Rent, Not Ban: Patel Puts Nvidia Profit at $7M a Megawatt

Dylan Patel of SemiAnalysis argued on an emergency episode with ChinaTalk, published on 13 November, that the United States earns more by letting Chinese companies rent American-run compute abroad than by banning it. In Patel’s estimate, Nvidia makes “$7 million of profit per megawatt that gets deployed” and the cloud provider renting it out makes another $7 million to $10 million. Patel called the MATCH Act, which targets chipmaking tools, “far more important for US national interest” than RASA, the remote-access bill.

Congress has moved the other way so far. The House passed RASA 369 to 22 on 12 January, and the bill has sat at the Senate Banking Committee for 305 days since. The House version of MATCH cleared the Foreign Affairs Committee 36 to 8 on 22 April and has had no floor vote in the 205 days since, by our count of the bill status files. Our map of all 15 chip export bills sets out the rest.

Business Pill · Gross Margin Explained

An evergreen Business Engineer explainer on gross margin, what a company keeps from each dollar of sales, the idea behind the margin stack discussed here. It teaches the concept only and does not discuss Nvidia, export bills or this episode.

The Case for Letting China Rent

Patel set two goals for US policy: to “maintain a lead on AI as large as possible” and to “control China’s access to AI as much as possible”. A total ban works against both, in Patel’s view, because China can answer it through the supply chain: “the largest PCB supplier to Nvidia is a Chinese company”, Patel said.

Patel put the risk plainly: “if you ban chips completely from China, including rental and deployment, there is no reason for China to not just blow the whole thing up”. The question for policy, Patel said, is “what is the minimum amount of compute that China can have with the most amount of American oversight”.

Patel’s answer is rental through US companies outside China, where America keeps the power to say “turn it off at any moment”. Cutting off remote access, Patel argued, adds to the incentive to smuggle chips and to build a domestic industry. “They’re going to do that anyways, by the way”, Patel added, “which is why you need the match act”. Patel summed up: “the best thing for America to do is allow remote access of GPUs because the alternatives are much worse”.

The Margin Stack Behind the $7 Million

The second argument was about money. If Chinese firms rent compute outside China, Patel said, “they are now paying humongous margin dollars to the US ecosystem”, with “multiple layers of margin stacking” on each megawatt. Patel put Nvidia’s profit at $7 million per megawatt deployed and the cloud provider’s at another $7 million to $10 million “if it’s selling at today’s pricing”. Together that is $14 million to $17 million per megawatt, by our arithmetic.

Patel gave a competitive example: if ByteDance rents the chips rather than buying them, then “meta has a fourth of the cost of compute potentially” when the two compete on video recommendation.

Nvidia reports its margins by quarter. In its 10-Q for the quarter to 26 July, gross margin was 75.0% on revenue of $96.2 billion, and Data Center revenue was $89.0 billion. If Patel’s $7 million were gross profit at that margin, it would match about $9.3 million of Nvidia sales per megawatt, by our arithmetic. Patel did not say on the episode which measure of profit was meant, and the arithmetic does not test the estimate.

The same filing shows how little of that business now comes from China directly. Shipments of Data Center Hopper products to China were “less than 1% of Data Center revenue” in the quarter, it says. On RASA, Nvidia told investors the bill “could prohibit the provision of cloud services to any company with an ultimate parent headquartered in China” and “could have a material impact on our business, operating results, and financial condition”. Our RASA explainer compares the House and Senate texts.

The Pushback: Retaliation and Scale

A co-host challenged the premise that China retaliates. Beijing has been unsure about its companies buying US chips at all, the co-host said: “even when Trump is offering them, the Chinese government is like, I’m not entirely sure we even want these companies to buy it”. In the co-host’s reading, parts of the Chinese government might see a RASA-style ban as a push toward domestic chips.

The co-host also cited a press report that China is “already breaking the trade deal” and argued that “escalatory dominance” should not be assumed. Patel answered that earlier rounds of tit for tat ended in compromise: “Tariffs got reduced”.

