Nvidia Can't Fill $35B China GPU Order From ByteDance
Three of China's biggest technology companies have quietly placed what looks like the largest single chip order in Nvidia's history — more than three million workstation GPUs with a sticker price above $35 billion. There is just one problem: Nvidia has told its own partners it cannot guarantee it can build them.
That is the twist hiding inside Reuters' exclusive on Tuesday. ByteDance, Alibaba and Tencent are not buying the datacentre monsters Washington spent four years banning. They are buying desk-side workstation cards — and that is exactly why the order slipped through the net.
The order nobody saw coming
Update, Oct. 21: Since Reuters' exclusive, the specs picture has filled in: Nvidia has quietly listed the RTX PRO 5500 on its official website with preliminary specs — Blackwell architecture, 84GB of GDDR7 ECC memory, 1,398 GB/s bandwidth, a 600W ceiling and PCIe Gen 5 (PNY/VideoCardz put it at 21,760 CUDA cores, the same count as the flagship RTX 5090). Partners were told to expect early samples in December with commercial deliveries from the first quarter of 2027. In China the card is priced around RMB 85,000–90,000 (The Information, via TrendForce) — comparable to Huawei's Ascend 950PR. An Nvidia spokesperson blamed "outdated US export controls" plus China's own limits on US imports; ByteDance, Alibaba and Tencent declined to comment. One squeeze to watch: Reuters reports Rubin's 2027 ramp will absorb much of Nvidia's TSMC allocation, leaving less room for the RTX PRO 5500 — the card that isn't even shipping yet.
Two people familiar with the matter told Reuters that ByteDance, Alibaba and Tencent — the three giants that anchor China's cloud and AI economy — have ordered more than 3 million Nvidia RTX PRO 5500 cards between them, at roughly $11,400 apiece. Do the maths and you are past $35 billion: a figure that dwarfs most countries' annual chip imports and, by Reuters' account, a scale of ordering that had never been reported before.
Total Chinese demand is even bigger. Reuters reports that overall orders from Chinese customers have topped 6 million cards — double what the three giants requested on their own — while Nvidia has privately indicated it expects to make no more than 3 million RTX PRO 5500s for the entire world in 2027. Some partners, the sources said, may receive just 5 to 10 per cent of what they asked for.
Nobody at Nvidia would say any of this on the record. The two sources spoke anonymously because they were not authorised to discuss the orders — which tells you something about how politically radioactive a $35 billion China sale has become.
Not a datacentre chip — and that is the point
The RTX PRO 5500 is not, strictly speaking, an AI datacentre chip at all. It is a workstation GPU: the kind of card that sits inside a powerful desktop used by engineers, designers and researchers. Reuters describes it as among the most powerful Nvidia chips still available to Chinese buyers, and that word "available" is doing an enormous amount of work.
We checked Nvidia's public workstation line-up: the RTX PRO Blackwell family stretches from the entry-level Pro 2000 all the way to the Pro 6000 Max-Q with 96 GB of memory. The 5500 sits in the upper middle of that range — serious silicon, built for desks and studios rather than server halls. Which is precisely why it can be sold where the B300 Blackwell datacentre chips cannot.
This is the loophole, hiding in plain sight. Washington banned the chips that train frontier models in giant clusters. It did not — could not, without crippling every CAD workstation and rendering farm on earth — ban professional graphics cards. So China's tech giants simply ordered the most powerful "professional graphics card" money can buy, three million times over.
What is the RTX PRO 5500? The specs behind the $35B order
The card at the centre of the $35 billion order is a Blackwell workstation GPU: 84GB of GDDR7 ECC memory, 1,398 GB/s of bandwidth, 21,760 CUDA cores — the same count as the flagship RTX 5090, on the GB202 die — up to 600W, and PCIe Gen 5 x16. It sits between the RTX PRO 5000 and 6000 in Nvidia's workstation stack, and Nvidia's own product page lists it as "Coming Soon". The specs matter because they explain the order: this is not a cut-down gaming card, it is a full workstation GPU that happens to sit outside the datacentre export-control line — which is exactly why ByteDance, Alibaba and Tencent want it.
