The Financial Times reported on November 3, 2026 that Nvidia has told its data-centre and cloud clients to reduce their exposure to Chinese customers. Bloomberg, summarising the FT, added a specific number: Nvidia reportedly told customers to ensure at least 51% of their services are sold to companies outside China.
That is the whole confirmed shape of the story. Reuters carried the headline and said it could not immediately verify it. We found no on-record Nvidia statement and no written policy. This post separates what is reported, what is inferred, and what is missing.
Quick facts
| Question | Answer |
|---|---|
| Who reported it? | The Financial Times, citing people with knowledge of the matter, on November 3, 2026 |
| What is the claim? | Nvidia told cloud and data-centre clients to cut their exposure to Chinese customers |
| The 51% figure | Relayed by Bloomberg from the FT: at least 51% of services sold to non-Chinese companies |
| Has Nvidia confirmed? | Not in anything we could read. Reuters said it could not verify |
| Is it a law? | No. It is reported as an Nvidia instruction to customers |
| How was it communicated? | Described by some outlets as verbal. The FT text is paywalled, so we could not check |
| What is unclear? | How the 51% is measured, which customers it covers, how it is enforced, and whether it is new |
What exactly was reported?
The FT story is behind a paywall, so we rely on outlets that summarise it. They agree on the core point. Nvidia, the dominant supplier of chips used to train and run large language models, has pushed its cloud customers to shift business away from Chinese buyers.
Bloomberg gives the 51% number. Other aggregators repeat it. One newsletter, FourWeekMBA, quoted the FT as saying that a majority of clients must be non-Chinese, and noted that the sentence carrying the 51% figure was cut off in the text it could read. So even the measure the number applies to is not settled in public.
We use Bloomberg's phrasing, not a stronger one: at least 51% of their services sold to companies outside China. We could not tell whether that means revenue, GPU hours, number of customers or contracted capacity.
Why Reuters' caveat matters
Reuters wrote that it could not immediately verify the FT report. That is standard wording when a wire service relays a competitor's scoop. It is not a denial. It does mean there are not yet two independent confirmations. Treat the 51% number as the FT's reporting, relayed by Bloomberg, until Nvidia or a customer speaks.
Why would Nvidia do this?
Nvidia has not said. The sources do not give a reason, so what follows is context, not fact.
Export controls limit which chips can be sold to Chinese buyers. They do not cleanly cover renting compute. A GPU can sit in a data centre outside China and still serve a Chinese customer over the internet. If a cloud provider's customer base becomes mostly Chinese, that looks to regulators like indirect access to restricted hardware.
Nvidia has reasons to avoid that picture. Our earlier coverage shows how much attention the channel has drawn:
- A Super Micro contractor pleaded guilty in an alleged $2.5 billion Nvidia server diversion case, with a co-founder still facing trial.
- Reuters reported that ByteDance, Alibaba and Tencent ordered 3 million RTX PRO 5500 cards, though Nvidia might deliver only a fraction.
- The Pentagon's CTO floated export controls on AI labs, which we covered as claimed versus verified.
A customer-side rule would let Nvidia show it limits indirect access without waiting for a new law. That is our reading of the incentives, not something any source states.
What does a 51% floor do and not do?
Taking the report at face value:
| It would | It would not |
|---|---|
| Cap the share of a provider's business that can go to Chinese buyers | Ban Chinese customers from renting overseas GPUs |
| Push neoclouds toward customers in other regions | Stop a Chinese firm from using a different provider that stays under the cap |
| Give Nvidia a contractual or informal lever on allocation | Change what chips may legally be exported |
A majority rule is a blunt tool. A provider that is 49% Chinese would be compliant, so large rental volumes could still flow to Chinese buyers. It is a throttle, not a wall.
Who feels it first?
If the instruction is real, the exposed group is cloud providers whose customer base leans Chinese or who serve Chinese firms operating abroad. We have been tracking pressure on that group from other directions: Gartner's forecast that most neoclouds could vanish by 2030 is in our neocloud shakeout post, which also notes how much Nvidia depends on them.
