Semiconductor analyst Dylan Patel forecasts that OpenAI and Anthropic will control the majority of global AI compute capacity by 2028.

By the end of 2028, two private companies—OpenAI and Anthropic—could control the majority of the world’s usable computing power. That is the central prediction from Dylan Patel, founder of semiconductor research firm SemiAnalysis, delivered during a wide-ranging conversation on the Dwarkesh Podcast. Patel argues that a fundamental shift in lab economics has made this concentration not just possible, but financially inevitable. The frontier labs have crossed a critical threshold: their inference businesses are now so profitable that each dollar spent serving customers generates enough margin to fund the next round of model training.

Anthropic’s revenue has reached $50 million per megawatt of compute against a base cost of $10–15 million—a 3–5x return that creates a self-funding flywheel no competitor can match. As a result, these two labs will absorb 40–50% of the world’s incremental compute next year, rising to 70–80% by 2028. At the start of 2026, each lab controlled roughly 2 gigawatts of compute. By year’s end, that figure is projected to hit 5 gigawatts—a 3 to 4x increase in twelve months. Global AI infrastructure capital expenditure already crossed $1 trillion in 2026 and is on track to exceed $2 trillion by 2028. The labs’ share of incremental compute is rising from about 30% this year to a projected 40–50% next year.

The raw gigawatt figures understate the concentration, because each new generation of chips delivers 3–5x more performance per watt. When the labs take 45% of new physical compute in 2027, they are absorbing a far larger share of the world’s effective floating-point operations. Patel’s projection is stark: “By the time you’re in towards the end of 2028, if this trend continues, which I see nothing that’s stopping it, you’ve got them just controlling most of the usable flops in the world on their own.”

For years, frontier AI labs burned venture capital to train models, hoping inference revenue would eventually justify the spending. That day has arrived. In the second quarter of 2026, Anthropic turned its first adjusted operating profit, and OpenAI is expected to follow in the third quarter. The mechanism is simple arithmetic, which Patel spells out: “If I spend $10 on inference capacity, actually generate $50 of revenue, and then I can turn around and incrementally spend all of that profit on training.” The labs no longer need to beg investors for every dollar; they can self-fund from operations.

What prevents the rest of the market from keeping pace? According to Patel, the answer lies in the bullwhip effect. A small increase in demand at the customer level triggers a delayed, amplified reaction further down the supply chain. By the time the signal reaches mirror makers, turbine manufacturers, and memory fabs, demand has grown again, and the signal never catches up. The economics of this delay are staggering: a $6 billion fab investment produces a gigawatt of compute annually, generating $100 billion in revenue per year. The return on investment sits at roughly 100x over five years. Anyone who can secure physical components—EUV tools, high-bandwidth memory, gas turbines—can capture enormous value.

Patel’s frustration with the manufacturing pace is palpable: “We’re like Anthropic and OpenAI, like we could make a trillion dollars right now but we’re just bottlenecked on the mirrors that go into the ASML machines.” Carl Zeiss, the German optics company supplying those mirrors, has committed to producing enough for 100 EUV tools per year by 2030—an upgrade from earlier plans, but still far below what current economics justify. This bottleneck has shifted market power unexpectedly. SpaceX, which has built massive compute capacity, has sold to Anthropic and Google at $25–40 million per megawatt—well above the $10–15 million base cost. Meta and SpaceX are now positioned to hoard compute, choosing whether to use it internally or sell at a premium. Companies with strong enough balance sheets to build without pre-committed customers now control pricing.

If AI infrastructure investment continues to compound, the consequences will extend far beyond the technology sector. Patel’s modeling suggests the AI ecosystem will require approximately $11 trillion in capex from 2024 to 2029, of which $5 trillion must be funded through credit issuance. When Anthropic, OpenAI, and the hyperscalers raise hundreds of billions in debt, they compete with every other borrower—governments, consumers, and corporations. Interest rates rise across the entire economy. Meta’s recent bond issuance at 5–6% could face 8%+ rates. The U.S. government’s debt servicing burden, currently about 20% of tax revenue, could balloon to 40–60% if rates rise by 5 percentage points.

The ripple effects are severe. A 2–3 percentage point rise in rates crashes the present value of long-duration cash flows. Traditional safe-haven equities—Johnson & Johnson, railway operators, dividend aristocrats—would see valuations collapse. Meanwhile, developing countries with high debt loads and frequent rollovers face default risk. Both Patel and host Dwarkesh Patel reference a potential “second Volcker shock,” nodding to the early 1980s when Paul Volcker’s rate hikes triggered a wave of Latin American defaults. Patel argues the U.S. will likely survive because it can tax the data centers located on its soil, whereas other nations are less fortunate.

The geopolitics of compute centralization are inseparable from its economics. Since 2022, U.S. export controls have dramatically shifted the global balance: 70% of new AI compute watts are now deployed in America, while China accounts for less than 10% of incremental compute—a stark reversal from 2022, when China added 30–35% of the world’s total. The reasons are twofold: export controls on advanced NVIDIA chips, and a financial system less willing to fund speculative startups. However, Patel notes that once China’s financial system decides to back an industry, it subsidizes it more aggressively than any other country. The Chinese semiconductor industry already receives more subsidies than the rest of the world’s semiconductor industries combined.

The trajectory follows a delayed hockey stick pattern. In 2026, China still relies on smuggled chips and TSMC-made silicon for companies that turned out to be Huawei fronts. But by 2027–2028, domestic fabs from SMIC and CXMT will begin producing at scale—millions of units per year. Patel projects China could add 5–10 gigawatts of domestically produced compute in 2028, and potentially 50 gigawatts of incremental compute by 2029. The catch is quality: a gigawatt of Chinese compute might be worth only 0.4 gigawatts of American compute in usable flops. Dwarkesh Patel draws a striking conclusion: by 2028–2029, a single leading U.S. lab could have more effective compute than all of China combined—a medium-term validation of the export control strategy.

The most provocative segment of the conversation addresses the effective “population” of AI laborers. The framework operates on a straightforward calculation: if frontier compute grows 4–5x per year, and the compute required for a given capability level decreases 3x per year, then the effective AI population at the frontier expands 4–5 times annually.