Three Key Issues for NVIDIA H200's Export Approval to China

"Under the premise of ensuring national security, the United States is about to allow NVIDIA to export H200 products to customers in mainland China and other regions." Trump announced this "deal" on social media on December 8, U.S. time. In return, NVIDIA needs to give the U.S. government a 25% sales share.

In response, NVIDIA stated: "Supplying H200 to commercial customers is a commendable initiative."

There are three issues worth noting regarding the lifting of the export ban on H200: When will the process be implemented, who will benefit, and what impact will it have on the domestic ecosystem?

The lifting of the ban on H200 has actually been rumored for some time, with previous controversies centered around the tug-of-war between the U.S. executive and legislative branches. In other words, how to alleviate concerns in Congress about "maintaining the U.S. AI leadership capability."

"The U.S. Department of Commerce and Congress may not necessarily welcome this; now that this proposal is on the table, the probability of it being finalized is questionable," said a technology policy researcher when discussing the rumors of the lifting of the H200 ban.

How long will the relevant process take?

From Trump's recent announcement on social media allowing the export of H200, the aforementioned concerns seem to have been alleviated. A crucial point is that in recent times, Jensen Huang has been hinting in various occasions, 'It's uncertain whether mainland customers will want it,' and 'H200 is no longer advanced enough.'

However, Huang was merely embellishing; the more important signal is Trump's official announcement.

Previously, H20 received its export license through a similar announcement by Trump on social media—this indicates that there is basically a consensus among U.S. executive and legislative branches. "Even if someone disagrees, Trump still has tools to impose constraints," emphasized the aforementioned technology policy researcher.

From the rhythm of Trump's announcement, it appears that there is already a consensus on the lifting of the H200 ban; what remains is how to resolve the policy-related processes and execution issues, which will still require time to address.

"It's not as simple as saying it can be sold and then starting to sell; following normal regulations also takes time, and directly granting licenses does not align with export control regulations," emphasized another researcher who has long focused on technology policy.

Is H200's performance lagging behind?

In fact, the consensus on the export of H200 is based on its performance being "relatively lagging."

In November 2023, NVIDIA officially released H200, which will start supplying global customers and cloud service providers in the second quarter of 2024, with mass production expected in late second quarter and large-scale deliveries starting after the third quarter. The price per card is around $30,000 to $40,000, and an 8-card server is approximately $300,000.

This chip uses TSMC's 4N advanced process, with the core being the GH100 GPU, containing 80 billion transistors, a thermal design power of 700W, and equipped with NVLink 4 interconnect technology, featuring 18 links and 900GB/s interconnect bandwidth. The GPU with HBM3e has 141GB of video memory, with a memory bandwidth of up to 4.8TB/s In 2024, the H200 is an absolutely advanced product, with an FP16 computing power of up to 1979T, while the H20, specially supplied for the Chinese market, only has 148T. The FP8 computing power is as high as 3958T, while the H20 only has 296T. Additionally, the interconnect bandwidth of the H200 is also double that of the H20, reaching 900GB/s.

However, by the end of 2025, products based on the Blackwell architecture, such as the B200, will be launched and become industry leaders, with the H200 falling to a secondary position, becoming a "relatively outdated" product.

"Meets expectations," said an industry researcher when discussing the topic of the H200 export ban lift, "Releasing the Hopper architecture while not releasing the Blackwell architecture can still explain to the domestic market, 'We are still one and a half generations ahead,' and Chinese customers can also buy what they want."

Overall, Trump's announcement on social media allowing the export of the H200 essentially eliminates concerns from all parties and fundamentally indicates that the H200 no longer represents advanced computing power.

Can the "thaw" of the Chinese mainland market be achieved?

Of course, the previously mentioned transaction price—25% of sales revenue—has increased by 10 percentage points compared to the 15% for the H20 export license. Based on an estimated quarterly revenue contribution of $10 billion from the Chinese mainland, the approval for H200 exports could bring at least $10 billion in "benefits" to the U.S. government annually.

In terms of benefits, NVIDIA, which has always been the "suffering party," will face a "thaw" in sales in the Chinese mainland market.

Previously, Jensen Huang has stated on many occasions that "the market share in the Chinese mainland is 0." The approval for H200 exports will bring new opportunities, especially compared to the compromised H20, which significantly lags in performance and is more likely to attract customers.

"The models of Chinese customers are all adapted for GPUs based on the Hopper architecture," the aforementioned industry researcher emphasized.

In his view, the current appeal of the Hopper architecture is greater than that of the Blackwell architecture. "The B card has not been adapted; otherwise, the operators, toolchains, and underlying software would all need to be redone, which would require even more engineering effort."

In simple terms, for model developers, migrating from the Hopper architecture to other new architectures requires redeveloping computational modules, building dedicated tool processes, and reconstructing underlying integration programs, all of which demand significant manpower, engineering, and time costs.

From NVIDIA's perspective, the profit margin for H200 sales is also more ideal than that of the H20, which is based on a compromised H100 and increases manufacturing costs, while the H200 does not require any compromises. As an older product, its average gross margin is expected to approach or even exceed 80%.

It can be said that Jensen Huang has been shuttling between China and the U.S., waiting for a piece of good news worth billions of dollars each year. After all, during the past period, NVIDIA has been troubled by the narrative of Google's TPU, even having to personally step in and shout, "We are stronger," "We are not competitors." According to previously released data, in the 2024 calendar year, NVIDIA's total revenue in the mainland China market is expected to be $17.1 billion. If the export of the H200 goes smoothly, revenue from data center products in the Chinese market is expected to reach a new high.

For this reason, NVIDIA's stock surged in after-hours trading on Monday, rising over 1%, with a market value increase of over $45 billion.

However, it is important to note that NVIDIA is not the only company that has received clearance; AMD's Chairwoman and CEO Lisa Su has also confirmed that she has received approval to export MI308 chips to mainland customers.

Regarding the export of the H200, Jensen Huang has expressed a concern—uncertainty about whether mainland customers will want the H200, implying that the performance of this product is no longer advanced.

However, from the perspective of computing power, the H200 clearly has advantages over the previous H20.

Industry estimates generally suggest its computing power is about 8 times that of the H20, and according to FP16 theoretical data, the H200 is also 13 times that of the H20. It should be noted that the H20 already lacks an advantage in computing power compared to domestic products, which also explains why there were no customer purchases of the H20 despite its earlier clearance, and why the U.S. is now proposing a new export plan for the H200.

According to early disclosures in 2025, several major companies placed a total of $16 billion in H20 orders with NVIDIA, but ultimately failed to deliver. With the clearance of the H200, this demand will convert into new orders, to be released in 2026.

"Domestic major companies have suppressed demand for two quarters, and capital expenditures will accumulate into next year, leading to significant demand next year," emphasized the aforementioned industry researcher.

However, the approval for the H200 to be sold to mainland Chinese customers has raised concerns about its impact on domestic products. Several practitioners related to domestic computing power chips have stated that the release of the H200 does not directly conflict with domestic products.

"The H200 has practical application value for Chinese customers; currently, domestic cloud providers are mainly used for training, and the cost-effectiveness for inference is not high. Most domestic AI chips are currently used in inference scenarios," emphasized a computing power distributor.

Han Lijie, a partner at Kaiteng Law Firm, said: "It mainly helps NVIDIA make money; competing with local products is secondary."

Risk Warning and Disclaimer

The market has risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Investing based on this is at one's own risk