Chinese demand for domestic silicon is also growing. Moore Threads, for example, reported huge growth and a planned 100,000-GPU cluster, as in our Moore Threads Q3 post. China also proposed its own export catalogue for AI technologies, covered in Announcement 48. Control tightening from one side tends to speed substitution on the other.
A short timeline from our own coverage
The FT report lands in a busy month for AI chip policy. These are the dates of our posts, each built on the sources named in it:
- October 11: the Super Micro contractor guilty plea in the alleged server diversion case.
- October 16: China's proposed export catalogue covering 23 AI technologies.
- October 19: the Pentagon CTO's remarks about export controls on uncooperative AI labs.
- October 22: Reuters on Chinese giants' large orders of Nvidia workstation-class cards.
- October 28: Moore Threads' quarterly results and the planned domestic GPU cluster.
- November 3: the FT report on Nvidia's instruction to cloud clients.
None of these items proves a link to the others. Together they show attention on chips moving, chips being used, and chips being replaced, from three directions at once: law enforcement, government policy and the vendor itself.
Claimed versus verified
| Claim | Status |
|---|---|
| Nvidia told cloud clients to reduce China exposure | Reported by the FT; relayed by Bloomberg and Reuters; Reuters could not verify; no Nvidia confirmation found |
| The floor is 51% non-Chinese | Bloomberg's figure from the FT; the measure is not described in anything we could read |
| It was delivered verbally | Stated by some secondary outlets; not checked against the FT text |
| It stems from US government demands | Not established in any source we read |
| It bans Chinese customers | Not claimed by the reports |
| It applies to all Nvidia cloud customers | Not established; scope is unclear |
Be careful with anything that goes beyond that table. Several aggregators added background statistics about past customer-list cuts and compliance inspections that we could not trace to a primary source, so we left them out.
What this means for developers who rent GPUs
Nothing has changed for you today. No provider has announced new terms, and no Nvidia policy has been published. Still, three practical points follow if the report holds:
- Check your provider's customer mix risk. A neocloud with a high share of Chinese customers could be pushed to rebalance. That may show up as capacity changes or pricing.
- Read the acceptable-use and export-compliance terms in your cloud contract. Providers often restrict access from sanctioned or restricted parties and may add verification steps.
- Plan for portability. If your workloads depend on one small provider, keep a second option. Our GPU price tracking post shows how fast rental markets move.
If you are a Chinese-based developer or a company with Chinese ownership, expect questions during onboarding. We cannot say more without the actual terms.
How to check the story yourself
- Read the FT original if you have access, and look for the sentence carrying the 51% figure and what it measures.
- Watch for an Nvidia statement or a customer disclosure. Public cloud providers file results; a change in China revenue disclosure would be a signal.
- Compare outlets. Bloomberg and Reuters both cite the FT. Wait for independent reporting from a source that does not trace back to the FT.
- Look for government comment. The US Commerce Department or Chinese ministries could confirm or contradict.
We will update this post when any of these arrive.
What happens next
Three outcomes are plausible, and none is confirmed. Nvidia may confirm the policy and publish the rule. It may deny or soften it. Or the story may fade with no public follow-up, leaving customers to deal with it privately. Because the instruction was reportedly verbal, the third outcome is the likeliest to leave no paper trail.
For readers tracking AI hardware policy, the larger point is that chip controls are shifting from "who can buy the chip" to "who can use the compute." A rule like this, if real, is an attempt to close the rental loophole by contract. Whether it works depends on enforcement that nobody has described yet.
Related reading
- Super Micro contractor pleads guilty in Nvidia server smuggling case
- ByteDance, Alibaba and Tencent order 3 million Nvidia RTX PRO 5500 cards
- Emil Michael on export controls for AI labs
- China Announcement 48: AI export catalogue
- Moore Threads Q3 and the 100,000-GPU cluster
- The neocloud shakeout and Nvidia exposure
- Jensen Huang on halting chips to an unsafe AI customer
Reports reflect November 3, 2026. The FT article is paywalled and Nvidia has not been quoted; details may change.