China Turns Trump’s Nvidia H200 Deal Into Another Tool for Self Reliance
by Jaffa Levy
·
Published
· Updated
In our earlier Telegraph Online investigation,
“China’s Nvidia Ban Is Pushing Alibaba, ByteDance and DeepSeek Offshore for AI Training,”
we showed how United States export rules and Beijing’s clampdown drove Chinese giants to train their biggest models abroad while building a second stack on Huawei and domestic chips at home.
This article is the next turn of the screw. Donald Trump has now allowed Nvidia’s H200 chips back into China for approved customers. Beijing’s answer is to ration that access, keep state systems on domestic hardware and use the short term relief to accelerate long term independence.
Trump has confirmed that Nvidia can sell its H200 chips to approved customers in China, under licences cleared by the United States Commerce Department. The H200 is Nvidia’s second tier data centre workhorse: more powerful than the China specific H20, but still a step behind the company’s top Blackwell and coming Rubin lines.
The deal is framed as a trade that secures both American revenue and supposed security. Approval is limited to vetted buyers. Sales attract a twenty five percent levy that Trump says will be paid to the United States. Blackwell and Rubin remain off the table. In political language, this is described as a balanced compromise. In technical reality, it is a partial reversal of the previous blockade that Beijing can exploit.
The H200 decision in one glance
Chip tier – H200 sits below Nvidia’s most advanced Blackwell line but above the H20 that was designed for China.
Licensing – Only approved customers in China can buy, through United States Commerce Department oversight.
Levy – Trump insists that twenty five percent of sales value will flow to the United States, most likely structured as a customs duty.
Scope – No permission yet for Blackwell or Rubin chips, which remain the true frontier hardware.
On Wall Street the decision looks like a win for Nvidia. On the Washington security circuit it is already contested. A group of senators in both parties has tabled a bill that would bar licences for China for thirty months for advanced chips including the H200 and Blackwell.
How Beijing plans to ration Nvidia’s return
The United States can decide which chips leave its ports. It cannot decide who actually uses them inside China. That is Beijing’s move.
Reporting from Beijing describes regulators preparing to allow only limited access to H200 clusters. Purchases would require explicit approval, with buyers explaining why domestic hardware cannot serve their needs. Public sector institutions could be barred from using the H200 entirely, locking state systems into Chinese chips.
The inward control regime on H200
Regulators such as the National Development and Reform Commission and the Ministry of Industry and Information Technology are reported to be shaping the policy frame.
Corporate buyers may need case by case approval, including justification for not using Chinese accelerators.
Public entities could be required to stay entirely on domestic chips, even if H200 units are available in the market.
Energy subsidies and regulatory preference would continue to support data centres that run Huawei and other local providers.
This is the mirror image of United States export controls. Washington restricts what leaves its territory. Beijing restricts what its own institutions are allowed to buy. Nvidia is squeezed between both regimes, tolerated as a tool for China’s private platforms but fenced off from command systems and state infrastructure where strategic data sits.
Why Chinese giants still train offshore even if H200 returns
The first article in this series set out the offshore pattern clearly. As Washington tightened export rules and Beijing restricted fresh Nvidia deployments in domestic data centres, companies such as Alibaba and ByteDance shifted the heavy training runs for their largest models to leased capacity in Southeast Asia.
Singapore, Malaysia and other regional hubs provide a simple route around the choke. The hardware belongs to foreign data centre operators. The Chinese client rents compute as a service. On paper this sits outside the direct scope of United States export licences aimed at China based end users. Offshore training has become a parallel rail.
Offshore training, domestic deployment
Train abroad – Large language models and recommendation engines are trained on Nvidia clusters in Southeast Asian facilities operated by non Chinese companies.
Run at home – Once trained, the models are distilled and redeployed on domestic hardware, often on Huawei’s Ascend line or other local accelerators, inside China.
Data limits – Strict rules on exporting personal data mean that firms either anonymise or synthesise training inputs, or keep the most sensitive data on shore and use hybrid methods.
Regulatory grey zone – The approach preserves formal compliance while undermining the strategic intent of the original export rules.
The return of H200 does not erase this habit. Offshore routes have already been built. Contracts are signed. Engineers are in place. It is safer for large Chinese groups to keep multiple paths open rather than depend again on a single supplier whose access is controlled by Washington and whose domestic use is constrained by Beijing.
How Beijing uses H200 access to harden its own stack
The important question is not whether Chinese firms want to use Nvidia. They do. The important question is how Beijing uses that desire to extract concessions and to buy time for domestic champions.
