Garry Tan Passed Up $4 Billion to Hit Level 60 at Microsoft

Source: a16z | Published: 2026-08-12T14:33:35Z

When Peter Thiel handed Garry Tan a check to start a company, Tan turned it down to chase a Microsoft promotion. He now estimates that decision cost him $4 billion.

Peter Thiel called him into his office and handed him a $70,000 check — the same as his annual salary at Microsoft — and told him this was what he should be doing. The two men who'd invited him — his Stanford fraternity brothers Joe Lonsdale and Stephen Cohen — were among the smartest people he knew. He said no.

The reason: he was due for a promotion to Microsoft Level 60 that year.

He got it.

Never Follow the Map. Look at the Ground Beneath Your Feet

This was, by his own reckoning, the second time he'd made the same mistake. The first came when he graduated from Stanford — the Nasdaq had just crashed, and Bay Area tech jobs were nearly impossible to find. He'd already been doing web development for five years, years ahead of everyone else. Zuckerberg was probably building FaceMash around that time. Garry decided to leave the web and chase the "next hot thing" — mobile.

"If I were a time traveler, the message I'd send to myself is: just stay on the web. You're five full years ahead of everyone. You could have lapped them all."

He later named this quality "courage." In an age where a single blog post or tweet can shake your convictions, trusting your own direct experience is itself rare. He describes it with a scientist's analogy: don't just read the published papers — go verify the truth yourself, talk to the people who lived it firsthand, and you'll end up with knowledge nobody else has.

YC Does One Thing Others Don't. It's 12 Questions.

Breaking into Silicon Valley used to require navigating an entire social infrastructure — the right school, the right people, an invitation to the right party. Garry calls those people "sceners." YC's move was to skip that layer entirely.

What Paul Graham and Jessica Livingston did was deceptively simple: post the application on a website — twelve questions, and later a one-minute video — then have people who had actually built things give every applicant a fair read. This year's Startup School brought 7,000 people to San Francisco, most of them for the first time. Garry says they've started calling it "the Birthright for tech."

But the most important thing YC gives founders, he says, is a community where you can tell the truth. At a TechCrunch event, you ask someone how things are going and they say "killing it," and you say it back. But on the day your best customer leaves, on the day your best engineer quits, you need someone you can actually call. Even harder: on the day your co-founder loses hope — and you're not allowed to.

Pure SaaS Still Works, But You Need a Plan for What Comes Next

Garry recently ran an analysis whose conclusion surprised even him: he's not sure pure seat-based SaaS will exist in its current form five to ten years from now.

That's not a eulogy for SaaS. His point is that if you're building it in 2026, it can only be a wedge — behind it, you need a moat built on data or network effects. Two years ago, SaaS was god — 10–20x ARR multiples on any given twelve-month horizon felt like physics. He says it's hard to believe that world existed.

Cheap code comes with a side effect: in the past, you had to believe what you were building mattered — with a capital M — to have the motivation to push something to completion. But many of the most important things don't look important early on. The Homebrew Computer Club wasn't an exciting place. It was just a room full of people chasing strange obsessions.

A Markdown File Is an Employee

Garry recorded every YC office hours session for months. He started using Claude to analyze the data, looking for the conversational moves that caused a founder's ambition to step-change — what he calls a "step function." He ended up with a markdown file he considered the ideal coaching script. He used Claude to compress it by 90%, then open-sourced it.

The logic behind this is bigger than the example: a markdown file is an employee. It does the job perfectly, as many times as you need it done.

He calls the methodology "skillify" — you do something, turn it into a markdown file plus code plus tests, drop it into a cron job. The first run is messy and expensive and takes many iterations, but eventually you have a reusable capability unit. Every future failure is just a bug fix — fix it once and it doesn't happen again. This applies to sales, marketing, customer support — nearly every business process. He's seen companies go from zero to $15M ARR in four months: two or three people, a few hundred skill files.

How Pedro Knows What Happened in Meetings He Never Attended

Brex CEO Pedro Franceschi is the most aggressive user of agents for organizational management Garry has ever seen.

Pedro uses a tool called OpenClaw — which has a reputation in some circles for being so dangerous it will wipe your entire machine. Pedro's solution was to wrap it in an open-source monitoring layer he calls CrabTrap, which watches all of OpenClaw's network traffic and operations. That's how a heavily regulated fintech company can actually use it.