"We can't promise we'll build them"
Nvidia's response, according to Reuters, was to warn customers and partners that it cannot guarantee the orders will be filled. Chinese demand, the company said, far exceeds planned production.
The arithmetic is brutal. Six million cards wanted by Chinese customers; at most three million to be made for the whole planet next year; allocation cuts that could leave partners with a tenth of what they requested. Per dpa-AFX's reporting, early samples are expected in December, with commercial deliveries starting in the first quarter of 2027 — and much of TSMC's capacity is already spoken for by Nvidia's next-generation Rubin platform from 2027 onward.
That last detail matters more than it looks. Every wafer TSMC dedicates to Rubin is a wafer that cannot become an RTX PRO 5500. Beijing's order book is colliding head-on with the foundry industry's tightest bottleneck.
The 5–10% problem: how much of the $35B order Nvidia can actually build
One source says partners may receive only 5–10% of the quantities they requested. Nvidia can produce no more than 3 million RTX PRO 5500 cards globally in 2027, while total Chinese orders have topped 6 million (Reuters via TrendForce). Early samples are expected in December, with commercial deliveries from the first quarter of 2027 — and from 2027, Rubin's ramp absorbs TSMC capacity, squeezing the 5500 further. The $35 billion order, in other words, is a claim on a production line that does not yet exist at the required scale.
The ban that built this market
None of this happened by accident. It is the latest chapter in a four-year game of regulatory whack-a-mole that Washington keeps losing.
In October 2022 the US barred Nvidia's A100 and H100 from China. Nvidia responded with the cut-down A800 and H800. In October 2023 Washington banned those too. Nvidia built the H20, a compliance-designed Hopper chip. In April 2025 the H20 was restricted — then in July the Trump administration agreed to license it again, in exchange for 15 per cent of the revenue. Trump had originally demanded 20 per cent and only lowered it after meeting Jensen Huang; he called the H20 "obsolete" at an August press conference. Nvidia's own filings now show zero H20 sales to China in the second quarter of 2026, with none forecast for the third. The "obsolete" chip, it turns out, was also unsellable.
The H200 got a different deal. Last December Trump announced on Truth Social that he had told Xi Jinping the US would let Nvidia ship H200s to "approved customers in China... under conditions that allow for continued strong National Security," adding that "President Xi responded positively!" and that "25% will be paid to the United States of America." In January, the Commerce Department flipped H200 licensing from a presumption of denial to case-by-case review. The B300 Blackwell datacentre chips, meanwhile, stay banned — "super-duper advanced," as Trump put it in August, being Trump. "I wouldn't make a deal with that."
An Nvidia spokesperson told CNN at the time: "We applaud President Trump's decision to allow America's chip industry to compete to support high paying jobs and manufacturing in America. Offering H200 to approved commercial customers, vetted by the Department of Commerce, strikes a thoughtful balance that is great for America."
So here is where four years of controls have landed: the datacentre chips are banned or taxed at 25 per cent, the compliance chips are commercially dead — and the workstation card nobody thought to restrict is staring at a $35 billion order book.
AI inference, not training: why China wants these workstation GPUs
Why would anyone want three million workstation cards? Because the AI economy has quietly shifted beneath the regulators' feet.
Training a frontier model — the thing export controls were designed to stop — needs the most powerful datacentre silicon on earth. Running a model for millions of users, the inference stage, needs far less: plenty of memory, decent throughput, and above all sheer quantity. Bloomberg has noted that chips like the H20 are "particularly well-suited to the inference stage: the point at which AI models recognise patterns and draw conclusions." The RTX PRO 5500 fits the same mould — and dpa-AFX reports the Chinese orders are indeed aimed mainly at inference.
