Huawei Just Won a Captive Market It Can't Feed
Nvidia's China exit cost it 2% of revenue. The remote-access rule signed a day earlier cuts off about 40% of China's AI compute and hands Huawei a captive market it lacks the memory to supply.
OPC Community
OPC Community

Nvidia's China exit cost it 2% of revenue. The rule signed the day before will cost Chinese labs far more, and it hands Huawei a problem that subsidies never could.
On Tuesday Nvidia walked away from China and, on the same call, raised its guidance by $14 billion.
Interim CEO Colette Kress told analysts the company has stopped shipping licensed H200s to Chinese customers. Then she guided next quarter to $130 billion with zero China data center revenue in the number. For scale, China bought $2.4 billion last quarter out of $115.8 billion in total sales. The raise alone is almost six times that.
So "Nvidia abandons China" is a strange headline. Nvidia left a room it had already been shown out of. I'm more interested in who did the showing, and in what got locked behind the door.
Beijing Fired Nvidia in January
Go back to last December and the deal Jensen Huang brokered with Trump. Nvidia could sell H200s to China, and Washington would take a 25% cut. Huang sold it as strategy: keep Chinese developers on American chips, because a lab that trains on CUDA stays on CUDA.
Commerce Secretary Howard Lutnick had said the quiet part in public the summer before. The goal, he explained, was to sell China enough that its developers "get addicted to the American technology stack."
Beijing can read. Washington cleared the H200 for export in January, and one day later customs officials in Shenzhen were telling logistics firms they could not file clearance paperwork for the chip. Tech companies got summoned and warned off buying it unless they had to. Suppliers of H200 components paused production.
Jacob Feldgoise at Georgetown's CSET put it flatly this week: China did not generally accept H200 deliveries. That $2.4 billion from last quarter is the residue of a channel China itself had squeezed nearly shut.
Huang built the export channel on the theory that dependence is a weapon. Beijing agreed with him, and that's exactly why it refused the chips.
Kress didn't end the China business on Tuesday. She stopped pretending it was coming back.
The Rule That Matters Was Signed on Monday
The day before the earnings call, the Bureau of Industry and Security published an interim final rule under a title built to be skipped: "Addressing Remote Access to Advanced Semiconductors."
Until this week, a Chinese lab that couldn't buy a Blackwell could rent one. Data centers in Malaysia, Singapore and Thailand filled their racks with chips that were banned in Shenzhen, then leased the compute to companies headquartered there. ByteDance reportedly went through a Singapore cloud provider to reach capacity in Malaysia. Johor turned into a construction site on the strength of Chinese tenants.
ChinaTalk estimates this rented backdoor supplied about two-fifths of China's AI compute.
Now hold that number up against the H200 story. The chip everyone is writing about was 2% of Nvidia's revenue and a rounding error in China's training capacity. The loophole almost nobody is writing about was 40% of the compute.
Under the new rule, any data center, anywhere, needs a license to provide more than 12 million TPP of compute to a China-headquartered company. That threshold works out to roughly 760 H100s, and a serious training cluster is fifty times that. Facilities already serving Chinese customers have until February 17, 2027 to get licensed. Five countries, Malaysia and Singapore among them, now need export licenses for advanced chips at all.
Rented Blackwells were where Chinese labs kept their CUDA habit alive. You train the frontier model abroad on Nvidia, then serve it at home on whatever you can get. As long as that split existed, nobody had to bet a flagship model on Huawei's software stack. They could hedge.
The hedge expires in February.
There's a third door, smuggling, and it's closing for a different reason. Treasury Secretary Scott Bessent gave Nvidia's board ten days to explain whether advanced chips reached China through shell companies, and he asked in public whether the company pushed "these seemingly illicit sales to goose their revenues and their earnings." The new board-level national security committee is the answer to that letter. A gray market needs a seller who doesn't look too hard, and Nvidia's board just made looking hard a fiduciary duty.
A Captive Market With Empty Shelves
Add it up and Huawei gets every Chinese lab as a customer, by law on one side and by customs order on the other. The stock take is that this is a gift.
Some gift. Huawei can't deliver on it. Demand is the one thing it has plenty of, and chip fabrication isn't really the constraint either. The constraint is memory.
Every AI accelerator needs high-bandwidth memory stacked next to the processor. SemiAnalysis estimates SMIC has wafer capacity for over a million Ascend dies a year, but the memory to go with them doesn't exist. CXMT, China's leading memory maker, was reported at roughly 25% yield on its HBM3 this summer. Epoch AI's analysis says Huawei has been living off a stockpile of foreign HBM bought before the December 2024 controls, and that the stockpile runs out this year.
Estimates for Ascend 950 output in 2026 run from 200,000 to 300,000 units. ByteDance alone could absorb that.
Call this what it is. A captive market served by a supplier who can't fill orders is a ration line. For the next eighteen months, the practical effect of this week is that Chinese labs have less compute than they had in October and no legal way to buy more.
