Dylan Patel – Anthropic & OpenAI will have most of the world’s compute by 2028

The episode opens on a running joke — this is the annual Dylan Patel conversation, their version of a family Thanksgiving, with no actual relation — and then states the premise. Where the world economy goes is increasingly a function of where lab economics go and where the compute market goes, so the question is what the extreme version of that looks like within a few years.

The starting point is stark. Last year, even at the end of it, most of US GDP growth was AI infrastructure. This year roughly a third of the compute coming online is for the labs — OpenAI and Anthropic — whoever builds it and rents it to them, they are the end customer.

CapEx is a little over a trillion dollars this year and headed past two trillion by 2028, and the labs take a rising share of it. The shape of the change is that these go from companies spending tens of billions a year to hundreds of billions, with contracts already signed that forecast trillions a year by the end of the decade.

That requires the economics to be reshaped, and Patel's claim is that it already has been. Until recently the labs mostly lost money on venture capital; Anthropic turned profitable in Q2, and OpenAI is believed to be close behind on the strength of Codex and the current model generation. New capital still arrives, but it now accelerates growth rather than funding the business.

The mechanism is margin. Serving a model costs roughly ten to fifteen million dollars per megawatt of capacity, and that used to be the whole story: GPT-4 on Hopper generated negative gross margin for OpenAI.

The change is that revenue per megawatt has moved well past the cost. For the current models, and for Anthropic in particular, revenue has run as high as fifty million dollars per megawatt — which means ten dollars of inference capacity generates fifty dollars of revenue, and the profit can be turned straight back into training. That single ratio is what converts a loss-making research lab into a self-funding infrastructure business.

Dwarkesh asks the centralisation question in the form of a date. If a third of marginal compute goes to the labs now, when does more than half of incremental compute go to them, and when do they hold most of the world's compute outright?

The trajectory: OpenAI started the year at two gigawatts, Anthropic just under; both end it above five, a three-to-fourfold increase, which is about thirty percent of the compute added this year.

Next year is already contracted, and it is more dramatic — forty to fifty percent of new compute. Centralisation is not slowing; it is accelerating.

What changes is the builder, not the concentration. SpaceX becomes a large new entrant and will lease much of its capacity to the labs, because they are the ones with the marginal ability to pay the highest price. The labs also start building for themselves: OpenAI with its own chips, Anthropic with TPUs bought from Google and deployed through Fluidstack.

So the answer to the date question is the end of next year, when half of incremental compute is already going to two customers. And because compute is growing so fast, incremental compute is most of compute — which means within about eighteen months to two years, most of the world's compute either belongs to or serves the demand of two labs.

Dwarkesh puts the trends side by side. World compute in gigawatts might double each year; frontier lab compute triples. From two at the start of this year to nearly six at the end, then eighteen by the end of 2027 and fifty-four by the end of 2028 — at which point the arithmetic must break against the available supply.

Patel's answer is that the arithmetic is misleading because a new watt is not an old watt. Thirty gigawatts added this year, fifty next, seventy after that; but this year's watts are GB300s, TPUv7s and Trainium3s, three to five times the performance per watt of the previous generation. The installed base is not doubling, yet its usable capacity is.

Stack the effects and the picture inverts. If the labs take forty-five percent of new compute next year, then by December 2027 they have taken half of the world's incremental compute — and that half is the most efficient half ever deployed. Continue the trend and by late 2028 they control most of the usable flops in the world. Patel says he sees nothing stopping it.

Dwarkesh's objection is the natural one: in a world where the value of a watt rises this much, why is the 2028 addition only eighty gigawatts? Patel's reply is that eighty is the upper bound — the number you get from being maximally bullish — which reframes the constraint as physical rather than economic.

They then work the supply chain from the bottom. A gigawatt of Vera Rubin, on Patel's earlier numbers, needs 55,000 N3 wafers, 6,000 N5 wafers and 170,000 DRAM wafers — a figure the conversation interrupts for a joke about Patel's accent and a childhood as a vegetarian in North Dakota, and then resumes.

Dwarkesh had a model run Patel's wafer fab equipment numbers: the tooling for a gigawatt of annual output costs three to four billion dollars. Add cleanrooms and shell and it is six billion of fab CapEx to produce a gigawatt every year — and a gigawatt currently produces a hundred billion dollars of revenue a year.

Compound it and the ratio becomes absurd. The first gigawatt earns five years of profit, the second four, and so on; six billion of fab CapEx supports more than a trillion dollars of end revenue. Patel grants the OpEx, the data centre, the power, the installation, the R&D paid to OpenAI, the many hands in between — and even halving everything for the middlemen leaves a hundredfold gap between fab CapEx and end revenue.