That is the shoddy logic at the heart of the ban. Washington built a fence around training; the money moved to inference, which is where the actual revenue — chatbots, video generation, enterprise copilots serving hundreds of millions of Chinese users — now lives. The controls did not stop China's AI buildout. They just changed its shopping list.
Meanwhile the domestic alternatives are coming, just not fast enough. Huawei's accelerators are "the closest competitor to Nvidia's and AMD's offerings, though they still lag significantly in terms of performance," in the words of Tomo Kinoshita, global market strategist at Invesco Asset Management — who added that the curbs are "part of the ongoing trade war" but also a response to China's rise in electronics, and "in that sense, it is likely to become a permanent policy." Huawei pulled its "Nvidia killer" Ascend 960 nine months early to the first quarter of 2027, Cambricon's first-half revenue jumped 4,348 per cent, and Alibaba is testing its own AI inferencing chip. But as CNBC noted, Alibaba keeps buying Nvidia anyway — because when you need three million cards next year, hope is not a supply chain.
Washington's cut — the loophole economy
Here is the part nobody in Washington wants to say out loud: the export-control regime has become a business model, and the US government is now a silent partner in it.
Think about what the last twelve months established. Sell the H20 to China? That will be 15 per cent, please. Sell the H200? 25 per cent, announced by the president himself on social media. The Treasury now takes a cut of the very transactions the national-security establishment spent four years trying to prevent — a colossally strange arrangement in which the state profits from the trade it claims to fear.
And the controls keep manufacturing the demand they were meant to kill. Every ban taught China's giants the same lesson: buy whatever is still legal, as fast as possible, before that gets banned too. The A800/H800 rush, the H20 stockpiling, and now a $35 billion workstation order that exists only because the B300 is forbidden — each wave bigger than the last. If this is a strategy, it is a botched one: a policy that produces record orders for the restricted party's suppliers while collecting a commission on the side.
The honest position is the one Kinoshita stated plainly: these controls are permanent. Not a negotiating tactic, not a temporary lever — a lasting feature of the US-China technology relationship. Once you accept that, the $35 billion order stops looking like a loophole exploited and starts looking like the market pricing in permanence: buy everything you can, forever, because the window only ever narrows.
What happens next
Three dates matter. December 2026: early RTX PRO 5500 samples reach Chinese customers, per dpa-AFX. The first quarter of 2027: commercial deliveries begin — into a queue where buyers may get a tenth of what they ordered. And throughout 2027: TSMC's capacity decisions, with the Rubin platform swallowing the wafers these cards would need, while the foundry giant scrambles to add supply, including its $23B bet on Japanese 3nm capacity.
Watch two things. First, whether Beijing blesses these purchases at all: Chinese officials have reportedly told firms that H200-class buys need special approval, and state media has spent months pushing homegrown silicon. A $35 billion vote for Nvidia is also a $35 billion vote of no confidence in the domestic supply chain — awkward for everyone. Second, whether Washington notices the workstation aisle. The controls have a documented habit of expanding after exclusives like this one. If the RTX PRO 5500 ends up on a restricted list, the only winners will be the lawyers.
Nvidia, for its part, has said nothing on the record — which, given the politics, is the only statement it can afford. Nvidia's China business has already shrunk dramatically under successive curbs; a $35 billion order it can only half-fill is both a lifeline and a liability.
Sources: Reuters exclusive (Oct 20, 2026), via The Hindu BusinessLine, Economic Times and Business Standard; TradingView/dpa-AFX; CNN; Caixin Global; Data Center Dynamics; Bloomberg News; The Information; Council on Foreign Relations. Order figures and Nvidia's supply warning rest on two anonymous Reuters sources; colour on samples, delivery timing and Rubin capacity is dpa-AFX's reporting.
By TechPulse USA Newsroom
TechPulse USA's newsroom covers the global AI and semiconductor race — from export controls to earnings — for readers who need the signal, not the noise.
Comments
Post a Comment