enesptjazh-Hanszh-Hantfrdeitkoruarhiidtrnlplvithsvhemsukcselrohudafinobn
A gigawatt of AI compute now costs $60 billion, and chip designers keep $40 billion of it
Dylan Patel says a gigawatt costs $60 billion, or $20 billion without chip design margin. OpenAI's chip and Microsoft's Maia 300 want that gap.
At the CASPA CEO Summit in Santa Clara on October 24, Dylan Patel of SemiAnalysis put a price on the AI buildout that most people skip past. A gigawatt of AI compute cost about $40 billion to build in 2024. Today it costs more than $60 billion. Take out what the chip design houses earn and he puts the number near $20 billion. That gap, roughly $40 billion a gigawatt, is what every buyer of compute now wants back. The same week, a chip founder, a search founder and two of Wall Street's biggest allocators each gave a different answer to who ends up keeping it.

A gigawatt now costs half again what it did in 2024
Patel was on a panel with Bernstein's Stacy Rasgon, moderated by Austin Lyons of Chipstrat. His point was that nearly everyone in the chain has raised prices. Memory makers went from negative gross margins about two years ago to margins Lyons put near 90%. TSMC climbed from about 45% to 60% and, in his telling, never gave much back.
So when you look at this sort of market, right, to build a gigawatt, you know, in 2024, it was roughly $40 billion, right, across chips, land, power, shell, everything. Now it's north of 60.
Dylan Patel, October 24, 2026, CASPA 2026 CEO Summit
He added that Nvidia has kept its own margin roughly flat while it absorbs higher memory, substrate and optics costs, so its rack prices have soared anyway. Rasgon gave Nvidia's own reference figures from the stage, about $25 billion a gigawatt for Blackwell and $40 billion for Rubin, and argued that is fine as long as performance rises faster than price. That is the bull case in one line, and it rests on Nvidia keeping up its pace.
The biggest customers are building their own chips
The number that should worry Nvidia holders came next. By his math, the chip designers' margin is about two thirds of what a gigawatt costs.
If you remove all the margin out of the chip design houses, right, the Nvidia and Broadcom, Marvell, etc., you're actually at like $20 billion a gigawatt.
Dylan Patel, October 24, 2026, CASPA 2026 CEO Summit
Then he described what his firm saw when it tested OpenAI's first chip, the inference part that OpenAI and Broadcom unveiled on June 24 with servers due online by year end. His claim was that "the performance was better than Blackwell" on SemiAnalysis's open-source InferenceX benchmark. That is one lab's test on one class of models, and I would want to see the numbers. But it is a first-generation chip, and he said Anthropic wants to build its own in half the time. Microsoft announced Maia 300 on October 12, pitching 40% more tokens per megawatt and multi-gigawatt deployments starting in 2027. Patel also said Nvidia's share of TSMC's CoWoS packaging capacity and of HBM supply fell in 2026 and will fall again next year, even as its revenue keeps doubling.
His sharpest point was about software. It took SemiAnalysis six to nine months to get inference running on AMD and about a month and a half on OpenAI's chip. If AI models make chips easier to program and then easier to design, the design margin is the part of the $60 billion that can be competed away.
Naveen Rao says the margin stays in hardware anyway
Naveen Rao, who sold MosaicML to Databricks and now runs the chip startup Unconventional AI, told 20VC that commoditization is coming for nearly every other layer. Open models, he said, will carry about 80% of tokens within three years, with spend split roughly 50/50 against frontier models. Inference hosts will see token prices fall and margins shrink. And he does not think the labs will own the stack.
I actually don't think the model companies will be the ones who own the margin of the whole stack. It's not even true today. Like Nvidia owns the margin of the stack today, right?
Naveen Rao, 20VC, published October 24, 2026, source
But the hardware side of it probably won't collapse. I think that part of it is still going to have an advantage because you can still squeeze more juice out of like, hey, I convert energy to intelligence more efficiently.
