China’s Nvidia Ban Is Pushing Alibaba, ByteDance and DeepSeek Offshore for AI Training

The popular story runs in one direction. Washington restricts advanced graphics processors, so Chinese companies take their training runs to foreign soil. That is true, but it leaves out half the picture. In the last few months Beijing has quietly told its own champions to stop ordering new Nvidia parts and to use domestic accelerators in state funded infrastructure. At the same time, the United States has abandoned the only rule that would have tried to police indirect access to GPUs through cloud and third countries.

The result is a strange equilibrium. The heaviest training runs that depend on Nvidia silicon are drifting to rented racks in Singapore and Malaysia. Inside China, teams led by DeepSeek and Huawei are trying to turn constrained hardware into competitive models. China’s big platforms now live in two systems at once, split between offshore Nvidia and domestic chips that Beijing wants to harden into a self sufficient stack.

Why Alibaba and ByteDance are training offshore

The immediate catalyst was the decision in Washington earlier this year to tighten controls on Nvidia’s China bound chips. A special line of processors that had been designed to sit within earlier rules was caught again. Sales of the H20 to Chinese buyers were frozen, and Nvidia warned investors that billions in revenue were at risk. The message was simple. Any chip that gave China serious training capability would eventually face a curb.

Within weeks, the Financial Times reported that Alibaba and ByteDance had shifted the training of important models to South East Asian data centres. Their Qwen and Doubao systems, among the strongest Chinese large language models, were now being trained in facilities in Singapore and Malaysia that host high end Nvidia hardware. Reuters confirmed the broad picture on the same day, noting that the firms were leasing capacity from foreign owned operators rather than building their own new sites. (Financial Times; Reuters)

This is not a theoretical move. Malaysia’s own trade ministry has acknowledged that it is investigating reports of a Chinese company training large language models on Nvidia equipped servers in the country, after stories of engineers flying in with suitcases of hard drives for training runs. Singapore, meanwhile, has seen a surge of new data centre project finance explicitly linked to artificial intelligence workloads.

From the point of view of the platforms, offshore training has three attractions. It gives them access to the newest Nvidia parts that they cannot easily import. It lets them scale quickly by renting from established operators. And, crucially, it allows them to claim formal compliance with both United States export rules and Chinese data law when the training corpus does not include sensitive Chinese personal data.

How Beijing joined Washington in squeezing Nvidia

What makes this moment different is that Beijing has started to add its own controls on top of Washington’s. For most of the last decade Chinese regulators criticised United States export policy while quietly welcoming access to foreign chips. This year that changed.

According to reporting based on insiders at ByteDance, Chinese regulators have now barred the company from deploying Nvidia accelerators in new data centres. The Information and Reuters both describe a shift in which authorities first urged Chinese companies to stop ordering new Nvidia artificial intelligence chips, then went further by mandating that state funded projects use only domestic processors. ByteDance, which had become Nvidia’s largest customer in China in 2025, found itself suddenly constrained at home. (The Information; Reuters; Al Jazeera)

At the same time the Trump administration in Washington decided to scrap the broad “AI diffusion rule” drawn up under the previous White House. That rule would have tried to regulate global use of certain training grade GPUs, including cloud access in friendly countries, through a tiered system. It would have made offshore leases more complicated for Chinese firms, even when the data centres were in allied states. Under heavy lobbying from Nvidia and major cloud providers, the rule was withdrawn before it took full effect. What remained were direct controls on shipping the most advanced parts to Chinese buyers.

That combination is what matters. Washington still restricts deliveries of true frontier GPUs to China. Beijing is now actively discouraging, and in some cases blocking, fresh Nvidia deployments even when they technically comply with foreign rules. Yet no one has stopped Nvidia from selling its fastest parts into Singapore, Malaysia or the Gulf, nor have regulators barred Chinese firms from renting time on those racks.

In other words, it is Beijing as much as Washington that is now nudging Alibaba and ByteDance to look for offshore training sites. The more domestic regulators insist that new national infrastructure must run on Chinese accelerators, the more attractive it becomes to put Nvidia based training runs in a neutral jurisdiction where neither capital controls nor industrial policy are as strict.

Singapore and Malaysia as GPU havens

Singapore’s advantages are structural. It has long been the regional hub for Alibaba Cloud and other Chinese platforms that serve customers in South East Asia. It has dependable power, predictable regulation and a financial sector willing to fund very large data centre projects. The Straits Times has already described a boom in data centre related finance, with AI workloads as a central driver.

Malaysia has followed a different path. It is pitching itself as a data centre and chip assembly hub, welcoming investment from United States and Chinese firms at the same time. The Wall Street Journal and regional press have reported on Chinese engineers who flew to Malaysia with cases of storage drives to run training jobs on Nvidia platforms, prompting Malaysia’s trade ministry to open an inquiry. Within months, Kuala Lumpur moved to tighten controls on the movement of United States made AI chips in and out of the country, seeking to close loopholes and avoid a backlash from Washington. (WSJ; Reuters; Arab News; The Edge Malaysia)

For now both countries sit in an uncomfortable middle. They welcome investment and cloud infrastructure that serves South East Asian clients. At the same time, they know that excessive reliance on Chinese demand for Nvidia compute could paint them as transit points for sanctions evasion. Their regulators are walking a narrow line between being open for business and being drawn into the United States China chip confrontation.

