Nvidia: Chinese AI companies rush for global GPU capacity ahead of potential new US restrictions
🔎 Why everyone is rushing, right now
On November 12, 2026, Reuters publishes an exclusive that sums up the week: Chinese AI companies are trying to urgently rent Nvidia computing capacity in data centers located outside China. Not to buy chips — to secure computing hours before the door closes.
Why this timing? Two threats are converging. The espionage scandal surrounding Nvidia CEO Jensen Huang is fueling Washington's desire to strike hard. And the regulatory arsenal is finally targeting the big loophole in export controls: remote capacity rental.
The market didn't wait for official statements. In Hong Kong, Alibaba dropped 7.1%, Z.AI and MiniMax at least 7%. What's at stake this week goes beyond semiconductor geopolitics: it's the question of who will be able, tomorrow, to run the models that matter.
The essentials
- The rush: according to three sources cited by Reuters (November 12, 2026), Chinese AI firms are urgently reserving Nvidia capacity in data centers outside China, fearing new US restrictions.
- The trigger: the scandal surrounding Jensen Huang and the fear that Washington will restrict remote access to advanced chips.
- The framework: the RASA bill, passed by the House in January 2026 by 369 votes to 22, awaits the Senate; the Commerce Department is finalizing a rule targeting 7 additional countries, including Malaysia, Singapore, and Thailand.
- The stock market shock: Alibaba -7.1%, Z.AI and MiniMax -7% or more in Hong Kong.
- The context: more than 6 million chips ordered by Chinese customers, with possible deliveries limited to 5-10% of the requested quantities (Reuters, October 20, 2026).
Recommended Tools
This news story is playing out between states and major labs. But it has a very concrete consequence for developers: scarce GPU capacity is going to cost more, everywhere. If your need is limited to hosting an application that calls models via API, there's no point chasing after H200s.
| Tool | Main use | Price (November 2026) | Ideal for |
|---|---|---|---|
| Hostinger | VPS to host your AI apps, APIs and interfaces | from a few €/month (check hostinger.com) | Deploy without depending on scarce GPU access |
| DeepSeek V4 Pro (API) | Leading general-purpose model (up to 88/100 in general use) | usage-based billing (November 2026, see deepseek.com) | Replace a GPU server with a simple API call |
| GLM-5.1 (Z.AI) | Self-host, open weights | free (open weights), server costs on you | Companies that want to control their entire stack |
| Kimi K2.6 (Moonshot AI) | Self-host oriented toward agentic use | free (open weights), server costs on you | Labs that already have infrastructure |
An honest caveat: open weights don't run on an entry-level VPS. The VPS hosts your application and your business logic; the model itself lives with an API provider or on your own GPU infrastructure.
What the Reuters exclusive says: a race against the clock
According to three sources close to the matter cited by Reuters on November 12, 2026, Chinese AI companies are urgently looking to rent Nvidia computing capacity in data centers outside China. The reason: fear that the scandal surrounding Jensen Huang could push Washington to restrict remote access to advanced processors.
Three distinct moves are taking shape. A major Chinese cloud provider has discussed internally locking up — in the sense of firmly reserving — Nvidia capacity abroad. Companies have reached out to data center operators to lock in reservations. And the smaller labs, for their part, are turning to compute rental providers in Southeast Asia, unable to play in the big leagues.
My take: this isn't panic, it's hedging. When a rule could drop "as early as November," every week of waiting mechanically shrinks the available options. The companies reserving today are above all buying time.
The detail that says it all in the exclusive: a Chinese university linked to the People's Liberation Army bought H200s last month, despite the American blacklist. When controls leak this badly, you can see why rational players are speeding up.
RASA and the seven-country rule: the American hammer rises
The Remote Access Security Act (RASA) is the bill that would give Washington its future lever: it would place authority over remote cloud access to restricted technologies in the hands of the Commerce Department. Passed by the House in January 2026 by 369 votes to 22, it is still awaiting Senate review.
In parallel, The Information reports, relayed by MLQ.ai, that the Commerce Department is finalizing a rule barring Chinese AI firms from remote access to advanced chips via data centers in seven additional countries. Malaysia, Singapore, and Thailand are explicitly named. The announcement could come as early as November 2026.
The Achilles' heel of this rule is its legal foundation. As long as RASA has not been adopted, the department is relying on a contested framework — and one can safely bet on legal challenges as soon as the text is published.
For data center operators in Malaysia, Singapore, and Thailand, the dilemma is brutal: honoring existing contracts with their Chinese clients means losing that business overnight. Some will have to choose between their order book and their access to the US market.
A final signal, too little discussed: Nvidia has asked its cloud partners to keep Chinese business below 50%. The chipmaker is rationing its own exposure, even before the regulator acts.
And the calendar is packed: Nvidia's results on November 17, Tencent's on the 18th, Alibaba's on the 19th. Three numerical readings of the same wall.
Capacity leasing, the loophole Washington wants to close
Renting GPU capacity abroad is precisely the route by which Chinese labs access the latest chips without having to import them, as AsiaOne and The Star summarize, picking up the Reuters exclusive.
