Nvidia halts China-bound H200 production, shifts TSMC capacity to Vera Rubin

[Daniel Nenni]

Credit: Nvidia

Nvidia has halted production of artificial intelligence (AI) chips intended for the Chinese market and redirected manufacturing capacity at TSMC to its next-generation Vera Rubin platform, as regulatory barriers in both the US and China continue to cloud prospects for sales to Chinese customers, the Financial Times reported.

The shift away from H200 chips toward the Vera Rubin architecture suggests Nvidia no longer expects significant demand for the processors in China in the near term after months of uncertainty over export approvals from Washington and potential restrictions in Beijing.

Regulatory uncertainty drives strategic pivot​

The H200, one of Nvidia's earlier-generation AI processors, had been positioned to comply with US export controls on advanced semiconductors. Vera Rubin, unveiled earlier this year, represents the company's latest chip architecture designed to support more complex AI systems and is expected to see strong demand from major US technology companies, including OpenAI and Google.

Washington has tightened restrictions on exports of advanced semiconductors to China, while Beijing has signaled it may curb imports to support domestic chipmakers.

"Instead of waiting in limbo, Nvidia has to move on to what it can achieve with certainty, especially when there's a shortage of supply for its advanced stuff," one person familiar with the plans said. "This could, in a way, accelerate the Vera Rubin delivery and roll out."

Nvidia had previously lobbied both Washington and Beijing to allow sales of H200 chips in China. After US President Donald Trump indicated in December that such sales could be permitted, the company began ramping up production in anticipation of orders from Chinese customers.

The company had expected demand of more than one million units from China, with suppliers preparing for deliveries as early as March. In early January, Nvidia CEO Jensen Huang said demand for the chips was "very high" and that the company had ramped up its supply chain.

H200 exports stalled despite limited approvals​

The approval process later stalled as US officials sought tighter safeguards to prevent Chinese use of advanced chips in ways that could threaten national security. At the same time, Beijing has considered restricting purchases of H200 chips to encourage local AI developers to adopt processors from domestic semiconductor companies.

During an earnings call last week, Nvidia CFO Colette Kress said that while the US government had granted licenses allowing "small amounts" of H200 chips to be shipped to China, the company had not yet generated revenue from those approvals.

"We do not know whether any imports will be allowed into China," Kress said.

A US Commerce Department official told Reuters last month that none of Nvidia's H200 chips had been sold to Chinese customers despite export licenses. Although the Trump administration formally allowed sales of H200 chips to China in January, shipments remained stalled due to guardrails built into the approval process.

Nvidia has already produced about 250,000 H200 chips. If both Washington and Beijing ultimately allow only limited orders, the existing inventory could be sufficient to meet demand.

China's President Xi Jinping and US President Donald Trump are scheduled to meet later this month, raising speculation that export controls on advanced chips could be revisited. If restrictions were eased, Nvidia could take up to three months to reallocate supply chain capacity to resume H200 production.

Vera Rubin reshapes AI supply chain​

The shift to the Vera Rubin platform is also reshaping the broader AI semiconductor supply chain.

According to Korean media reports, Samsung Electronics and SK hynix are emerging as leading suppliers of high-bandwidth memory (HBM) for Nvidia's next-generation AI accelerators based on the Vera Rubin architecture.

The companies are expected to supply HBM4, the sixth generation of high-bandwidth memory, which Nvidia has identified as a key component supporting the performance of its next-generation AI accelerators.

Industry sources cited by Korean media said Micron Technology was not currently listed among suppliers for the flagship Vera Rubin accelerator, although the company could supply memory for mid-range products within the broader Rubin-series lineup.

Memory suppliers compete for next-generation AI accelerator​

Nvidia's dominance in the AI accelerator market has intensified competition among memory manufacturers seeking to join its HBM supply chain.

The Vera Rubin platform is expected to feature 16 stacks of HBM4 with a total memory capacity reaching 576GB, higher than the 432GB capacity planned for AMD's next-generation MI450 accelerator.

