and democratize coding, is simultaneously creating the shadow of a cybersecurity catastrophe that is spreading uncontrollably throughout our product ecosystem.

The Spear Gets Stronger Because of AI

But the bigger problem is that while our shields are weakening, hackers have been handed a more powerful weapon. Generative AI has given developers and startups speed and productivity, but it has given the exact same benefits to hackers. In the past, hackers who lacked the technical skills to properly plan an attack can now easily create sophisticated code and automated attack scenarios with large language models like ChatGPT. For attackers, AI is no longer just a "support tool"; it has become a core weapon that is changing hacking itself.

• In a real-world phishing experiment, an IBM research team revealed a shocking result. The click-through rate for AI-written phishing emails was 11%, which was virtually no different from the 14% for phishing emails written by a human. The difference was starkest in production speed: AI completed an attack email in just 5 minutes, which would have taken a human an average of 16 hours. This has enabled hackers to carry out more attacks in less time.

• Even more shocking data came from the Unit 42 research team at Palo Alto Networks in 2025. They developed an Autonomous Attack AI (Agentic AI) and conducted a simulation, with results that were beyond imagination. The AI agent completed the entire process—from infiltration and privilege escalation to data exfiltration—all by itself in just 25 minutes. The data breach cycle, which used to take an average of 9 days, now takes just a few hours. According to the report, the average time to detect and respond to a breach in 2021 was over 200 hours, but by 2024, the time it took for an attacker to steal the data they wanted was shortened to just a few hours. This ability to execute various attacks at high speed has created a gap that defenders currently cannot close.

And The Cost of the Spear Has Gotten Cheaper

AI doesn't just improve the quality of attacks; it has also broken down the economic barriers that hackers faced. In the past, with limited labor and time, hackers could only target "big fish" like large corporations or financial institutions. They would only move if the potential monetary gain was high enough, a positive ROI.

But the situation has changed. Thanks to AI, attacks can be replicated at almost zero cost and executed in parallel. Automated phishing campaigns, indiscriminate malware distribution, and even tailored social engineering attacks are no longer driven by human labor but by replicable AI resources. The ROI equation has been flipped, and cheap, fast attacks now target not only large corporations but all companies, including small and medium-sized businesses.

The bigger problem here is that small companies, including startups, are often poorly prepared for cybersecurity. Since hackers didn't target them in the past due to low ROI, these small companies didn't particularly focus on security. Startups, for a similar reason, have thin security, and now, by rapidly building code with vibe coding, they have become perfect prey for hackers. It's no longer a matter of "if" they will get hacked, but a matter of "when."

Recent real-world examples clearly show that startups relying on vibe coding are highly vulnerable to hacking. For instance, WIRED reported that an AI attack agent called RunSybil hacked a Claude Code-based website in just 10 minutes. Lovable, once hailed as Europe's fastest-growing vibe coding startup, was noted for its innovation in allowing customers to create apps with natural language, but it was later revealed to have neglected severe security vulnerabilities for several months until a competitor reported it. Furthermore, the popular "one-person startup" AI platform Base44 was found to have an authentication bypass vulnerability that allowed malicious users to access corporate apps, which was only patched after the Wiz research team pointed it out.

When startups blindly pursue speed and productivity by adopting vibe coding, security often takes a back seat. As a result, the ingredients for a cybersecurity "perfect storm" are steadily accumulating.

What Should the Shields Do?

So, how can we strengthen our shields again? First, we need a new infrastructure. It's not just about creating code quickly, but about having tools—like a hypothetical "VibeCheck Cyber"—that can check in real-time where the code came from, if it was AI-generated, and what vulnerabilities are hidden within it. Speed alone is not enough; a verification system must support productivity.

Second, we must not blindly trust vibe coding. Code automation is a powerful tool, but it's not the complete solution. What's ultimately important is working with people who understand the overall principles of a program and the fundamentals of security. The moment a startup trusts AI code without a deep understanding of code quality, design principles, and security architecture, it becomes a perfect target for hackers. A single security incident, such as the photo and personal information leak at the startup "tea" in August 2025, can end a startup in an instant.