The hawks are right about that. If anyone tells you export controls "don't work," send them back to the two-fifths number.
What the hawks get wrong is what a ration line does to the supplier standing at the front of it.
Subsidies Buy Fabs. Bans Buy Bug Reports.
China has poured money into domestic chips for a decade. Until this year, what it got for the money was hardware that Chinese engineers avoided whenever they had a choice.
You can see it in the gray market. When the H200 became legal in January, a Chinese reseller told reporters that customers were cancelling H200 orders. They weren't switching to Ascend. They were chasing smuggled B200s and B300s. Offered a legal, mediocre Nvidia chip and a domestic alternative, the market picked a third option: illegal, better Nvidia chips.
No subsidy fixes that. A grant can build a fab. It can't make an engineer at Alibaba spend six weeks porting a training pipeline to Huawei's CANN toolkit when a CUDA cluster in Johor is a contract away. Software ecosystems get built by users who have no alternative and a deadline. CUDA got built that way too, by a generation of grad students who had one card that worked.
Starting in February, the best-funded ML engineers in China have no alternative and a deadline.
Each of them becomes an unpaid QA department for Ascend. Their failed training runs turn into the bug reports Huawei's compiler team has wanted for years and couldn't get, because the people capable of writing them were busy on Nvidia hardware. And every yuan of guaranteed demand goes into CXMT's yield curve, which matters because memory yield is the kind of problem that improves with volume.
“Washington spent three years trying to deny China chips. This week it gave China's chipmakers the one thing Beijing couldn't buy them: customers with nowhere else to go.”
The part I'm least sure of is the ceiling. Next-generation memory and logic both lean on lithography tools China can't import, and no amount of captive demand conjures an EUV machine. If CXMT stalls, the ration line just stays a ration line and this essay ages badly. My bet is that it doesn't stall, because for the first time the demand side of the equation is locked.
Your Stack Just Picked a Side
If you build on models, this lands on your desk in three ways.
First, Chinese open weights are about to be shaped by scarcity again. The last time Chinese labs were compute-starved relative to their ambitions, DeepSeek shipped a model so cheap to train that Nvidia lost nearly $600 billion of market value in a day. Expect more of that, and less frontier-scale brute force. The gap at the top widens. Meanwhile the cheap tier you run in production keeps getting better, and it comes from the people with the strongest incentive on earth to squeeze a GPU.
Second, the reference implementation will move. Today every major Chinese open model lands on Hugging Face running on CUDA first, because that's what it was trained on. Within two years I expect a flagship Chinese release tuned for Ascend first, with the CUDA path as the port. If your product depends on those weights, you inherit someone else's second-priority backend.
Third, there's a short-term trade. The Blackwells racked in Johor don't evaporate on February 17. They need tenants who aren't headquartered in China, in a region that just lost its anchor customers. If you buy compute, the first quarter of 2027 is a good time to ask Southeast Asian providers what their capacity costs now.
The bigger shift is harder to price. For fifteen years, "AI hardware" meant one software stack, and you never had to think about it. That assumption ended this week. There are now two stacks, each with a home market of over a billion people and a government making sure the other one can't get in.
Nvidia gave up $2.4 billion a quarter and its stock barely noticed. That's the number in the headlines.
The number that matters is the one Huawei can't ship yet.
Frequently Asked Questions
Why did Nvidia stop selling H200 chips to China?
Nvidia's interim CEO Colette Kress said the company has stopped shipping licensed H200s to Chinese customers and guided next quarter to $130 billion with zero China data center revenue. Beijing had already squeezed the channel nearly shut: a day after Washington cleared the H200 in January, Shenzhen customs blocked clearance paperwork and tech companies were warned off buying it. China was $2.4 billion of $115.8 billion in sales last quarter, about 2% of revenue.
What does the BIS "Addressing Remote Access to Advanced Semiconductors" rule change?
Any data center, anywhere, now needs a license to provide more than 12 million TPP of compute (roughly 760 H100s) to a China-headquartered company. Facilities already serving Chinese customers have until February 17, 2027 to get licensed. This closes the rented-compute route through Malaysia, Singapore and Thailand, which ChinaTalk estimates supplied about two-fifths of China's AI compute.
Can Huawei's Ascend chips replace Nvidia for Chinese AI labs?
Not yet. The constraint is high-bandwidth memory: CXMT was reported at roughly 25% yield on HBM3, and Huawei's stockpile of foreign HBM is expected to run out this year. Ascend 950 output for 2026 is estimated at 200,000 to 300,000 units, which ByteDance alone could absorb. The article's bet is that captive demand will force Chinese engineers onto Huawei's software stack and improve it over time, though it flags lithography limits as the open risk.
Join the OPC Community
Connect with solo founders worldwide. Get daily insights, curated opportunities, and peer support.
Join free