The case for letting China rent American compute
- Source
- video.twimg.com
- Author
- SemiAnalysis
- Date
Context
The argument in this clip is that when Chinese companies rent American AI chips in data centers outside China instead of buying them, much of the money they spend ends up with American companies, and that this could be a reason to allow it. SemiAnalysis posted the clip on X with an edited quotation and labels it "the devil's advocate case"; the speaker also calls it devil's advocate, so it is offered as an argument worth hearing rather than a declared position. The post does not name the speaker, and the clip is a segment of a longer show that has not been identified. The backdrop is that the United States has restricted exports of advanced AI chips to China since 2022, while, as an Epoch AI analysis of chip smuggling notes, current rules generally allow Chinese customers to rent computing power abroad.
The core of the case is that profit stacks up at each layer between the chip and the customer. By the speaker's figures, Nvidia makes about $7 million of profit for every megawatt of computing capacity deployed, a measure of data-center size by the power it draws, and the cloud provider renting that hardware out makes another $7 million to $10 million at today's prices. The clip gives no time period or source for either figure. A company that rents pays both layers of profit. The speaker draws the competitive consequence with ByteDance, TikTok's parent company: if it can rent chips but not buy them, Meta, which competes with it on video recommendation worldwide, potentially pays a quarter as much for computing. The clip was read through an automatic transcript that is garbled around this point, but the speaker appears to contrast renters with large American companies, naming Anthropic, OpenAI, Google, Amazon and Meta, whose costs differ because they pay less of that margin, in some cases by designing their own chips.
The second strand concerns open models, whose weights are published so anyone can run them. The speaker says application companies built on open models together bring in well over $10 billion in annual recurring revenue, and that DeepSeek, the Chinese AI lab, earns less from its models than American hosting companies such as Baseten and Fireworks earn by serving them. From that the speaker concludes that America is making a great deal of money from Chinese models. The $10 billion figure, as stated, covers open models in general, not Chinese ones alone.
The clip makes an economic case and leaves the security case aside. Its figures are the speaker's own estimates, unsourced in the available material, and its reasoning is about who collects the revenue, not how much computing Chinese companies gain by renting it, which is the compute advantage that supporters of export controls, Anthropic in a 2025 policy post among them, say the controls protect. Because only this segment was available, it is not known whether the speaker takes up that question elsewhere in the show.
Checking sign-in…
Loading comments…