And finally, the last line of defense is cyber insurance. In the U.S., cybercrime losses hit a record high of $16.6 billion in 2024, a 33% increase from the previous year. 60% of small and medium-sized businesses without insurance went out of business within six months of being hacked, while those with insurance were able to survive by recovering significant portions of ransom payments, legal settlement fees, and business losses. Now, cyber insurance is no longer an option; it is essential infrastructure for survival.

So, What Now?

Disease and treatment are an eternal battle of spear and shield. We have conquered many diseases, but others remain unconquered, and new diseases like COVID continue to emerge. That's why we have medical insurance.

Cars and accidents are slightly different but a similar battle of spear and shield. While we try to reduce the accident rate with various safety features, we can't completely eliminate car accidents. That's why we have car insurance.

Cybersecurity is no different. AI continues to sharpen the hacker's spear, and vibe coding thins our shield, but cybersecurity solutions continue to thicken our shield. Ultimately, cybersecurity is also an endless battle of spear and shield.

Therefore, given the recent advancements in AI-related hacking methods and the emergence of new problems like vibe coding, I believe the time is not far off when cyber insurance, like car and medical insurance, will become essential infrastructure for a company's survival. I wanted to share my thoughts in this newsletter because I think it's important for smart entrepreneurs and investors to remember this trend.

In fact, I had this thesis as soon as ChatGPT appeared (hacking becomes easier and cheaper = a cybersecurity nightmare). In 2023, I was interested in an American company called Resilience, which received joint investment from Peter Thiel's Founders Fund, Lightspeed, General Catalyst, CRV, and the large Canadian insurance company Intact. Now that two years have passed and it's 2025, I don't know the company's specific situation, but I feel that this company is even more perfectly aligned with the market trend now than it was then.

I usually do my deal sourcing outbound rather than inbound. I look at overall tech trends, get inspiration for my theses, and then focus on finding the companies needed for the future I imagine based on those theses. For example, the idea of a Personal CRM that came to mind after seeing a note-taking app is in a similar context. I plan to continue incorporating these emerging startup ideas into my newsletters.

If you are an entrepreneur building a startup that fits the future I imagine, please feel free to email me at ian@ianpark.vc. Thank you for reading today.

Sincerely,

Ian

What Even Is a “One-Person Unicorn”?

The narrative goes like this: one person, armed with AI, builds a $1B company. On the surface, it sounds seductive. AI handles coding, QA, customer support, sales, copywriting — costs stay minimal, productivity looks explosive.

One of the loudest champions of this idea is Sam Altman, who’s mentioned it multiple times, filling people with dreams and hope. But think about his incentives for two seconds:

More people trying → more users flowing into the AI ecosystem. More awe and fear around AI. More demand for OpenAI’s API and compute.

This story simultaneously inspires founders and drives revenue growth for platform and model companies. It’s brilliant marketing.

I actually talked about this a year ago — consider this the extended director’s cut. My position hasn’t changed: this trend is a temporary phenomenon that appears at the dawn of every technology shift. Here’s why.

Why One-Person Unicorns Are Impossible

Competition: Your Margin Is My Opportunity

The essence of free markets and capitalism is competition. High margins are market signals, and market signals attract new entrants. The moment a solo operator generates outsized returns, startups, big tech, and global capital come running. Software has near-zero replication costs, and in the AI era, feature copying is faster than ever. When everyone in the world has access to the same tools, any one-person business that reaches unicorn-level growth will immediately face superior teams building better products. Growth stalls. It’s not a question of if — it’s when.

Rising Customer Expectations

We’re at the very beginning of the AI era, so simple wrappers and clones can still get traction as solo ventures. But as time passes, every developer will be using AI to build faster and better, and customer expectations will rise dramatically. The market will demand products that leverage AI’s strengths in fundamentally new paradigms — not just clones of existing tools. These products are technically difficult and require sustained collaboration from multiple people over long periods. The era of “you can’t build this alone” is coming back.

Operations: The Sustainability Problem

Operational risk isn’t about “does it work?” — it’s about “does it work safely, all the time?”