The H20 episode is the template. Under earlier pressure, the administration in Washington authorised sales of the cut down H20 chip and floated a revenue share to the United States Treasury. Beijing’s regulators responded by questioning security and then quietly guiding big platforms away from the new product, saying its performance was not meaningfully better than Chinese alternatives.
The same logic applies now. Limited H200 access gives Chinese firms a ceiling of foreign performance they can target. At the same time, rationing keeps volume and margin with domestic suppliers for government work and much of the cloud market. In that environment Huawei and other local designers are not just substitutes. They become the default infrastructure for critical systems.
Domestic leverage from foreign dependence
H200 clusters can be concentrated in commercial cloud offerings where global compatibility is useful.
State and public systems stay on Chinese chips, which stabilises demand for Huawei and peers.
Any supply shock, whether from Washington or Nvidia, will hurt the foreign stack first and vindicate Beijing’s insistence on redundancy.
Subsidies and procurement rules can quietly favour local vendors even as H200 is permitted on paper.
This is not romantic talk about self sufficiency. China still trails in many aspects of advanced chip design and extreme ultraviolet manufacturing. But each cycle of restriction and workaround shifts more capital, talent and political will toward breaking that constraint.
Why Congress is trying to slam the door again
The H200 decision has exposed a split inside the United States establishment. The White House and Nvidia frame it as a controlled opening that keeps China just far enough behind. A bloc of senators sees it as the abandonment of what they call a narrow edge in military relevant AI.
The proposed Secure and Feasible Exports Chips Act would force the Commerce Secretary to deny export licences for advanced chips to China for thirty months, specifically targeting Nvidia’s H200 and future Blackwell processors. Similar language reaches for Russia, Iran and North Korea.
Inside the United States backlash
Security focused lawmakers argue that any H200 sales will accelerate Chinese progress in military, cyber and intelligence applications.
Economic voices warn that cutting China off entirely risks pushing it to fully independent chip supply, eroding long term American leverage.
The compromise camp wants to keep China on last generation American hardware while reserving the frontier for domestic use.
Trump’s deal falls in that middle category, but the enforcement and political durability of this arrangement remain uncertain.
For now the bill is a threat rather than law. Many of the same politicians who demanded tougher export rules under the previous administration have been quieter about challenging Trump directly. But the message to Nvidia and its peers is clear. Every licence to China now sits on a moving political fault line.
What this new export control race really means
Strip away the slogans and the pattern is simple. Washington is trying to freeze China a fixed distance behind the frontier in AI hardware. Beijing is trying to reduce the cost of that handicap and shorten the gap over time, while making sure that critical systems do not rely on assets that can be switched off from abroad.
H200 exports under licence, paired with Chinese rationing and domestic preference, are not an exception to that pattern. They are its purest expression. Offshore training shows that even coordinated controls leave gaps, and that firms on both sides will use those gaps while they exist.
The uncomfortable point for policy makers in the West is that this arrangement may actually suit Beijing more over the medium term. China can still tap Nvidia for incremental performance and ecosystem compatibility. At the same time, sustained pressure forces capital and engineering into local champions, builds political support for stricter data localisation and gives Chinese regulators practice at running inward facing control regimes.
The uncomfortable point for Beijing is that its domestic stack still depends on foreign tooling, legacy licences and fragile supply chains for equipment and design software. Rhetoric about complete independence sits far ahead of capacity. The H200 decision buys breathing space, not victory.
For Nvidia this is a short term revenue boost that deepens long term structural risk. The company sits in a gap that is slowly closing. Both Washington and Beijing will continue to treat it less as a private firm and more as an instrument in a widening contest over who writes the rules for the next layer of machine intelligence.
References
This article draws on open source reporting, official statements and prior Telegraph Online investigations. Key sources are summarised below.
Trump’s H200 export approval
Wire and agency reporting on Trump’s decision to permit H200 exports to approved Chinese customers with a levy, while keeping Blackwell and Rubin off limits.
Beijing moves to limit H200 access
Accounts of plans by Chinese regulators to restrict H200 purchases, require approvals and keep public sector procurement on domestic chips.
Offshore training by Chinese platforms
Coverage describing Alibaba, ByteDance and peers training large models in Southeast Asian data centres to access Nvidia clusters and then redeploying models at home.
United States congressional backlash
Reports on proposed legislation to block advanced AI chip exports to China for thirty months, including Nvidia’s H200 and Blackwell lines.
Telegraph Online telegraph.com prior work
Telegraph Online articles including “China’s Nvidia Ban Is Pushing Alibaba, ByteDance and DeepSeek Offshore for AI Training”, “China Turns United States Chip Sanctions Into a Technological Triumph”, “China Strikes Back in United States Chip War With Trade Probe” and our wider work on sanctions, reserves and de dollarisation.