Naveen Rao, 20VC, published October 24, 2026
Read Rao next to Patel and they agree more than they seem to. Rao means hardware in the broad sense, the chips, systems and networks that turn power into tokens, and he is building a chip that competes with Nvidia. His case is that whoever converts energy to intelligence most efficiently keeps the margin. Patel's case is that the efficient converter may be the buyer itself. Rao also told the host that "if you're a pure software investor, you're kind of screwed" because SaaS will stop earning the margin. He is talking his book, and he admits he is "pretty AI bullish."
Nobody on stage doubted demand
None of this matters if the tokens go unsold, and the demand evidence that week came from an odd place. Parag Agrawal, the former Twitter CEO who now runs Parallel Web Systems, sells web search to AI agents. He said he once pitched agents using the web a thousand times more than humans, and now thinks that was too low.
But my best estimation right now of where we are is agents are using the web 10 to 20% of humans. Which is to say if my thousand-times number is right, we have from here at least 5,000 times or 10,000 times growth left.
Parag Agrawal, Sourcery, October 24, 2026, source
His rule of thumb is that "for every 10 GPUs the world spends on model inference, we'll spend one to two on web search." He also said his team has been through several GPU fire drills in three months, with capacity offered over iMessage before the finance people sign the papers. A forecast of 10,000 times growth is a founder's number. The scramble for GPUs is a fact, and it is why pricing has held so far.
Financing decides who can afford the margin
Patel's other point was that money now limits who can build. A cloud without a contract from a hyperscaler or a top lab cannot get financed. He said Volta paid for its Anthropic cluster with debt at about a 12% coupon sold at 95 cents, close to 15% all in. Semafor reported on October 13 that the $5 billion Norway deal was expected to price at "at least 11%," with a $1.5 billion anchor order from Apollo. Nvidia has started guaranteeing rent on clusters in exchange for revenue shares, which keeps its chips moving and moves some of the landlord's risk onto Nvidia.
Bill Ackman drew the same line on CNBC on October 22, owning Microsoft, Amazon and Meta and avoiding the neoclouds.
the neocloud play is a bit like investing in a real estate developer, you know, in a highly levered format. And if you get it right, you know, the levered equity returns are enormous. If you get it wrong, you can get wiped out.
Bill Ackman, October 22, 2026, CNBC
Stephen Schwarzman owns the real estate. On Bloomberg's Leaders, published October 25, he said the data center company Blackstone bought in 2021 for $10 billion is now 15 times larger.
We're the largest owner of data centers in the world. That puts us in touch with the six hyperscalers because we work for all of them, so we know what's happening sort of almost everywhere in this ecosystem.
Stephen Schwarzman, October 25, 2026, Bloomberg Leaders
He also called the panic over software companies going bankrupt "way overdone" in the same interview. That is a useful counterweight to Rao, from a man whose private credit funds lend to those companies.

Nvidia already trades as if it is losing the argument
The market has a view. Nvidia's forward multiple fell to about 19 times at the end of September, against a long-run average near 32, which the Motley Fool noted was its lowest in a decade. Patel said Nvidia and Broadcom trade near their lowest multiples ever because investors fear Meta, Google and others will take design in house. Earnings are still growing at 2x a year. The multiple is pricing what happens after.
So who keeps the $40 billion?
For the next two years, the suppliers of whatever is scarcest. Patel ranked substrates first, then memory, then optics, with power behind them, and Rasgon expects supply to stay tight through 2028. Those vendors set the price, and Nvidia passes it on. After that the design margin is exposed. If OpenAI's first chip really beats Blackwell and Anthropic builds one in half the time, Patel's math says a lab building at $20 billion a gigawatt "could pay for 5 gigawatts with 1 gigawatt worth of compute running inference." My read is that Jensen Huang keeps his margin only as long as each generation delivers far more than it costs. The landlords and lenders get paid either way, at 11% to 15%, and Schwarzman and Ackman have both chosen that seat.