Vibe coding: Prompt-based development is convenient but creates critical vulnerabilities like arbitrary code execution and memory corruption. Wiz and Databricks have issued repeated warnings.

Replit AI: An agent hallucinated and deleted a production database. A symbolic incident exposing the absence of guardrails and permission controls at the design level.

Tea app: A fast-growing vibe-coded social app leaked tens of thousands of ID photos, selfies, and DM images, exposing over a million messages. Trust was destroyed instantly.

The lesson is universal: no matter how impressive the features, weak operational and security foundations kill trust and sustainability.

Durable moats are built on data, network effects, regulatory expertise, trust, partnerships, ecosystems, distribution, and brand. These require long timelines, significant capital, and team effort. For B2B and enterprise products generating stable revenue, regulatory and policy capabilities are non-negotiable. One person cannot sustain these demands long-term. The Tea and Replit incidents prove that operational risk isn’t a feature problem — it’s a trust and moat problem. And that’s where the one-person unicorn model structurally fails.

Who Actually Benefits?

The narrative persists because specific groups profit from selling it:

Model and cloud companies: More individuals flooding in means explosive API and compute revenue.

Platforms and tool providers: Hosting, runtime, auto-deployment, marketplace fees — steady income from the army of solo builders.

The course industry: Package a handful of extreme outlier success stories from overseas, sell courses, newsletters, and workshops.

Obviously, Sam Altman and other AI company CEOs are the biggest beneficiaries, which is why this narrative keeps getting amplified.

The course industry is where it gets ugly. I think of it as the digital chicken shop factory. In Korea, chicken restaurants pop up in every neighborhood and most go under quickly. Same thing here: course graduates flood the market with vibe-coded products that have similar ideas, similar quality, and similar marketing. They get the title of “founder,” but reality hits fast — they can’t survive competition and shut down. Even if they make money temporarily, a faster and better team takes that market, forcing them to endlessly chase the next thing. Digital daeyang castella — the trendy cake shops that exploded and collapsed overnight.

The bigger problem is the message these course sellers push. “You can be a one-person unicorn too” sounds sweet, but it’s fundamentally no different from crypto trading groups promising “make millions in months.” They plant delusional expectations, and the people who paid for the course end up as the biggest victims.

The key questions are obvious:

Has the instructor actually built a one-person unicorn?

Do they transparently share their own product metrics — revenue, retention, cohort data?

Or is the course itself their only product?

The answers are predictable. If so, this isn’t entrepreneurship education — it’s a multi-level structure that commodifies hope. The people at the top make money. Everyone below them ends up a mark. What this narrative mass-produces isn’t unicorns — it’s digital chicken shops with no shelf life.

The Real Future: Small, Elite Teams Built to Scale

The actual winners of the future are clear to me:

Start as an elite micro-team optimized for speed, with scaling as the explicit goal.

Expand into a structured, VC-backed company when the time is right.

Use AI not as a crutch, but as a fundamental force multiplier.

AI shortens the ladder, but it also makes the climb steeper. Entry got easier, but the number of teams that can actually reach the top is shrinking. The next Google will look like the current Google — not a one-person unicorn, but a company where AI runs on top of real scale and fundamentals. Definitely not through a $500 “AI Solopreneur Secret Masterclass.”

So what?

Predicting Trends Is Survival

Amazon dropshipping → crypto → metaverse → AI. Chasing trends is structurally disadvantaged. I’ve seen too many founders who reinvent themselves as “experts” with every hype cycle, and I remember all of them. (I actually follow and use them as negative reference checks...)

Fundraising is not success. Exit, cash flow, and DPI are success.

The teams winning right now are the ones who started preparing long before this moment.

This is what makes startups genuinely hard. You have to understand trends early, predict the future, hold a vision, absorb doubt and ridicule from everyone around you, and push through alone. That’s the founder’s path, and the startup’s fate.

The bottom line: don’t chase what’s already here — anticipate what’s next and move early. Right now, we need to stop using AI to speed up old things and start building things only AI can make possible. When the iPhone first launched, the default was porting PC websites to mobile. Most AI startups today look like that.