The mechanism deserves explaining, because it is disarmingly simple. The chip stays physically in a data center in Malaysia or Singapore. The Chinese client, for its part, pays for computing time consumed remotely. No cross-border transfer of hardware, and therefore — on paper — no violation of export controls.
This is where the American logic becomes legible: if cloud access is restricted, access to high-end compute becomes nearly impossible without importing the chips directly. And importing is exactly what export controls already lock down. The net closes in from both sides.
The Megaspeed case illustrates the gray zone. This Malaysian company is suspected of having helped Chinese clients access export-controlled chips — which it formally denies. Guilty or not, its case shows just how difficult the leasing chain is to audit.
Let me be clear about my reading: export control is always one step behind. We lock the door while the compute goes out through the windows — rented windows, precisely.
Gray market: H200 smuggling gets organized
A direct consequence of the uncertainty: traders and intermediaries have begun stockpiling restricted Nvidia chips, in anticipation of a crackdown on smuggling, Reuters reported on November 12, 2026. When a commodity risks becoming impossible to find, it becomes a speculative asset.
The case of the PLA-linked university says it all: H200s purchased last month despite the American blacklist. This is no isolated incident — it's proof that the circumvention channels already exist and are working.
We documented this in Chip war: the United States authorizes the sale of H200 to 10 Chinese companies, but Beijing blocks deliveries: the H200 has become the chip at the center of every tension — authorized on a case-by-case basis on the American side, held back on the Chinese side, courted everywhere else.
This speculative stockpiling has a side effect that everyone will pay for: every chip locked away in a warehouse is a chip that will never run in any open data center. The scarcity premium is building right now, and it's the legitimate buyers — startups, universities, regional clouds — who will foot the bill.
Hong Kong Sanctions: Alibaba -7.1%, Z.AI and MiniMax -7%
On November 12, 2026, the market made its verdict crystal clear: Alibaba plunged 7.1% in Hong Kong, while Z.AI and MiniMax fell at least 7%. Investors are immediately pricing in the risk that Chinese clouds could lose access to high-end compute.
This plunge comes in an already scorching context. As we detailed in The Chinese AI bubble has already burst: Star50 -40% in 4 months, $480 billion gone up in smoke, the sector had already lost hundreds of billions in a matter of months. When a valuation rests on the promise of compute access, every rumor of restriction becomes a solvency problem.
The week ahead looks decisive, with three events that will tell the same story from three angles: Nvidia's results on November 17 (what the chipmaker can deliver), Tencent's on the 18th (what Chinese clouds can order), and Alibaba's on the 19th (what the markets will still tolerate).
My prediction: expect very cautious language on these earnings calls. No one, whether in Santa Clara or Shenzhen, has any interest in publicly confirming the scale of the blockade.
6 million chips ordered, 5 to 10% delivered: the wall of Chinese orders
ByteDance, Alibaba, and Tencent have ordered more than 3 million RTX PRO 5500 units — over $35 billion at the listed price — and total orders from Chinese customers exceed 6 million chips. The catch: Nvidia doesn't guarantee delivery, and shipments could be limited to 5-10% of the requested quantities (Reuters, October 20, 2026).
To see things clearly, here is the actual status of Nvidia chips on the Chinese side:
| Chip | Status for Chinese buyers (November 2026) | Source |
|---|---|---|
| H200 | Case-by-case licenses; Beijing is slowing deliveries | Reuters |
| B300 (Blackwell) | Banned from sale to Chinese buyers | Reuters (Oct. 20, 2026) |
| RTX PRO 5500 | Compliant with export controls; >3M units ordered (>$35B), deliveries not guaranteed (5-10%) | Reuters (Oct. 20, 2026) |
| Rubin | From 2027 onward, Nvidia is expected to dedicate the bulk of TSMC capacity to Rubin, at the expense of products like the RTX PRO 5500 | Reuters (Oct. 20, 2026) |
The Rubin line is the most strategic. From 2027 onward, Nvidia is expected to reserve the bulk of TSMC capacity for its new architecture — at the expense of the "compliant" products destined for China. In other words: even the approved chips will be rationed, by the simple physics of the fabs.
It's in this context that you should read NVIDIA pours $3.5 billion into MediaTek: when the king of GPUs buys its defense against custom chips. Nvidia is securing its supply chain and protecting itself against custom chips — while rationing China. This is no coincidence; it's a strategy.
6 million chips ordered for 5 to 10% delivered: frankly, these aren't orders, they're lottery tickets worth $35 billion.
The American paradox: cutting off chips, depending on Chinese models
Washington is tightening access to chips while, on the applications side, American companies are sending roughly 46% of their AI tokens to Chinese models. Interdependence is not a future risk: it is already the reality of the market, as shown by 46% of American companies' AI tokens go to Chinese models — the CNBC investigation that changes everything.
