H200 China Market Volume Countdown! CUDA Drives Strong Demand NVIDIA (NVDA.US)AI Empire Welcomes "Incremental Good News"

“NVIDIA’s AI chip superpower” NVIDIA (NVDA.US) CEO Jensen Huang said that the chip giant is kicking off mass production of H200 AI training/inference accelerators based on the Hopper architecture introduced in March 2022 for customers in the China market, which also shows that the U.S. chip company’s efforts to return to China—one of the most crucial AI compute infrastructure markets—are making positive progress.

There is no doubt that if the H200 really can flow to the China market at a much larger scale despite an additional 25% fee/U.S. government tariff cost, it would be a substantial incremental positive for NVIDIA’s fundamental growth prospects against the backdrop of the company’s current stock price trading sideways and oscillating. After all, neither NVIDIA’s official quarterly earnings guidance nor the “super AI blueprint” of at least $1 trillion through 2027 that it just unveiled at the GTC conference on Monday has accounted for the China-market revenue outlook.

After Jensen Huang delivered a major keynote at NVIDIA’s GTC conference and made a major global investor announcement of the next-generation AI compute infrastructure—the Vera Rubin architecture AI compute system—at a news briefing on Tuesday local time, Huang said NVIDIA has already obtained U.S. government approval to sell H200 AI chips to “many large customers in the China market,” and is currently in the process of “restarting our large-scale manufacturing.” He emphasized that this outlook is dramatically different from just a few weeks ago.

“Our H200 supply chain is being restarted,” Huang said during an event held at NVIDIA’s annual GTC conference in San Jose, California. The company’s CEO, Jensen Huang, had on the previous day at the GTC keynote ceremony made a major announcement of a series of new products and provided investors with an updated set of guidance on financial fundamentals.

In recent years, NVIDIA has been working to restore its AI chip sales in the China market. Because the U.S. government has imposed long-standing restrictions on chip exports to China, this massive market that NVIDIA once relied on has in fact been effectively shut down for such AI compute infrastructure products in the long term.

H200 under a 25% U.S. government tariff burden

However, since this year began, the Trump administration has started allowing NVIDIA and its strongest competitor AMD (AMD.US) to sell weaker performance versions of AI chips to the China market, but this still requires official U.S. government permission and faces an additional 25% tariff imposed by the U.S. government.

The U.S. government allows NVIDIA H200 to be exported to China under certain conditions, with a 25% fee/tariff as the “exchange condition.” In practice, this is a policy compromise—one that allows exports while using them to collect returns. By contrast, more high-end AI chip products, such as the Blackwell series architecture and AMD Instinct MI450 series, are still regarded as more sensitive technologies under U.S. policy and are therefore not yet included in, or not within, the current export approval scope. This means they are simply not allowed to be exported, and thus such tariff policies do not apply.

It should be noted that this semiconductor tariff policy aimed at NVIDIA and AMD excludes chips used for U.S. domestic data centers, consumer devices, and industrial applications—meaning these tariffs do not apply to H200/MI325X directly deployed for use in the U.S.

At present, NVIDIA has not incorporated any revenue outlook from data center types in the China market into its financial forecasts. The data center business unit is currently NVIDIA’s most core business unit, and it is precisely the AI GPUs provided by that division—H100/H200 and the Blackwell/Blackwell Ultra architectures—that deliver extremely powerful AI compute infrastructure to data centers worldwide.

In a recent earnings call last month, the company said that at the time it had only received a preliminary permit from the U.S. government to ship a small number of H200 AI chips to the China market. Although the H200’s overall performance is far inferior to the Blackwell/Blackwell Ultra architecture AI chips NVIDIA currently uses for training and running large AI models, it remains popular in the China market under sanctions constraints thanks to its strong AI inference performance, the CUDA ecosystem sweeping across the global AI developer community, and its convenient deployment capabilities.

China previously accounted for a quarter of NVIDIA’s total revenue, but now it makes up only a very small portion. Although demand for NVIDIA AI chips remains extremely strong across countries worldwide, this Asian country is unquestionably the largest single-scale semiconductor market globally, which makes it crucial to NVIDIA’s long-term fundamentals flourishing.

NVIDIA obtained verbal clearance from U.S. President Donald Trump as early as December last year to sell H200 to some China customers, but so far the chip company has not confirmed any H200 revenue from the China market from these types of licenses. Washington’s makers of manufacturing and tariff rules have also set several additional hurdles, slowing down the formal approval process, which also makes the possibility of a full resumption of sales without sanctions constraints unlikely.

With Jensen Huang’s latest remarks stating that H200 AI chips are currently in the process of “restarting our large-scale manufacturing,” NVIDIA may soon confirm H200 revenue data from the China market.

Media previously reported that H200 AI chips shipped to the China market need to undergo additional routine inspections by the U.S., and are subject to tariffs as high as 25%. U.S. government officials are also considering limiting the number of H200 AI chips each China customer may purchase to 75,000 chips, with total shipments capped at up to 1 million processors.