But we’re entering the native app era — where products are built from the ground up to leverage everything AI can do, in paradigms we haven’t seen before. That’s where one-person operations can’t compete, and where traditionally strong teams with real talent become necessary again. Better tools make easy problems easier, but when everyone scores well, the test gets harder.

One more thing: some AI companies right now are getting massive attention for rapid revenue growth or viral marketing. Impressive, sure. But if they’re structurally dependent on other models or infrastructure companies and running major operating losses — I don’t care how high the valuation is or how much revenue they’re generating. It could all be a house of cards. (Windsurf, anyone?)

I saw a great analogy recently: “Current AI companies are like chickens being shot into the sky with cannons.” My interpretation: the chickens go up high on AI-powered virality, but they can’t fly, so they gradually fall back to earth while new chickens get launched above them. Of course, if one of them turns out to be an actual bird — like Cursor — it takes flight. (Though even Cursor eventually needs to solve its margin problem.)

The Fundamentals Don’t Change in the AI Era

AI is leverage, not a shortcut. “One-person unicorn” stimulates imagination and plants dreams, but it collides with the reality of competition, moats, and operational weight. When the bubble deflates, we’ll see who was actually swimming versus who was naked.

What’s sad is that content about consistently building wealth and generating real insight always gets less attention than clickbait promising “millions in months” or “solo millions.” People are still soaking in dopamine from the liquidity party and paralyzed by FOMO. It’s genuinely heartbreaking. The courses and trends that exploit this psychology — selling dreams the instructors themselves have never achieved, using cherry-picked foreign success stories — I don’t see what positive contribution they make to our economy or ecosystem. They inflate expectations and create bubbles in the startup world.

Consumer value as a fundamental doesn’t change. Competition in a free market capitalist economy doesn’t change. You might get lucky and make money on a short-term shift, but that’s not building a world-changing company. What matters is not chasing illusions, but building on fundamentals and trust — using AI to create long-term consumer value with a real team and a real product.

If you’re building that kind of company, reach out anytime: ian@ianpark.vc

Thanks for reading, as always.

— Ian

Anticipated Comments Section

“You’re a VC, so you’re just scared people won’t need fundraising.” VCs want founders who use capital for explosive growth to change the world — not people building digital chicken shops that generate modest income. Solopreneurs are, by definition, not building the generational rocket-ship startups I’m looking for.

“Didn’t Base44 sell for $80M as a one-person company?” No. They had 8 employees. And $25M of that $80M went to retention bonuses for the team.

“Selling courses isn’t illegal.” Of course not. But some instructors present wildly unrealistic dreams as achievable reality, and I’m trying to provide balance.

“I teach a solopreneur course and you’re being rude.” I’ll counter with an offer: come on my YouTube channel for a 60-minute debate. Let’s go.

So, how will I survive as a VC?

What does it actually take to win in venture capital? Having seen the industry from both sides, during my time at KIC, one of the world's largest LPs, where I helped decide which VC funds to back. And now and prior to KIC, from my seat here in Silicon Valley as a VC.

After meeting with and analyzing over 500 VCs, I’ve started to recognize a few distinct archetypes—the different 'playbooks' that top investors use to get an edge.

Understanding the US Venture Capital Game

First, let me quickly break down how I see the VC business. For a primer on the basic VC functions—source, pick, win, support, and exit—you can check out my previous post.

A Market Where the Seller Picks the Buyer

Venture capital is fundamentally different from public market investing. First, because it’s private, information and access are extremely limited. In direct opposition to public stocks: (1) you can't invest, no matter how much you want to, unless the founder chooses you, (2) luck and timing play a huge role, and (3) information is not distributed fairly among all investors.

This means you’re dealing with an asset class that is high-risk, reliant on limited information, and has low accessibility. I believe it's an arena where individual capabilities are paramount. A successful venture fund, in my view, is one that builds an incentive and decision-making structure that empowers exceptional individuals to work harder and smarter. I’ve been thinking a lot about what an optimal VC fund structure looks like and plan to write about it soon.