Samsung Electronics has reportedly passed key stages of Nvidia's HBM4 qualification testing, while SK Hynix is continuing optimization work with Nvidia as part of final testing procedures.

Given that HBM4 production typically requires more than six months from DRAM wafer input through packaging, industry observers expect the two companies to begin mass production as early as this month.

Nvidia declined to comment on the report, while TSMC(2330.TW) also declined to comment when contacted by Reuters.

Nvidia halts China-bound H200 production, shifts TSMC capacity to Vera Rubin

Nvidia has halted production of artificial intelligence (AI) chips intended for the Chinese market and redirected manufacturing capacity at TSMC to its next-generation Vera Rubin platform, as regulatory barriers in both the US and China continue to cloud prospects for sales to Chinese customers...

www.digitimes.com

 

[Daniel Nenni]

This one is on Jensen. I thought for sure he had the China situation handled. I guess politics will always beat common sense.

 

[user nl]

This one is on Jensen. I thought for sure he had the China situation handled. I guess politics will always beat common sense.

In this geopolitical "chess" probably not in Xi's interest to make concessions to USA by letting in all these older NVIDIA chips. Especially, now that POTUS shot himself in the foot (or worse).

 

[chillfire]

In this geopolitical "chess" probably not in Xi's interest to make concessions to USA by letting in all these older NVIDIA chips. Especially, now that POTUS shot himself in the foot (or worse).

US always shot its foot according to China. Don't know which foot DJT shoot himself this time

 

[user nl]

US always shot its foot according to China. Don't know which foot DJT shoot himself this time

Some people argue it is the middle-sized one:

https://www.foreignaffairs.com/iran/curse-middle-sized-wars

In 1988, the military historian James Stokesbury observed that democracies are best at fighting either little wars, which are reserved for “professionals” and don’t involve ordinary citizens, or really big wars that mobilize all of society. Those democracies, he continued, have “very real problems trying to fight a middle-sized war, where some go and some stay home.”

Middle-sized wars are big enough to cause immense destruction and bloodshed but small enough that they do not engage the full home front. They should not be confused with what the military theorist Carl von Clausewitz called a limited war, in which the goal may be only to hurt the enemy, not to destroy it. A limited war is by design, whereas a middle-sized war grows out of what was intended to be strictly a small war. Generals and political leaders know what they are doing in a limited war. U.S. leaders in today’s middle-sized wars do not.

It may be uncomfortable to consider the so-called forever wars in the Middle East—which have killed or wounded tens of thousands of U.S. soldiers and left countless dead on all sides—as merely middle-sized. But Stokesbury’s point is one of comparison. The wars in Afghanistan and Iraq, as well as those in Korea and Vietnam, as gruesome as they were, cannot be equated to the two big world wars of the twentieth century. Nor can they be grouped with little wars, such as the invasion of Grenada in 1983 and of Panama in 1989, which made headlines for a few days but were essentially imperial policing actions. U.S. military interventions in Bosnia in 1995 and Kosovo in 1999 also had exceedingly few American casualties and were mainly air operations conducted within strict limits.

For the United States, middle-sized wars present a unique problem. They ruin presidential administrations along with the American public’s regard for the U.S. government’s ability to conduct foreign policy. It would seem that the American people are finished with middle-sized wars and never want to repeat them. In fact, after each of the United States’ recent middle-sized wars, the public and politicians alike declared an end to them. This was especially true after the wars in Vietnam and Iraq, which destroyed the reputations of top policymakers. Yet the United States may be on the brink of another. The Trump administration’s war in Iran has the potential to evolve into a middle-sized war if the clerical regime does not surrender, as U.S. President Donald Trump demands, and continued U.S. and Israeli bombing leads to anarchy in Iran and destabilizes the Persian Gulf. The gap between toppling an existing order and erecting a new, more pliable one can be vast.

 

[siliconbruh999]

There is nothing common between H200 and VR both uses different Node different packing different Arch different HBM the only thing i see is the tooling might have been moved to VR such a bad article saying nonsense for info that is valid