On one side, sales are banned; on the other, chips are arriving. This chip saga is more surreal than a TV drama.
Reuters reported today that procurement records show Northwestern Polytechnical University, which is on the U.S. Entity List for export controls, still managed to buy NVIDIA H200 chips last month. A tender notice dated October 21 listed two H200s for 874,700 yuan, supplied by a company in Shaanxi—not directly by NVIDIA. On September 4, the university bought another two for 796,000 yuan. If you can’t buy the whole server, rent computing power instead: in January this year, the university signed two rental contracts for two servers, each equipped with eight H200s, to fine-tune models and run reinforcement-learning training. The total came to more than 880,000 yuan. Reuters reviewed 26 procurement and company records and found that this kind of arrangement isn’t limited to just one institution.
Meanwhile, IDC’s figures are even more striking: in the first half of this year, NVIDIA shipped 678,000 AI accelerators in China, down 44% year over year, and its market share fell from 55% last year to 34%. Huawei shipped 572,000, taking a 28% share. A year ago, the gap between the two was 31 percentage points; now it’s just six. More importantly, shipments of non-GPU accelerators surpassed GPUs for the first time. The inference market is rewriting the rules of the game.
My take: export bans can control what’s openly available on store shelves, but they can’t control intermediaries and rental contracts—and they certainly can’t control demand. What really determines the winner isn’t whose name appears on a blacklist, but who can scale up capacity and build out an ecosystem. Huawei’s catching up not because one chip can go head-to-head with NVIDIA, but because inference demand is surging and domestic alternatives are getting real orders. NVIDIA is in an awkward spot, too. The H200 was only cleared for conditional exports to China in January this year, with NVIDIA taking a roughly 25% cut, yet the chips still ended up where they weren’t supposed to. Congress has already asked Jensen Huang to give a briefing on November 18.
What does this have to do with ordinary people’s wallets? Two things. First, the scarcer and more restricted computing power is, the harder it is to bring down AI costs. Price cuts for large models and the spread of AI applications will both be slowed, and ultimately everyone who uses AI for work will pay the price. Second, the tighter the controls, the more people will buy into the narrative of decentralized computing power. That could repeatedly reignite enthusiasm for AI-related tokens—but hype is just hype. Few projects can actually fulfill orders, so don’t get carried away just because you hear the words “computing power.”
So here’s the question: in this chip battle, do you think the winner will be determined by export bans, or by capacity and ecosystems?
$TAO $RENDER
This is not investment advice. Do your own research.
Reuters reported today that procurement records show Northwestern Polytechnical University, which is on the U.S. Entity List for export controls, still managed to buy NVIDIA H200 chips last month. A tender notice dated October 21 listed two H200s for 874,700 yuan, supplied by a company in Shaanxi—not directly by NVIDIA. On September 4, the university bought another two for 796,000 yuan. If you can’t buy the whole server, rent computing power instead: in January this year, the university signed two rental contracts for two servers, each equipped with eight H200s, to fine-tune models and run reinforcement-learning training. The total came to more than 880,000 yuan. Reuters reviewed 26 procurement and company records and found that this kind of arrangement isn’t limited to just one institution.
Meanwhile, IDC’s figures are even more striking: in the first half of this year, NVIDIA shipped 678,000 AI accelerators in China, down 44% year over year, and its market share fell from 55% last year to 34%. Huawei shipped 572,000, taking a 28% share. A year ago, the gap between the two was 31 percentage points; now it’s just six. More importantly, shipments of non-GPU accelerators surpassed GPUs for the first time. The inference market is rewriting the rules of the game.
My take: export bans can control what’s openly available on store shelves, but they can’t control intermediaries and rental contracts—and they certainly can’t control demand. What really determines the winner isn’t whose name appears on a blacklist, but who can scale up capacity and build out an ecosystem. Huawei’s catching up not because one chip can go head-to-head with NVIDIA, but because inference demand is surging and domestic alternatives are getting real orders. NVIDIA is in an awkward spot, too. The H200 was only cleared for conditional exports to China in January this year, with NVIDIA taking a roughly 25% cut, yet the chips still ended up where they weren’t supposed to. Congress has already asked Jensen Huang to give a briefing on November 18.
What does this have to do with ordinary people’s wallets? Two things. First, the scarcer and more restricted computing power is, the harder it is to bring down AI costs. Price cuts for large models and the spread of AI applications will both be slowed, and ultimately everyone who uses AI for work will pay the price. Second, the tighter the controls, the more people will buy into the narrative of decentralized computing power. That could repeatedly reignite enthusiasm for AI-related tokens—but hype is just hype. Few projects can actually fulfill orders, so don’t get carried away just because you hear the words “computing power.”
So here’s the question: in this chip battle, do you think the winner will be determined by export bans, or by capacity and ecosystems?
$TAO $RENDER
This is not investment advice. Do your own research.