Because access is limited and you have to be chosen by the seller (the founder), I've never met a successful VC who lacked personal charisma. I’m not just talking about looks. Every single one of them was someone I wanted to spend time with, someone whose network I wanted to be a part of, someone I were eager to work with.

It's a constant reminder that you have to be that magnetic—that capable of demonstrating your value to a founder—to get into the best deals. And frankly, you need that same energy to convince LPs to give you their money. A survey on how founders choose VCs confirms this, with Personal Chemistry ranking at the very top.

As you can see on the left, the most important factor for founders when choosing a VC is Personal Chemistry.

Basically, It's a Sales Job

From that perspective, a VC is not nearly as glamorous as it looks. You're not just elegantly sitting in an office, reviewing inbound deals like in the movies. You are a salesperson, constantly peacocking to two sets of stakeholders: founders and fund investors (LPs). You're also an entrepreneur running your own business in a ruthless market, in a profession where pivoting to another industry after failure isn't easy.

And the higher you climb, the less time you spend investing and the more time you spend fundraising. There's a running joke in the industry for a reason: "Being a VC is 70% fundraising, 20% fund management, and 10% investing."

Like Other Investors, It's an Endless Grind

Furthermore, since the job is not about simple pattern-matching but about finding true outliers, seniority and experience don't guarantee success. In fact, you have to constantly worry about becoming old and missing new trends, or getting stubborn from past wins. It’s a career that demands constant learning and effort, and the workload seems to increase over time.

You can only be happy if you genuinely enjoy the grind of learning, fundraising, and persuading founders. If you're in it just for the money or the prestige, the risk, stress, and exhaustion will be overwhelming.

I've gone on a bit of a tangent because I see so many romanticized ideas about the VC analyst role. Anyway, let's get back to the main topic. Here’s my fun, unofficial breakdown of the playbooks I’ve seen successful VCs use in this unique asset class.

The Playbooks of Successful VCs

1. The Operator

Operators are typically former founders who provide practical advice to a new generation of entrepreneurs, drawing from their own hard-won experience.

Strength

They offer actionable solutions to immediate problems and can genuinely empathize with the immense psychological challenge of building a company. This is a very common archetype in Silicon Valley. Armed with advice and empathy, they have a high probability of demonstrating good judgment in sourcing, picking, winning, supporting, and exiting. Statistics also show that founder-VCs often generate better returns.

Weakness

Their startup experience is in the past. A success formula from yesterday might not work today, and past glory can sometimes lead to stubbornness. While the big picture may be similar, times change, and every company's situation is different. They can also sometimes "forcefully provide" way more advice than the founder actually wants.

Upgraded Version: The Unicorn Builder

Of course, all founder experiences are valuable, but those who have gone through multiple funding rounds and built a unicorn are naturally more sought after. Among them, those who have managed a successful exit via an IPO or a major M&A are at the top of the food chain.

• Examples: Marc Andreessen (a16z), Ben Horowitz (a16z), Reid Hoffman (Greylock)

Andreessen and Horowitz, the founders who help founders with massive resources. Hoffman, the founder of LinkedIn and the ultimate connector.

2. The Networker

This is the player who hits the ground running, attending every networking event, knowing everyone in the industry, and having the widest coverage. It might seem like the easiest path, but it requires an innate social talent—a playbook for the true extroverts.

Strength

They dominate sourcing by getting there first. They win deals the same way. Their support comes from sharing information others don't have and connecting their portfolio companies with the right people.

Weakness

While not always the case, their picking and exiting abilities can be relatively weaker. They can leverage market gossip and references to make good picks, but they are also more susceptible to getting swept up in market-wide FOMO.

Upgraded Version: The Matchmaker

This VC excels at creating the perfect connection. In Silicon Valley, matchmaking is a critical VC skill. It's about knowing who to connect with whom for maximum synergy, how to frame the introduction, and how to manage the follow-up to ensure it's a success for both parties.

This may sound simple, but it requires a natural instinct and quick thinking. As I've said many times, a VC's most scarce resource is time. Making an introduction is asking for someone else's time, which is like taking on a small debt. The recipient will judge you based on the quality of that connection and decide whether to accept future intros from you.