Here's what he actually does with it: he has agents read every meeting transcript from his direct reports, two levels down. More importantly, he uses the meetings he wasn't in to find where things are breaking, who's in conflict with whom. Sometimes he walks into a conference room already knowing everything that's happened with a compliance team over the past three weeks, and says: "You're right. Do it your way." Then he leaves.

The real problem this solves, Garry says, is fundamental to any company at scale: the company gets too big to fit inside one person's head. Agents with memory and retrieval don't just help you manage someone's opinion — they give you ground truth at the data layer. He calls it "clairvoyance."

Walking Across the Highway With a Baseball Bat

When Garry worked at Microsoft, his team needed integration support from the Windows team. The Windows team was in a building across the highway. They didn't return emails, didn't fix bugs, didn't even mark bugs as won't-fix — just ignored them entirely.

Eventually Garry and his PM mentor walked over in person. The mentor brought a baseball bat. Garry says he still doesn't know exactly what it was for — "it wasn't like he was going to hit anyone, it was just a vague sense of menace." He's thought about this a lot since: two smart people, with decent salaries, mired for four hours in bureaucratic back-and-forth over a P3 bug affecting a thousand users, ultimately needing physical intimidation to get something done. The person who could make the call probably didn't even know the issue existed.

"How much human energy gets wasted this way. How much 2 AM anxiety, just because someone didn't want to have a hard conversation."

What that organizational structure couldn't do, a startup can — and must. That became his conclusion.

Bureaucracy Is the Moat, Not the Enemy

The public discourse around AI, Garry says, is full of apocalyptic framings: white-collar work disappears, a permanent underclass emerges. He thinks the analysis is pointing in the wrong direction.

Bureaucracy is the white pill. Every middle management layer in every company, institution, and organization is structural evidence that change will be slower than anyone expects. Microsoft isn't going anywhere — there are structural reasons for that. Government change takes 20 years.

He stood in front of 7,000 faces at Startup School, most of them 18 to 22, and called them AI natives — the same way his generation was web natives. Twenty years from now, society will run on the expectations these people formed between 18 and 22. What needs to come will come. It just takes time. For startups, that's exactly where the opportunity lives.

The "Arms Race" Starts in 2027

Garry now runs OpenClaw himself, fully loaded with tokens. He estimates he spends $50,000 to $100,000 a year running agents at full intensity, with a million-token context window per request. He says it makes him feel like he's "living in 2028." For most people, that cost is prohibitive.

His read: when the compute cost of today's frontier models drops to $50–100 in two or three years, an "arms race" will break out — like the browser wars, except this time the prize is a billion consumers. The consumer AI software layer, he says, hasn't been properly built yet. For the first time, models genuinely have memory, emotional texture — they can know a person's hopes, fears, and desires, and continuously help them. That war hasn't started yet. It's waiting for the cost curve.

"The token price of frontier models right now is trending toward infinity — you can't get access. But the model that was frontier last week is trending toward zero." He thinks this is a massive white pill.

The Algebra Class That Changed Everything

Garry got pulled into San Francisco local politics around 2022, and there was a specific moment that did it.

The San Francisco school board was debating whether to offer algebra in public middle schools. At the same time, a DA was ignoring crimes against elderly Asian residents. The Bay Area's ABC affiliate wasn't covering anti-Asian hate crimes.

The algebra issue hit Garry personally — he grew up in the public schools in Fremont. If he hadn't taken algebra in middle school, he couldn't have taken calculus in high school, couldn't have gotten into Stanford's engineering program, and nothing that followed would have happened. "They made it personal," he says.

Their approach was intensely in-person: going to farmers markets, telling people one by one what was happening. Once you map out who's actually driving these things, he says, you find it's the same people — the ones opposing data centers, opposing new housing anywhere in California, putting ideology ahead of concrete outcomes.

Through ballot initiatives and community organizing, they recalled the school board. The current San Francisco mayor, Lurie, may be one of the most popular politicians in America right now, Garry says. He doesn't follow national politics — he cares about housing, crime, treatment, and recovery in San Francisco. [garryslist.org](http://garryslist.org) is his attempt to replicate this playbook in Los Angeles, New York, Minneapolis, and Seattle.