TOPSHOT – Pedestrians walking through Waterloo Bridge with the skyline of the City of London in the background on October 27, 2016. – Britain’s economy won a double boost on October 27 on news of faster-than-expected growth following its vote for Brexit and a pledge by Nissan to build new car models in the UK. Gross domestic product expanded by 0.5 percent in the third quarter, official data showed. (Photo by Daniel LEAL / AFP) (Photo by DANIEL LEAL/AFP via Getty Images)
AFP via Getty Images
Elad Gil, the investor and entrepreneur whose early bets include Airbnb, Stripe, and Coinbase, published a wide-ranging AI thesis on April 20 arguing that artificial intelligence has grown from near-zero to a measurable share of U.S. GDP in just a few years and that the window for founders to exit their AI companies is narrowing fast. The analysis is dense; counterintuitive in places, and worth reading for anyone allocating capital or building in the space right now.
U.S. GDP sits at roughly $30 trillion. Gil estimates OpenAI and Anthropic are each running at approximately $30 billion in annualized revenue, putting each at 0.1% of GDP. Factor in cloud infrastructure and adjacent AI services and the sector has already reached 0.25% to 0.5% of GDP. If both AI labs hit $100 billion in revenue by year-end, as Gil and others have speculated, AI will account for roughly 1% of U.S. GDP on a run-rate basis before 2027. That trajectory has no comparable precedent in modern tech history.
The VC angle is where Gil’s argument gets most actionable; he advises founders running successful AI companies to consider exiting in the next 12 to 18 months. “Most companies, including those that are ramping revenue today, will see the market, competition, and adoption turn on them,” he writes, drawing an analogy to the dot-com era when roughly 2,000 companies went public between 1995 and 2001 and only a handful survived. The current moment, in his framing, is the upswing. Founders who wait for a clearer picture may find the picture has already changed.
That call to exit is calibrated against a broader structural argument: compute scarcity is becoming an artificial ceiling on model progress. Gil contends that memory supply from Hynix, Samsung, and Micron will constrain GPU buildouts for at least two more years, meaning no single lab can pull decisively ahead until 2028 at the earliest. The implication for investors is a durable oligopoly among the major labs for the foreseeable future, not a winner-take-all race. Concentration of capital in a small number of frontier model providers is, in Gil’s reading, structurally locked in by hardware supply chains, not just by talent or algorithms.
On labor markets, Gil distinguishes between noise and signal. Most announced “AI layoffs” are, he argues, deferred corrections from pandemic-era overhiring dressed up in AI language. The real displacement is happening quietly in outsourcing, not on corporate balance sheets. Companies cutting customer support headcount are canceling contracts with offshore service providers first, which means the employment shock is landing in India and the Philippines before it registers in U.S. jobs data. The Philippine IT-BPM sector employs roughly 1.7 million people and has historically depended on exactly the kind of voice and data work that AI agents are beginning to absorb.
Gil also introduces what may become a useful frame for thinking about AI adoption velocity: the tightness of the closed loop. Jobs where AI can test, fail, and iterate autonomously, like software engineering, will see the fastest displacement because the feedback cycle is short and measurable. Jobs with slow or ambiguous feedback loops, like management consulting or strategic planning, will take longer. Gil places software engineering at the top of this matrix, which explains why coding tools like Cursor have attracted funding despite being less than three years old.
The compute-as-currency framing is one Gil spends time developing. Token budgets, he argues, are becoming organizational inputs alongside salaries, raising the question of what the right ratio of compute spend to headcount looks like for different types of companies. Some firms, he notes, are already running the analysis. Cursor and other developer tools are, in his reading, effectively subsidizing inference as a user acquisition strategy, which compresses margins but builds the kind of daily usage habits that translate to retention.
For investors, the Gil thesis resolves into a few concrete bets: durable exposure to the major frontier labs while compute constraints hold, attention to companies building proprietary closed-loop data systems in high-value verticals, and skepticism toward any AI company whose moat depends on model performance alone rather than workflow lock-in. The exit advice for founders also carries an implicit message for growth-stage investors: secondary market liquidity and M&A activity in AI should accelerate through 2025 and into 2026 as founders who absorb this logic look for doors.
Gil has been right before on timing: he co-founded Color Genomics and served as VP of Corporate Development at Twitter, and his 2018 book High Growth Handbook became a standard reference for late-stage startup operators. Whether the compute ceiling holds, whether the exit window stays open through 2027, and whether the GDP math plays out as projected are all open questions. But the analytical architecture Gil is using, closed loops, compute as currency, oligopoly by hardware constraint, is rigorous enough to stress-test your own assumptions against whatever position you currently hold.