If you make connections that don't benefit both parties, people will start avoiding your intros. If you make great ones, your relationships will strengthen. Once you get a reputation as "the person whose intros are a waste of time," your life as a VC gets very difficult. That's why I believe introductions must be strategic, thoughtful, and executed with care.

• Example: Every VC. It's the most fundamental part of the job, and yet, perhaps the most difficult.

3. The Content Creator

This playbook involves producing content—memes, newsletters, podcasts—to source deals. Often, it takes the form of writing a newsletter or hosting a podcast to share one's thoughts, like I do.

Strength

The biggest advantage is that you can do the work of a Networker to some degree, but from the comfort of your own desk. While the depth of personal relationships is shallower, the potential reach can be even wider. It also allows for broader personal branding and relatively rapid growth. My own newsletter is a good example; it allowed me to quickly build a network in Korea, a community I barely knew since I live in Silicon Valley.

Weakness

As someone who has used this playbook for over two years, I worry about a few things: (1) Damaging your brand with inappropriate content. (2) Creating opposition—if you share an opinion, you will inevitably attract dissenters. (3) The danger of chasing clicks and becoming increasingly sensationalist.

While it's great that so many are trying this content-driven approach, I often see it backfire. For founders, too, their social media activity can negatively impact investor perception. Honestly, this fear makes me spend more and more time writing each post as my subscriber count grows, and I find myself becoming more cautious.

I also worry about creating an opposition. For any opinion, there will always be people who disagree, and sometimes they get organized. I've received tips that an "Anti-IanPark" group chat exists, and I know my writing can affect my reputation among founders and LPs, so I am very careful about what I write.

Furthermore, this strategy exposes you to people outside your target audience of founders and investors, which I think creates unnecessary risk. In my experience, it's fine when things are going well, but when something goes wrong, you can face a much bigger backlash. It feels a bit like you're borrowing from the future by making a high-stakes bet in the present.

Finally, I'm wary of the trap of writing increasingly provocative topics just to get views. When you create content, there's a moment when clicks and view counts become the KPI, rather than the original purpose of learning and thinking to support your main job. This can lead to a vicious cycle of (1) low-quality content and (2) creating more opposition.

Perhaps I'm more worried because I'm in the trenches myself. But is this playbook bad? If you can be more strategic and diligent to reach the "upgraded version," I believe many of these concerns can be resolved.

Upgraded Version: The Researcher

The final form is becoming a recognized expert who delivers deep insights on a specific sector or macro trend. In the founder survey mentioned earlier, "sector expertise" was the second most desired VC trait.

In this case, sourcing is primarily a top-down effort, using research to identify and proactively contact companies that fit their thesis. They pick and win by persuading founders with their superior understanding of the sector. They provide support by sharing the latest industry tech and trends. What founder would turn down intel on their competitors, like who's #1, how many new customers they're acquiring, and what their valuation is?

When content has this level of depth and objectivity, the problems I worry about—(1), (2), and (3)—are unlikely to occur. I'm working every day to get there myself.

• Examples: Mary Meeker (Bond), Tomasz Tunguz (Theory), Micky Malka (Ribbit)

Meeker of internet report fame, rising star Tunguz, and fintech legend Malka.

4. The Number Cruncher

While Researchers love the big picture and macro trends, Number Crunchers are laser-focused on the company level, investing based on micro trends.

Strength

This analytical type, common among hedge fund alumni but less so in traditional VC, relies on deep valuation analysis, often creating traditional hedge fund-style reports. They tend to be more flexible, sometimes investing in public stocks to capture more opportunities. They use interesting and novel strategies, like macro timing, and show strength in later-stage deals where traditional VCs are relatively weak. Some have shown great performance by carving out this niche.

Weakness

They are a bit removed from traditional early-stage VC and tend to be weaker at investing in pre-data startups. Even at Series B and beyond, startups carry far more risk and unpredictability than public companies. Furthermore, if their unique strategies don't pan out, they can lose the trust of LPs who invested for a traditional VC strategy.