On one side, the United States holds the physical bottleneck: the chips, TSMC's capacity, and now remote cloud access. Western frontier models — GPT-5.5, Gemini 3 Pro Deep Think, Claude Opus 4.7 (Adaptive) — run on clusters that Beijing can no longer hope to match in volume.
On the other, China holds a software bottleneck: models like DeepSeek V4 Pro (up to 88/100 in general use), GLM-5.1, or Kimi K2.6 rank high in global benchmarks and are already running in American stacks.
This asymmetry is precisely what makes the situation explosive. Each side can hurt the other, but not in the same way: compute is rationed by law, models are rationed by price and trust.
And that is why the battle over capacity rental is so important. As long as remote access remains possible, China keeps a material lifeline. The day the Commerce Department rule takes effect, the equation tips entirely toward those who manufacture the chips.
What this changes for cloud GPU prices (and for your roadmap)
Mechanically, this rush will push cloud GPU capacity prices upward, wherever Chinese customers fall back to — with Southeast Asia on the front line. Three pressures compound.
First pressure: capacity reserved in a panic is capacity pulled off the spot market for all other tenants. Second pressure: speculative chip stockpiling absorbs hardware that will never make it into an open data center. Third pressure: if the Commerce rule falls, Malaysian, Singaporean, and Thai operators will suddenly lose their Chinese customers — and will pass the bill on to their remaining clients.
For a Western startup, three reflexes are in order. Lock in multi-year commitments while capacity is still available at a reasonable price. Diversify across regions, because a rule can target seven countries overnight. And above all, ask the right question: do you really need a dedicated GPU?
In most cases, the answer is no. An API call to a top-tier model almost always costs less than renting a GPU server, and a Hostinger VPS is enough to run your application, your orchestration, and your interfaces. The scarce GPU should remain reserved for those who train or serve models at scale.
❌ Common Mistakes
Mistake 1: Believing the ban is already in effect
RASA hasn't cleared the Senate, and the Commerce Department rule hasn't been announced yet — it could come as early as November 2026. Confusing a bill, a rule in the works, and an active ban leads to rushed decisions, and even unnecessary spending. The solution: follow official announcements and keep the three levels distinct.
Mistake 2: Sourcing from the gray market
Chips stockpiled by intermediaries sell at a premium, with no warranty or support, and the prospect of a smuggling crackdown turns them into toxic assets. A university linked to the APL buying H200s despite the blacklist proves the legal risk isn't theoretical. The solution: go through licensed channels (H200 on a case-by-case basis) or compliant chips like the RTX PRO 5500 — while accepting that lead times are beyond your control.
Mistake 3: Building your roadmap on a single point of compute access
Whether you're in Beijing or Paris, depending on a single provider, a single region, or a single chip is the real risk this crisis has exposed. The solution: diversify your suppliers, keep a fallback API option, and host the application layer yourself — a Hostinger VPS is enough for the application, even when the model runs elsewhere.
❓ Frequently Asked Questions
What is the Remote Access Security Act (RASA)?
It is a US bill that would give the Commerce Department authority over remote cloud access to restricted technologies. Passed by the House in January 2026 by 369 votes to 22, it is still awaiting Senate review — which leaves the rule being drafted on foreign data centers on shaky legal ground.
Why do Chinese companies rent GPU capacity abroad?
Because renting makes it possible to use the latest chips without importing them: the hardware stays in a data center in Malaysia or Singapore, and only the compute is consumed remotely. If cloud access were to be restricted, this route would close — hence the current rush documented by Reuters.
Which Nvidia chips can Chinese companies still buy?
The B300 (Blackwell) is off-limits to them. The H200 goes through case-by-case licensing. The RTX PRO 5500, which complies with export controls, is seeing massive orders — more than 3 million units from ByteDance, Alibaba and Tencent alone — but Nvidia does not guarantee deliveries.
Is the Commerce Department rule already in effect?
No. According to The Information, it is being finalized and could be announced as early as November 2026. Its legal basis remains contested as long as RASA has not been passed by the Senate. Operators in the seven targeted countries are waiting for the text before rewriting their contracts.
Which stocks dropped after Reuters' revelations?
On November 12, 2026 in Hong Kong: Alibaba -7.1%, Z.AI and MiniMax down at least 7%. The market is pricing in the risk that Chinese clouds will lose access to high-end compute, even as their chip orders already exceed 6 million units.
Should we expect GPU cloud prices to rise?
The mechanics point in that direction: capacity booked in a rush, speculative chip stockpiling, and the potential exit of Chinese customers weakening Southeast Asian operators. For a developer, the workaround remains the API of a leading model or self-hosting open weights like GLM-5.1 or Kimi K2.6.
✅ Conclusion
By urgently reserving Nvidia capacity abroad, Chinese AI companies are merely shoring up against the closing of the last major loophole in export controls — capacity leasing — which Washington is in the process of locking down. What comes next plays out this week: Nvidia's earnings on the 17th, Tencent's on the 18th, and Alibaba's on November 19th.
And if your AI project fits within an application that calls models via API, there's no need to wait for the end of the chip war: a Hostinger VPS and good API access are all you need to get started today.