DeepSeek, Huawei and the limits of domestic chips

The one Chinese firm presented as an exception in the original Financial Times story is DeepSeek. Unlike Alibaba and ByteDance, DeepSeek is said to be training its leading models entirely inside China. It was able to do that because it acquired a significant stockpile of Nvidia hardware before the most recent export controls took effect, and because it has formed a tight partnership with Huawei to expand the available domestic stack.

A detailed report from the Center for Strategic and International Studies describes DeepSeek as a kind of stress test for export controls. The lab is credited with training strong reasoning models on what appear to be mid range Nvidia chips, pushed to the limit with unusual levels of efficiency. CSIS and other observers also note that Huawei’s current Ascend line, while improving, still struggles to match Nvidia in the heaviest training tasks, even if it is increasingly used for inference. (CSIS; MIT Tech Review; ORF)

What emerges is a hybrid picture. DeepSeek squeezes legacy Nvidia stockpiles and as much domestic compute as it can reach. Huawei embeds engineers inside DeepSeek’s headquarters in Hangzhou to tune frameworks and software so that more work can shift to Chinese accelerators over time. Government money flows into new foundry investment and subsidy programmes aimed at catching up on the hardware side.

Yet none of that removes the incentive to look abroad when the heaviest training runs arrive. When models move from tens of billions to hundreds of billions of parameters, and when training runs must be repeated again and again to refine behaviour, the advantages of high end Nvidia clusters become difficult to ignore. That is why talk of self reliance in Beijing runs in parallel with concrete moves by other firms to rent foreign capacity.

Forced offshore, forced to innovate

For policy makers in Washington, DeepSeek has become a Rorschach test. Some argue that its success proves that export controls have failed, since a top tier Chinese model appeared despite the rules. Others, especially in the CSIS and think tank community, draw a more mixed conclusion. They argue that DeepSeek has exploited gaps in earlier rules and legacy supplies, but that the available evidence still suggests a serious drag on China’s ability to scale.

The tension for Beijing is just as sharp. By barring fresh Nvidia deployments at home, regulators increase the short term pain for their own platforms. ByteDance and others must now pay to train abroad, accept higher latency for some cross border workloads, and shoulder political risk in host countries. At the same time, this pain is exactly what Beijing wants to create in order to force a shift toward Huawei and other domestic suppliers. The system is being pushed to innovate under constraint.

That innovation has several faces. Chinese engineers are trying to make models more efficient on a given number of chips, experimenting with new sparsity approaches and weight sharing. They are expanding the role of domestic chips in inference, where power efficiency and memory rather than absolute training throughput often dominate. And they are building a network of offshore training options that can be turned up or down as politics allows, treating foreign GPU clusters as a complement rather than a replacement for domestic infrastructure.

Seen from that angle, the offshore move is not simply an act of defiance against United States policy. It is also an unintended tool of Beijing’s own industrial strategy. By raising the cost and complexity of access to Nvidia, both capitals are collectively nudging Chinese firms toward a future in which the country’s largest models are no longer wholly dependent on foreign accelerators.

What this means for the next phase of the AI race

Three conclusions follow from this double squeeze. The first is that export controls are not a binary success or failure. They have not stopped China from building strong large language models. They have made that work more expensive, more fragile and more reliant on a patchwork of offshore capacity. The offshore trend described by the Financial Times and Reuters is a symptom of that friction, not evidence that the rules do nothing.

The second is that jurisdiction has become as important as hardware. The same Nvidia accelerator can be a forbidden object in one context and a permitted tool in another. For the moment, Singapore and Malaysia sit in the zone where Chinese firms can still reach that compute at scale. Their choices about screening and cooperation with United States agencies will matter as much as anything written in export control schedules.

The third is that China’s path now runs through a deliberately awkward mix. Offshore training on Nvidia, guarded by two sets of regulators. Onshore inference and some training on Huawei and other domestic chips, guarded by industrial policy. A domestic firm like DeepSeek sits at the hinge point of this system, using every watt of legacy foreign hardware it already owns while helping to bring an indigenous stack up to speed.

None of this guarantees that the United States will keep its lead in artificial intelligence, or that China will effortlessly close the gap. It does mean that China’s big platforms are being forced offshore by Beijing almost as much as by Washington. And it means that every serious training run that leaves Chinese territory now carries with it a reminder of the uncomfortable truth at the heart of this race. Even for a state that prizes control, sovereignty in artificial intelligence is being built on rented racks in other people’s data centres.

References

This reference list is indicative rather than exhaustive. It focuses on primary reporting and technical analysis that substantiate the specific claims made in this article.