Upgraded Version: Private Equity, Hedge Fund, Hybrid Fund

Since these types are accustomed to scrutinizing large datasets and public information, PE or public market investing often seems like a better fit. Some operate hybrid funds that invest in both public and private markets. I sometimes wonder why they take on VC risk with their skill set, but they clearly believe there's a gap in the market. While their flashy and unique ideas are impressive, I have questions about their long-term, consistent profitability.

• Examples: Lee Fixel (Addition), Neil Mehta (Greenoaks), Brad Gerstner (Altimeter)

Fixel from Tiger, Mehta from D.E. Shaw, and Gerstner, the super-connector with two soccer fields at his house.

5. The Visionary

To put it grandly, this is the philosopher or prophet VC who reads the currents of the world and discovers an investment thesis through original thought. It’s a bit of a results-oriented label, but legendary VCs who are seen as having an eye for the future fall into this category. It seems to require a combination of immense effort, diverse experience, a powerful ecosystem, and a healthy dose of incredible luck for everything to fall into place.

Strength

They make world-changing investments with a unique perspective that differs from everyone else. With great returns and immense industry respect, they can get into almost any deal they want and collaborate easily with other funds.

Weakness

While they may think about the fund's long-term continuity, they often can't shake the feeling of being a one-person team. Their personal brand and performance are so strong that if something were to happen to the Visionary, the entire fund's existence could be called into question.

Upgraded Version: The CEO

The ultimate upgrade is to be a star investor who also succeeds in building an enduring franchise fund that can carry on their legacy. These individuals invest heavily in the fund's structure and systems, experiment with different models, and spend significant time mentoring the next generation.

• Examples: Peter Thiel (Founders Fund), Vinod Khosla (Khosla Ventures), Bill Gurley (Benchmark), Hemant Taneja (General Catalyst)

Thiel of the PayPal Mafia, the one and only Khosla, Gurley of the lean elite, and the sage of Boston, Taneja.

So, what?

Okay, so which one are you?

I want to be a Matchmaker & a Researcher!

Unfortunately, I never had the chance to be a founder. I have a bit of a fascination with and an inferiority complex towards operator VCs, and I think many non-founder VCs feel the same way. I honestly considered starting a company just to get over it, but I realized that's a terrible reason to start a company, so I gave up that idea.

And based on my own gut feeling, I like to believe I'm a pretty decent salesperson and have a knack for matchmaking. I put a lot of effort into figuring out who to introduce our portfolio companies to for mutual benefit, which investors to connect to make everyone happy, and even things like how to connect with Professor Andrew Ng to build the right network to eventually meet the president, and what agenda would get his attention. I'm working hard to get even better at this!

I also personally love reading, summarizing, and debating the latest trends, so I'm trying hard to become a Researcher. Lately, I've been obsessed with the consumer space, which is seeing rapid growth at relatively reasonable valuations. I'd love to meet with interested investors and founders in this area! If you have any good reports or information, feel free to email me at ipark@sazze.vc.

So, what should I do?

This might sound presumptuous, but I believe everyone has natural talents and a unique personality. We all have things we are relatively better at and enjoy more. As you've seen, every playbook has its upsides and downsides (or is just plain hard to achieve), so it's difficult to say one is definitively better than another.

There are countless successful VCs in both Korea and the US who have never written a single line for a blog. Media activity is more visible, so it might seem appealing, but it comes with its own risks. Ultimately, your playbook has to fit you, and you have to enjoy it. I have many friends who write great content because they love it, and I cheer them on. At the same time, considering the risks, I don't think everyone should feel pressured to do it. It's the same logic I used when I decided not to start a company.

I believe we are in the service industry, managing our clients' money to grow it for them. In the end, we are judged and evaluated on our returns and our service. Everything else—all of it—is just a tool to do those two things well. The specific tool doesn't matter. Just make money and provide great service.

As you all feel, the VC world is an incredibly competitive market. I think it's far better to focus on your own strengths and talents, on what you truly enjoy, and to sharpen that into your chosen weapon. That's a much better use of your time than trying to compensate for your weaknesses or wasting time with a tool that doesn't suit you.

Our life is too short, and every single day is game day.