Three separate reports landed within about 48 hours of each other on November 2 and 3, 2026, and together they describe the same contest from three angles: who gets to rent advanced AI hardware, what China's leading open-weight lab will train on, and how restricted memory reaches Chinese chip designers. This post separates what each report says from what has actually been confirmed, because as of November 4 none of the three has been confirmed by the companies involved.
Report 1: Nvidia and the 51% rule
What was reported
The Nvidia story originates with the Financial Times, in a piece published on November 3, 2026 under the headline "Nvidia orders data centre clients to reduce China exposure" (reporters Zijing Wu, Demetri Sevastopulo and Eleanor Olcott). The FT is paywalled, so this post relies on summaries of it. Per those summaries, Nvidia has told cloud computing providers that a majority of their clients must be non-Chinese, with providers required to sell at least 51 per cent of their services to non-Chinese customers. Bloomberg, citing the FT's unidentified sources, said the 51% applies to services sold. AFP reported the FT cited five people with knowledge of the matter and that the rules reportedly affect only future contracts, not existing agreements.
What is not confirmed
Nvidia neither confirmed nor denied the report. In a statement to AFP it said: "America cannot win the AI race unless America's AI infrastructure wins worldwide." Reuters said it could not immediately verify the report. The provider list is also thinner than some headlines suggest: the Crypto Briefing write-up names neoclouds in Singapore, Malaysia and Japan and says non-compliant firms risk losing hardware access, but that article names no original source, and the FT text on those specifics was behind the paywall. Treat the country list, the access-loss consequence and the audit details as reported by a secondary outlet until Nvidia or the FT original is cited directly.
The July precedent
The rule does not arrive in a vacuum. On July 14, 2026, Techzine, citing FT reporting based on three people familiar with the matter, said Nvidia had sharply reduced the number of Asian companies authorised to buy its AI processors directly after a new vetting process. More than half of existing customers reportedly failed the first round, with stricter criteria for customers in Singapore, Malaysia and Japan, among others. Nvidia reportedly began inspecting data centres, reviewing contracts and speaking with end users. The same report said the US Commerce Department issued guidance in May on supplying AI chips to foreign subsidiaries of Chinese companies, flagging the risk of Blackwell chips reaching China through Malaysia, and noted that US prosecutors had earlier this year indicted a Supermicro co-founder and several employees over an alleged network that sent AI chips to China via Southeast Asia (the hardware was reportedly worth about $2.5 billion; Supermicro says it is cooperating).
The July reports restricted who may buy directly. The November report, if accurate, restricts who the buyer may then serve.
Metir analysis of the reported sequence
Why the mechanism matters
A vetted-customer list governs the first sale. A percentage rule governs the second: what the cloud provider does with the capacity it rents out. That is a different lever, because renting GPU time to a Chinese firm through an offshore cloud is legal in many cases and is the route the May Commerce guidance reportedly targeted. A 51% floor would not ban Chinese customers; it would cap their share.
Scale check from Nvidia's own filing
FourWeekMBA's reading of Nvidia's Form 10-Q for the quarter ended July 26, 2026 gives context on how much Nvidia has at stake in partner clouds. It lists $36 billion of cloud service commitments to AI cloud partners (typically six years in duration) and maximum gross exposure of $3,529 million on guarantees of partners' data-centre leases. It also lists revenue from China-headquartered customers (including Hong Kong) of $12,430 million out of $177,837 million in the first half of fiscal 2027, which is 7.0% (our check: 12,430 divided by 177,837 is about 6.99%), and $7,880 million of $96,221 million in the latest quarter, about 8.2% (our check: 7,880 divided by 96,221 is about 8.19%). Those filing figures come through a secondary site and should be checked against the 10-Q itself. Two readings are consistent with them: Nvidia has a large financial stake in partner clouds staying in compliance, and direct China revenue is already a small share of the total. Neither reading depends on the 51% report being accurate.
For background on how far Nvidia's position inside China has already moved, see our earlier post on the IDC data on Nvidia's 34% China share.
Report 2: DeepSeek V5 on Huawei Ascend
What was reported
On November 2 and 3, several outlets reported that DeepSeek plans to pre-train its V5 model on Huawei Ascend 950DT chips. The versions differ in sourcing, and that matters:
- STAR Market Daily, citing multiple industry sources, reported the plan to use Ascend chips for V5 pre-training; it was republished by the Geopolitechs newsletter on November 3. That report is unconfirmed.
- Crypto Briefing reported at least 160,000 Ascend 950DT chips at a new data centre in Ulanqab, Inner Mongolia, with pre-training possibly starting as soon as late 2026, and described the fleet as initially intended for inference. It names no original source.
