Elad Gil: Most AI Companies Should Sell Themselves Within 18 Months

Source: Tim Ferriss | Published: 2026-04-29T22:04:02Z

Top Silicon Valley angel investor Elad Gil argues that every tech cycle has wiped out 95%+ of companies, and AI will be no exception — founders should seriously consider exiting while the valuation window is still open.

50 to Several Hundred People Just Had a "Personal IPO" — All at Once

Over the past year, a phenomenon in AI has gone largely undiscussed. Meta began aggressively bidding up AI talent, forcing other Big Tech companies to match. The result: 50 to several hundred elite AI researchers scattered across Silicon Valley saw their compensation packages suddenly spike into the tens or even hundreds of millions of dollars — as a group, they experienced an IPO-level wealth event simultaneously.

Gil says the only historical precedent he can think of is the crypto wave around 2017, when a cohort of early holders and founders got rich all at once. The consequences of this "class-wide IPO" are predictable: some will keep their heads down and keep grinding, others will get distracted — pursuing scientific philanthropy, politics, personal interests, or simply checking out.

Memory, Not Chips, Is the Real Bottleneck in the AI Arms Race

Every AI lab — OpenAI, Anthropic, Google, xAI — is buying as much compute as it can for model training. But the constraint is no longer Nvidia's chip production capacity. It's a specific type of memory manufactured primarily by Korean companies. Gil estimates this bottleneck will persist for roughly two years, as these manufacturers underestimated demand and now need time to bring new production lines online.

This bottleneck creates a fascinating competitive dynamic: in the short term, it prevents any single lab from pulling decisively ahead. No one can buy ten times the compute of their rivals, so OpenAI, Anthropic, and Google will remain relatively close in capability. But once this constraint lifts, one lab could suddenly leave the others in the dust.

From Zero to 1% of GDP

Gil's team put together a chart tracking how long different generational companies took to go from zero to one billion, and from one billion to ten billion in revenue. ADP took thirty years to hit one billion. Google took about four. OpenAI and Anthropic each took roughly one.

Both companies are now reportedly at around $30 billion in annualized revenue. Gil does some quick math: that's already about 0.1% of U.S. GDP. And that doesn't include AI-driven revenue at cloud providers like Azure and GCP. If each hits $100 billion in revenue within a year or two, each company would represent 1% to 2% of GDP.

Every Tech Cycle in History Has Killed 95%+ of Companies

In a recent essay, Gil argued that founders running successful AI companies should soberly evaluate the option of exiting within the next 12 to 18 months — this may be the window that maximizes value.

His argument is historical. Around 450 companies IPO'd in 1999, another 450 or so in early 2000, plus those from prior years — roughly 1,500 to 2,000 companies went public in total. How many survived to today? About a dozen, maybe two dozen. Roughly 1,980 vanished.

"Every cycle is the same. SaaS, mobile, crypto. Most companies don't survive."

There's no special reason to think AI will be the exception. Every AI founder should ask themselves: are you one of the dozen that still matters in ten years, or is right now your peak?

A Few Tests for Durability

Gil believes the core labs — OpenAI, Anthropic, Google — will persist long-term, barring a major blowup. Three years ago on Substack, he predicted this would be an oligopolistic market aligned with cloud providers, and that's roughly how it's played out.

For application-layer companies, he proposes a few litmus tests: Does your product get meaningfully better when the underlying models improve? Is your product so deeply embedded in customer workflows that it's painful to rip out? Are you accumulating and leveraging proprietary data?

He makes one point in particular: the real adoption barrier for AI companies usually isn't whether the technology works — it's change management. Will customers actually change how they work? If you've successfully embedded yourself into workflows, that stickiness alone is the moat.

Every Investment Should Boil Down to One Core Belief

Gil does extensive due diligence — reviewing financials, making customer calls, vetting executive teams. His fund even runs cash audits. But he says all of it ultimately collapses into a single question: what is the one thing I need to believe about this company?

Coinbase: it's an index fund on crypto, and crypto will keep growing. Stripe: it's an index fund on e-commerce, and e-commerce will keep growing. Anduril: machine vision and drones matter for defense.

"If you need to believe three things, it's too complicated — probably won't work. If you don't need to believe anything, that doesn't make sense either. Usually it's one or two core insights that determine the outcome."

Perplexity's CEO Built What They Discussed Every Two Weeks

Gil's early investment in Perplexity illustrates his style well. Aravind (Perplexity's CEO) cold-messaged him on LinkedIn while still an engineer at OpenAI, back when nobody was paying attention to AI. They started meeting every two weeks to brainstorm.

What ultimately convinced Gil to lead the first round was simple: every time they discussed an idea, Aravind would show up a week later with a working product. Gil's exact words: "Who can do that?" This was a classic people-over-market bet — he was surveying every AI direction at the time and stumbled onto someone with extraordinary execution speed.