The co-host then asked whether the compute now reaching the Chinese ecosystem through Malaysia would do more for the US by “helping OpenAI, Anthropic and Deep Mind make better models”. Patel, playing “devil’s advocate”, set out what Patel called “sort of what Jensen’s argument was to some extent”: US companies earn money serving Chinese open models. Adding up application-layer companies on open models, “it’s actually well north of $10 billion”, Patel said, in annual recurring revenue.

Where the Two Sides Meet: Equipment

Later in the episode Patel turned the argument into a ladder: “Don’t let them buy chips. Make them rent the chips. Don’t let them rent the chips. Make them buy the AI.” The aim, Patel said, is to make China “transfer up the value chain”, where the US has the scale. “Focus on the equipment to make the chips”, Patel said. “That’s what’s more important.”

ChinaTalk’s Jordan Schneider agreed on equipment but put the weight on volume: “it’s a scale game, right? It’s not just that you can make a thousand chips, you need to make a hundred thousand.” China “already can’t buy the stuff from the Americans”, Schneider said, so the point of tool controls is to stop Japanese, Dutch and Korean suppliers from filling the gap.

That is the MATCH Act’s design. It gives allied countries that make chipmaking tools 150 days to adopt US-style curbs on sales to China and names deep ultraviolet immersion lithography machines as covered, as our MATCH Act explainer sets out. Our comparison of three other chip bills covers the licence-review and location-verification routes.

The Structural Read

As we read Patel’s case, it treats rental as a control tool as much as a revenue line: compute that sits in US-run data centres abroad can be switched off, and every megawatt of it carries Nvidia’s margin and a cloud operator’s margin, $14 million to $17 million in Patel’s estimates, by our sum.

As we read the pushback, it disputes the premise rather than the arithmetic: if Beijing does not retaliate, a full ban costs the US less than Patel assumes, and if the contest turns on scale, tool controls matter more than who rents which chips. Both sides of the debate put equipment first.

Three Implications

FOR CLOUD OPERATORS ABROAD Patel’s case rests on US-run data centres outside China renting to Chinese firms; Nvidia’s filing says RASA could prohibit cloud services to any company with an ultimate parent headquartered in China.

FOR CHIPMAKING-TOOL MAKERS MATCH, the bill Patel ranks first, gives allied toolmaking countries 150 days to adopt US-style curbs on sales to China, and has had no floor vote since its 36 to 8 committee vote on 22 April.

FOR NVIDIA INVESTORS Data Center Hopper shipments to China were less than 1% of Data Center revenue in the quarter to 26 July, the 10-Q says; Patel’s $7 million per megawatt is a spoken estimate, not a reported figure.

The Business Engineer Lens

This story maps onto the Business Engineer framework AI’s Geopolitical Chokepoint.

The framework opens: “The race for AI supremacy is being fought in the wrong arena.”

As we read it, the episode’s debate lands in the same place: Patel and Schneider disagree on renting chips but both put chipmaking equipment, the bottom of the stack, first. The framework writes about materials and processing rather than export bills, and does not discuss RASA, MATCH or Nvidia.

What Is Not Established

The per-megawatt figures are Patel’s spoken estimates; the episode gives no method for them, and we have not verified them. The $14 million to $17 million totals and the $9.3 million figure are our arithmetic, and the second assumes gross profit, which Patel did not specify.

Quotes come from YouTube’s automatic captions, which carry no speaker labels. Speaker attribution is ours, from context; where the captions do not make the speaker clear we write “a co-host”. The bill status files list no Senate action on RASA after the January referral and no House floor vote on MATCH. We did not contact SemiAnalysis, ChinaTalk or Nvidia for this piece.

The Bottom Line

Patel ranks tool controls first and remote-access limits last, and puts a price on the difference: $7 million per megawatt for Nvidia and $7 million to $10 million for the cloud operator, in Patel’s estimate, each time Chinese demand is served from US-run data centres abroad. The pushback is that China may not retaliate and that the contest turns on scale. Congress passed RASA in the House 369 to 22 and has not brought MATCH to the floor in 205 days, with 51 days left before the 119th Congress ends on 3 January 2027.