- Wccftech attributed the V5 training plan, a roughly 8 trillion parameter size and the 160,000-chip, 1 GW order (shipments ramping in Q4 2026) to two tipsters on X, and presented it as reported, not confirmed.
Neither DeepSeek nor Huawei has confirmed any of this. One claim in circulation, that DeepSeek's founder said four Ascend 950 chips are needed to match one Nvidia GB300, appears in one outlet only, with no named source, and we could not find it in any other report, so we have left it out of the analysis.
The hardware figures, with attribution
Wccftech reports the 950DT with 144 GB of HiZQ 2.0 high-bandwidth memory and 4.0 TB/s of memory bandwidth, on an SMIC 7nm-class process. Crypto Briefing says Huawei is targeting a commercial release in Q4 2026. These are secondary-outlet figures rather than a published Huawei datasheet.
Pricing: two conflicting numbers
Wccftech reported about 111,000 yuan (roughly $16,000) per chip, which at 160,000 chips gives about $2.56 billion (160,000 multiplied by $16,000 equals $2.56 billion, matching Wccftech's own estimate). TrendForce, citing The Information, reported on the same day that Huawei raised the suggested price of the 950DT to over RMB 250,000 (about $37,000). At $37,000 the same 160,000 chips would cost roughly $5.9 billion (our calculation, 160,000 multiplied by $37,000 equals $5.92 billion). The two price points cannot both describe the same order, so the reported total could differ by more than a factor of two depending on which is right. The spread is itself a reason to treat the headline order value cautiously.
Reported price points: smuggled HBM versus the Ascend 950DT
Approximate US dollars in thousands. Green bars are reported figures; grey bars are Metir arithmetic on the reported per-stack price. All inputs come from unconfirmed reports.
Sources: The Information via Wccftech and TrendForce (Nov 3, 2026); Wccftech (Nov 3, 2026). The 950DT suggested price was reported as above RMB 250,000. Hover a bar for details.
Why pre-training is the hard part
Reports describe DeepSeek struggling with stability and performance when it tried to train on earlier Ascend generations, and Crypto Briefing says Huawei engineers have reportedly embedded with DeepSeek's team. The Geopolitechs piece draws a distinction worth keeping: a Shenzhen research consortium has reportedly completed full-parameter post-training of DeepSeek V4-Pro (1.6 trillion parameters) on Ascend 910C chips, but post-training is not pre-training. Pre-training is the longer, more failure-sensitive phase, so a clean run on Ascend would be a stronger milestone than the post-training result. It is also the milestone that would be hardest to verify from outside, since the evidence would be the finished model and how it was trained, not a shipment record.
For readers following the model side, see our earlier coverage such as DeepSeek V4 Pro.
Report 3: HBM paired with throwaway server chips
What was reported
The third report is often attributed to Tom's Hardware in social posts, but the Wccftech article we could read credits The Information, reporting on November 2, as the origin, and does not mention Tom's Hardware. The Information is paywalled, so the details below come from Wccftech and TrendForce. Wccftech itself labelled the story plausible but with little corroboration.
According to those summaries, standalone high-bandwidth memory (HBM) cannot be exported to China, but export rules judge a packaged component by the processor's computing capability rather than its memory. The loophole is traced to December 2024 rules that exempted HBM already integrated into logic chip products. Smugglers reportedly have packaging firms bond top-tier HBM to deliberately weak, cheap server processors, declare the package under the processor's classification and ship it to Hong Kong. Couriers then reportedly carry the packages into mainland China, where workshops heat them to soften the solder, remove the memory stacks and discard the processor. Improper heating can crack the memory, and smugglers reportedly accept that failure rate.
The economics
One broker told The Information that HBM3E that cost a few hundred dollars a year ago now costs about $4,000. Wccftech says an AI chip package needs six to eight stacks, so memory alone would cost about $24,000 to $32,000 (our check: 6 multiplied by $4,000 is $24,000 and 8 multiplied by $4,000 is $32,000). That puts the memory inside one package in the same range as the higher reported price of the Ascend 950DT itself, which is why the loophole is reported to matter: it concerns the scarcest input to Huawei-class accelerators.
TrendForce adds that Huawei's Ascend 960DT is expected in Q1 2027 with roughly double the memory bandwidth and capacity of the current processor and a heavy reliance on Samsung and SK hynix HBM. It also reports, citing The Information, that US restrictions prevent Huawei from buying enough HBM directly from the Korean suppliers and that Huawei has asked some customers to source memory themselves. On the domestic side, TrendForce cites Reuters reporting that CXMT's HBM3E yields are 20% to 30% versus about 70% for SK hynix. None of this is confirmed by the companies named.