Demand for H200 AI chips in the China market is likely very strong itself; the core limiting factor for actual deals is not demand, but U.S. government policy and approval. Some media recently reported that China technology companies’ actual order demand for H200 AI chips, which they relied on starting in 2026, has already exceeded 2 million units, while NVIDIA’s inventory at the time was only about 700,000 H200 chips.

China market—NVIDIA’s major incremental positive

Tuesday, NVIDIA’s stock price fell 0.7% to $181.93 at the close of U.S. trading hours, bringing the stock’s decline since the start of the year to 2.5% and causing it to underperform the S&P 500 index.

From a fundamentals expectation perspective, if the H200 AI chip really can flow to the China market at a relatively large scale, it would be a substantial incremental positive for NVIDIA, especially since China once accounted for about a quarter of NVIDIA’s revenue but now is left with only a small portion. In addition, NVIDIA’s strong performance guidance for the current quarter given in February did not include any China data center revenue outlook, and the company’s recent outlook for China data center revenue is still zero. This means that as soon as H200 shipments begin to normalize—even if it is not fully opened—it will still create an incremental upward revision opportunity for the current NVIDIA valuation model and market growth expectations.

In terms of underlying combined performance, compared with today’s Blackwell—and especially Vera Rubin, which Jensen Huang has just announced will enter mass production by year-end—H200 is already clearly a generation or even two behind. H200 is based on the classic Hopper architecture, with a single-card spec of 141GB HBM3e, 4.8TB/s bandwidth, and about 4 PFLOPS FP8. Meanwhile, NVIDIA has already publicly demonstrated that GB200 NVL72 can achieve up to a 15x performance/revenue opportunity advantage over Hopper H200 in certain inference scenarios. Further, NVIDIA’s official stance for Vera Rubin is a 10x per-watt performance improvement and 10x lower token costs versus Blackwell. But these factors do not appear to prevent H200 from matching the real demand of the China market under U.S. sanctions.

H200 delivers nearly a 6x performance improvement over the earlier NVIDIA AI chip product specifically for the China market—H20. Amid the global AI inference wave, what enterprises truly need is a batch of mature chips that can be deployed immediately, can run large-model inference, and have larger memory and higher bandwidth.

On the AI training side, where NVIDIA AI GPUs almost dominate, companies need stronger generality across AI compute clusters and the ability to iterate rapidly across the entire compute stack. On the AI inference side, after cutting-edge AI technologies are scaled and deployed, unit token cost, latency, and energy efficiency matter more. “The AI inference era has arrived,” Huang said at the GTC conference on Monday. “And inference demand is continuing to rise,” he added.

Therefore, H200’s 141GB HBM3e remains highly attractive for long-context use cases, larger-scale batch processing, retrieval-augmented generation, and enterprise-scale high-efficiency batch deployment of AI inference clusters. Combined with the strong demand drivers brought by the CUDA ecosystem, it is still “high-end usable compute under constrained conditions” for the China market. At the same time, however, CUDA, CUDA-X, out-of-the-box model adaptation, development toolchains, and operations experience significantly reduce migration and deployment costs for China customers.

For Wall Street institutional funds, this is not a grand narrative that “NVIDIA will turn the tables in China.” Instead, beyond the already very strong global AI compute infrastructure main theme, NVIDIA has gained an additional piece of upward demand potential in the China market that may be severely underestimated.

At NVIDIA’s GTC conference in the early hours of March 17 Beijing time, Jensen Huang showcased NVIDIA’s “unprecedented AI compute revenue-super chart” in the AI compute infrastructure space. He told global investors that driven by strong demand for AI compute power from Blackwell-architecture GPU chips and even more explosive demand for the Vera Rubin architecture AI compute system that is set to move into mass production, NVIDIA’s future revenue in the AI chip arena could reach at least $1 trillion by 2027—far higher than the $500 billion AI compute infrastructure blueprint for 2026 that was floated at the previous GTC conference.

As model scale, inference pipelines, and multi-modal/agentic AI workloads drive compute consumption to expand exponentially, tech giants’ capital expenditure main theme is even more inclined to concentrate on AI compute infrastructure. Global investors are also continuing to anchor on the “AI bull market narrative” centered on NVIDIA, Google TPU clusters, and expectations for new-product iterations and AI compute cluster delivery from AMD—keeping it as one of the most certain macro investment narratives in global stock markets. This also means that investment themes closely tied to AI training/inference—such as power, liquid-cooling heat dissipation systems, and optical interconnect supply chains—will continue to remain among the hottest investment camps in the stock market as AI compute leaders such as NVIDIA, AMD, Broadcom, TSMC, and Micron face uncertainty amid geopolitical conditions in the Middle East.

In the view of Wall Street giants Morgan Stanley, Citigroup, Loop Capital, and Wedbush, the global wave of artificial intelligence infrastructure investment centered on AI compute hardware is far from finished—it is only at the beginning. Driven by the unprecedented “inference-end compute demand storm,” the scale of this global AI infrastructure investment wave running through 2030 is expected to reach as much as $3 trillion to $4 trillion.