Anduril was the opposite: Google shut down Project Maven (defense-related), and Gil reasoned that if big companies wouldn't touch it, that was a golden opportunity for a startup. Market gap first, then the right people.

91% of Global AI Private Market Value Is Concentrated in the Bay Area

Data analysis by Gil's team shows that 91% of global private-market value in AI is concentrated in a roughly 10-by-10-mile area in the San Francisco Bay Area. The "you can work from anywhere" narrative, at least in AI, doesn't hold up.

He draws parallels to other industries: want to make movies, go to Hollywood; want to do finance, go to New York; defense tech clusters around SpaceX and Anduril in Southern California. The gravitational pull of these industry clusters is real. The most fundamental reason Gil was able to invest early in Airbnb and Stripe was simply being in San Francisco, plugged into the right networks, naturally meeting the right people.

Distribution Never Happens Organically

Gil points out a fact that gets romanticized over and over: virtually every company that ended up worth tens or hundreds of billions employed brutally aggressive distribution strategies — stories conveniently omitted from later TED talks.

Google spent hundreds of millions in its early years getting its toolbar bundled into software downloads. Facebook bought ads in Europe targeting people's own names as keywords — because the most common thing people search for is themselves — funneling them into signups. TikTok spent billions on distribution to build network effects. Snowflake spent billions on sales teams and channel partnerships.

"Occasionally you see a company win not because they had the best product, but because they were stronger at sales and distribution. That frustrates technical people, but it's how it works."

From Selling Software to Selling Labor: AI Unlocked Previously Closed Markets

Why has selling software to law firms historically been a terrible business, yet Harvey is growing at breakneck speed? Gil argues the key shift is this: AI transformed the business model from selling seats or subscriptions to selling the equivalent of human labor. You're no longer selling a tool — you're selling billable hours of cognitive work.

This explains why AI has suddenly pried open so many long-closed markets — every domain involving language, data, and white-collar work became instant greenfield for AI. Code is a form of white-collar work. Email, documents, enterprise data — all of it. Language is everywhere, so the market is everywhere.

Gil also notes that unlike language, robotics — even with stronger models — faces a relatively small existing hardware market and lacks that plug-and-play path to commercialization.

When a Trillion-Dollar Company Can Buy You for 1%

The acquisition capacity of today's tech industry is unprecedented. Fifteen years ago, the world's largest market cap was around $300 billion. Now multiple companies sit between one and several trillion. One percent of $3 trillion is $30 billion — meaning a "small" acquisition can reach scales that were previously unimaginable.

The pool of potential acquirers is also larger than ever: major labs and hyperscale cloud providers, vertical industry giants (think Thomson Reuters for legal), and large private companies like Snowflake, Databricks, and Stripe. Gil also flags an underrated option: mergers between competitors. PayPal was born from the merger of X.com and Confinity — two companies doing the same thing that decided to stop bleeding each other dry.

Using AI Models for Deep Research on Autism Diagnosis Rates

Gil's approach to gathering information has fundamentally changed. Beyond X and conversations with smart people, he now runs two or three models simultaneously for deep research, asking them to aggregate clinical trial data, provide primary source citations and summary charts, then goes back to verify everything himself.

He gives the example of autism diagnosis rates: 30 years ago it was roughly one in several thousand; now it's approaching 3%. Intuitively you'd guess it's driven by older parents, but digging into the data reveals the primary driver is dramatically broadened diagnostic criteria, plus financial incentives within the school system to diagnose. In New Jersey, 60% of autism diagnoses aren't even based on clinical criteria — just a teacher's subjective judgment. He also found that maternal age's impact on risk is actually larger than paternal age in some datasets, but public discourse focuses almost exclusively on paternal age — likely influenced by social attitudes and political leanings.

Uploading Founder Photos for AI Personality Reads

One "weird thing" Gil has been doing lately is uploading founders' photos to AI models and asking them to predict personality traits based on micro-expression features. His prompts instruct the model to play the role of someone skilled at cold reading, making judgments based on crow's feet around the eyes (indicating genuine smiles), brow furrow patterns, and other micro-features, and to explain the specific visual evidence behind each conclusion.

He says the results are surprisingly good. The models produce highly specific assessments, such as "this person likely holds back in most social situations, then suddenly drops a witty remark no one sees coming." He acknowledges it's not necessarily accurate or predictive, but as a tool for augmenting intuition, he finds it fascinating. After all, investors already make similar snap judgments in the first few minutes of a meeting — they just do it subconsciously.

Ten-Year Planning — Something Gil Had Never Done Before

At the end of the podcast, Gil mentions he recently tried ten-year life planning for the first time. He never used to reflect on the past — in his own words, "I think about the past close to zero." But he discovered that stretching the time horizon to a decade changes how you define ambition and what you think you should be doing right now.

Some will say AGI arrives in two years and planning is pointless. Gil sees that as defeatism. The bigger the disruption, the more you need a sense of direction — plans can always be adjusted, but having no plan means you're just drifting.