How the three reports connect
NVIDIA
DeepSeek| Report | Layer of the stack | Lever | Status |
|---|---|---|---|
| 51% rule | Cloud rental of Nvidia GPUs | Vendor contract terms, audits | FT report, unconfirmed by Nvidia |
| DeepSeek V5 on Ascend | Model training hardware | Domestic substitution | Anonymous and tipster sourcing, unconfirmed |
| HBM loophole | Memory inputs | Export-rule definitions | The Information via secondary outlets |
Read together, the reports describe pressure on three different points. Access to Nvidia compute through third countries is reportedly tightening. China's flagship lab is reportedly moving its next large run onto domestic silicon. The domestic silicon, in turn, reportedly depends on memory that is itself export-controlled. A tighter first constraint makes the second more attractive and the third more important, which is the loop the reports imply. Whether any link holds depends on facts that are not yet public.
What to watch
- Primary confirmation. Whether Nvidia, Huawei, DeepSeek or the US Commerce Department says anything on the record about the 51% rule, the V5 plan or the HBM routes.
- Scope of the 51% rule. Whether it applies only to future contracts (as AFP reported), how compliance is measured, and which providers are in scope.
- Huawei delivery. Reports tie both the 950DT commercial release and DeepSeek's start date to Q4 2026, so any slip in one moves the other.
- Price resolution. Whether the RMB 111,000 or the RMB 250,000 figure holds, since it changes the cost of the reported 160,000-chip order by more than a factor of two.
- Rule changes on HBM. Whether regulators revisit how packaged memory is classified, given that the reported gap sits in the definition itself.
- The finished model. If V5 ships, its technical report would be the first real evidence of what hardware it was trained on.
Teams building on AI models across providers can treat all of this as a reminder that hardware access shifts underneath the model layer. Model-agnostic tools such as Metir, which let you switch between providers, are one way to keep a workflow stable while the supply picture changes, but the underlying questions here are about chips, not software.
Sources:
- Financial Times (summarised by FourWeekMBA), "Nvidia orders data centre clients to reduce China exposure," November 3, 2026: https://fourweekmba.com/ai-nvidia-tells-data-centre-clients-to-cut-their-china-exposure/
- Crypto Briefing, "Nvidia tells cloud clients to sell at least 51% of services to non-Chinese firms," November 3, 2026: https://cryptobriefing.com/nvidia-cloud-clients-51-percent-non-chinese/
- Techzine Global, "Nvidia cuts number of Asian customers in half due to export controls," July 14, 2026: https://www.techzine.eu/news/privacy-compliance/142878/nvidia-cuts-number-of-asian-customers-in-half-due-to-export-controls/
- Crypto Briefing, "DeepSeek plans to train its V5 model on Huawei Ascend 950 chips," November 2, 2026: https://cryptobriefing.com/deepseek-v5-huawei-ascend-950-training/
- Wccftech, DeepSeek V5 on Ascend 950DT-based Atlas SuperPoDs, November 3, 2026: https://wccftech.com/china-going-for-domestic-full-stack-ai-dominance-deepseek-will-train-its-upcoming-v5-model-from-scratch-on-huaweis-ascend-950dt-based-atlas-superpods/
- Geopolitechs (republishing STAR Market Daily), "DeepSeek Plans to Train Its Next-Generation Model on Huawei Ascend Chips," November 3, 2026: https://www.geopolitechs.org/p/deepseek-plans-to-train-its-next
- Wccftech, smuggled HBM and server-chip loophole (citing The Information), November 3, 2026: https://wccftech.com/smugglers-hbm-throwaway-server-chips-us-export-controls-loophole-china-south-korea/
- TrendForce, "China Reportedly Obtains Smuggled HBM via Server Packages; Huawei 960DT Faces Domestic HBM Shortage," November 3, 2026: https://www.trendforce.com/news/2026/11/03/news-china-reportedly-obtains-smuggled-hbm-via-server-packages-huawei-960dt-faces-domestic-hbm-shortage/
Image credits
- Jensen Huang with students at Stanford, April 30, 2026, by Anderseidesvik, CC BY-SA 4.0, via Wikimedia Commons: https://commons.wikimedia.org/wiki/File:Jensen_huang_stanford_2026-04-30_003.jpg
- Patent Wall at Huawei, Shenzhen, October 19, 2014, by RG72, CC BY-SA 4.0, via Wikimedia Commons: https://commons.wikimedia.org/wiki/File:Sidejo_de_Huawei_en_%C5%9Cen%C4%B5eno_05.jpg