On August 25, SemiAnalysis founder Dylan Patel laid out a set of projections on Dwarkesh Patel's podcast: by the end of 2028, Anthropic and OpenAI will together control most of the world's usable compute.
"By the time you're towards the end of 2028 — if this trend continues, and I see nothing that's stopping it — you've got them just controlling most of the usable flops in the world on their own."
Here's how the numbers unfold. Each lab's power footprint: about 2GW apiece by the end of 2025, more than 5GW each by the end of 2026, roughly 18GW each by the end of 2027, and about 54GW each by the end of 2028. Their share of new global compute: around 30% this year, 40% to 50% next year, roughly half of all new capacity by around December 2027, rising to 70-80% in 2028. On the global side, new capacity comes to 30GW in 2026, 50GW in 2027, 70GW in 2028, and 90-100GW in 2029, putting the 2028 global total at around 200GW.
How Much Revenue Per Megawatt
Patel argues the drive toward concentration is hidden in per-unit revenue. Current infrastructure costs run roughly $10 million to $15 million per megawatt, while Anthropic's revenue per megawatt has at times spiked to $50 million. He expects that by the end of 2027, revenue per megawatt at the top labs will climb past $50 million, possibly reaching $70-80 million. Whoever earns more per megawatt can outbid everyone else for rack space.
"To get to that 100 gigawatts in 2028, you have to believe that the labs can outpay for compute...Does compute now get to $25 million a megawatt? Does it get to $40 million a megawatt?"
He collapses three mechanisms into one line — "Every force is screeching towards centralization." Economies of scale in training, higher per-megawatt revenue for the frontrunners, and the purchasing power that revenue buys, all interlock and reinforce each other.
Where the Money Comes From
The capital table is the heaviest part of this projection. Between 2024 and 2029, total capex across the ecosystem comes to roughly $11 trillion, with $6 trillion paid in cash and $5 trillion financed through debt — requiring more than $5 trillion in newly issued credit. On rates, Meta's current cost of financing sits at 5% to 6%, and Patel sees no reason it wouldn't climb to 8%; that 250-basis-point jump would ripple through the credit market into the broader economy.
On China, he offers three judgments: China's AI compute will sit at 30GW or less by 2028; China currently accounts for less than 10% of global data-center AI compute deployment; and a conversion factor — "that 50 gigawatts is really worth as much as 20 gigawatts from American chips." That conversion ratio is his own estimate, its methodology undisclosed, and treating it as a firm conclusion would be risky.
Where's the fragility in this projection? It ties three curves into one: compute scale, revenue per megawatt, and credit supply. The first two feed each other — higher revenue buys more compute, more compute builds the next model, a stronger model drives revenue higher still — and once that loop starts, it can genuinely accelerate itself. The third is exogenous. Rough math: if an 8% financing cost holds, interest alone on $5 trillion in debt runs to $400 billion a year, a figure that requires AI-related revenue to reach a fairly extreme level over the same period. Patel himself leaves this part as a question — he asks whether the labs can keep outbidding for compute, without following up on what happens if they can't.
For everyone else in the industry, that ratio curve matters more than the absolute numbers. If 70-80% of new compute flows to two labs, the remaining 20-30% has to be split among every other lab, cloud provider, in-house enterprise build, and sovereign project. The global total keeps climbing — but the slice left over keeps shrinking.
Sources: Dwarkesh Podcast, CocoLoop, SemiAnalysis — worth verifying independently: the gigawatt ranges for both labs, the methodology behind per-megawatt revenue figures, and the cash/debt split within the $11 trillion capex total.