Strip Nvidia, Broadcom Margins and a Gigawatt Costs $20B

SemiAnalysis’s Dylan Patel says the cost of building a gigawatt of AI data centre capacity has risen from about $40 billion to more than $60 billion, and that it would be closer to $20 billion without the margins of the chip design houses. “If you remove all the margin out of the chip design houses, right, the NVIDIA and Broadcom, Marvell, et cetera, you’re actually at like $20 billion a gigawatt,” Patel said, according to a transcript Fireside Alpha posted on X on 25 October 2026.

Patel was speaking on a panel with Bernstein’s Stacy Rasgon at the CEO Summit of the Chinese American Semiconductor Professional Association (CASPA), hosted by Austin Lyons of Chipstrat, which Fireside Alpha wrote up on 24 October 2026. In the same remarks Patel said Anthropic “is starting their chip program, and they want to do it in half the time that OpenAI did.” The figures are Patel’s spoken estimates, not company disclosures.

Business Pill · VERTICAL INTEGRATION

A 45-second Business Engineer explainer of vertical integration: why each layer of a supply chain adds its own margin, and when owning more of the chain pays. It teaches the general idea, not this story’s figures.

What Patel Said

Patel framed the cost point as a chip race. “All of these companies are trying to make chips,” he said, and OpenAI “showed, I think for the first time ever in the industry, that you can make a first-generation chip that’s really good.” Patel said SemiAnalysis tested it: “They let us into the lab. We tested on our open-source InferenceX benchmark, running all these open-source Chinese models, and the performance was better than Blackwell.”

On Anthropic, Patel said “these are aggressive goals, but if models are able to help them, you know, really accelerate this process.” He listed Google “selling TPUs externally,” Amazon’s Trainium, “really quite good now and fully usable in PyTorch and such,” and “Meta’s very aggressive roadmap.”

Patel then linked the margin to inference revenue: “if you were able to do that, then you could pay for five gigawatts with one gigawatt worth of compute running inference.” He added that “these companies just have to run like crazy,” saying “You can’t afford to be 2x performance per dollar per year, because you’re going to get completely smashed by these other companies.”

What Was Posted on X

Fireside Alpha (@firesidealpha) posted on 25 October 2026 a transcript of Dylan Patel saying Anthropic aims to build its chip in half the time OpenAI took, and that without chip designers’ margins a gigawatt costs about $20 billion, not $60 billion. This piece quotes that transcript.

The Cost Figures

In Fireside Alpha’s write-up, Patel says a gigawatt cost “roughly $40 billion” to build in 2024 and “Now it’s north of 60,” and that it is “not clear who’s going to eat all the cost.” The same write-up says Patel sees Nvidia and Broadcom’s margins as safe for “the next two years,” with the period after that an open question.

Sources: as listed in each row. All figures are spoken estimates.

Anthropic’s Chip Effort So Far

Anthropic confirmed a custom silicon team on 5 August 2026, The Next Web reported, citing a company spokesperson’s comments to Business Insider. The spokesperson said Anthropic would co-design hardware and models, and that hardware from AWS, Google, Nvidia and AMD would remain central under a “multi-chip approach.” A job listing for the team sought someone who “has shipped silicon,” with a salary range of $320,000 to $485,000, The Next Web reported.

Our earlier Anthropic chip piece covered the hire of Clive Chan, who previously worked on OpenAI’s chip programme. Patel’s half-the-time goal is his account; Anthropic has not published a chip timeline in the sources we read.

Where It Meets Our Coverage

Our Jalapeño piece reported OpenAI’s inference chip with Broadcom, and that OpenAI ran it on SemiAnalysis’s InferenceX benchmark and reported the results itself. Patel’s account adds that SemiAnalysis ran its own tests in OpenAI’s lab. Our H100 piece tracked rent on Silicon Data’s index at $2.88 per GPU-hour on 24 October 2026. Our DeepSeek piece reported DeepSeek’s figure of US$38 per GPU-hour for one cluster at off-peak prices: two inputs to Patel’s claim that inference revenue can pay for capacity.

The Structural Read

As we read it, Patel ties the chip push to cost: by his estimates a gigawatt went from “roughly $40 billion” in 2024 to “north of 60” now.

As we read it, Patel cites OpenAI as proof that a first chip can work: SemiAnalysis tested it on InferenceX and “the performance was better than Blackwell,” he said.

As we read it, the timing question sits with Anthropic: Patel said it wants its chip in half the time OpenAI took, while Anthropic said in August that a “multi-chip approach” with AWS, Google, Nvidia and AMD would continue.

Three Implications

FOR CHIP DESIGN HOUSES Patel sees Nvidia and Broadcom’s margins as safe for “the next two years,” per Fireside Alpha’s write-up, with the period after that an open question.

FOR AI LABS Patel said OpenAI’s first-generation chip beat Blackwell in SemiAnalysis’s tests, and that Anthropic wants its own chip in half the time OpenAI took.

FOR DATA CENTRE BUILDERS Patel said a gigawatt cost “roughly $40 billion” in 2024 and “Now it’s north of 60,” and that it is “not clear who’s going to eat all the cost.”

The Business Engineer Lens

This story maps onto the Business Engineer framework Tokenomics: The Economics of AI.

The framework puts it this way: the token “is simultaneously the unit of cognition the model produces, the unit of compute the data center serves, the unit of price the lab charges, and the unit of value the enterprise extracts.”

As we read it, Patel’s claim that one gigawatt running inference could pay for five is an argument about the third of those roles paying for the second, and the chip designers’ margin sits inside the second. The framework was published in May 2026 and does not discuss Patel’s remarks.

What Is Not Established

The $40 billion, $60 billion and $20 billion figures are Patel’s spoken estimates. We read them in Fireside Alpha’s transcript and write-up, not in a SemiAnalysis report, and we did not watch the full panel. SemiAnalysis’s test results for OpenAI’s chip are not in the sources we read. Anthropic’s chip timeline is Patel’s account. We did not contact Patel, SemiAnalysis, Anthropic, OpenAI or Nvidia.

The Bottom Line

Patel says a gigawatt of AI capacity has gone from about $40 billion to more than $60 billion to build, and would cost “like $20 billion” without the margins of Nvidia, Broadcom, Marvell and other chip design houses. He says that is why labs including Anthropic are trying to build their own chips, with Anthropic aiming to do it in half the time OpenAI took.