In the midst of the geopolitical battle over Artificial Intelligence, NVIDIA is considering increasing the production of its H200 GPUs due to a “very strong” demand from China that already exceeds its current output level. The move, reported by Reuters, comes at a particularly delicate moment: the company is focusing its industrial muscle on the Blackwell generation and the upcoming Rubin, while advanced manufacturing and packaging capacity is more contested than ever.
The key turn hinges on Washington. According to Reuters, US President Donald Trump stated that the U.S. government would allow NVIDIA to export H200 to China, but under a striking condition: charging a 25% “tariff” on those sales. Still, the scenario is far from closed: Beijing has not yet given the green light for purchases, and urgent meetings are underway to decide whether to allow these chips into the country.
Why the H200 is “the desired chip” again in China
The H200 is part of the Hopper generation, and Reuters recalls it entered large-scale deployment “last year.” It is manufactured in TSMC on a 4 nm process, a detail that matters because it links the supply to the same advanced production lines that today are highly competitive among giant computing companies.
Chinese interest has a straightforward explanation: it is, by far, the most powerful chip they can aspire to in the current landscape. Reuters points out that the H200 is approximately six times more powerful (in compute capacity) than the H20, the scaled-down GPU NVIDIA designed specifically for the Chinese market and launched at the end of 2023.
This gap puts many companies at a crossroads: wait for the local ecosystem to reach comparable capabilities or “buy time” with foreign hardware if a regulatory window opens. According to Reuters, companies like Alibaba and ByteDance have contacted NVIDIA with plans to place large orders.
The bottleneck: limited capacity… and new priorities
This is where things get more technical (and more uncomfortable). Reuters notes that there are very limited quantities of H200 in production, because NVIDIA is focused on manufacturing Blackwell and preparing Rubin. If China wants H200 “in volume,” NVIDIA needs not only commercial willingness but also industrial capacity, which depends on TSMC.
Moreover, Reuters highlights a critical nuance: adding capacity for H200 is complicated just as NVIDIA is competing with other players — including Google (Alphabet) — for limited advanced capacity at TSMC. In other words: even if demand exists and policy permits, the factory cannot be stretched like an elastic.
NVIDIA, for its part, tries to downplay fears of a “vacuum effect” that would leave the rest of the world waiting. In statements to Reuters, a spokesperson assures they are managing the supply chain so that licensed sales to China do not impact their ability to serve U.S. clients.
A “conditional” way out: China wants AI… but also domestic industry
The great paradox is that Beijing, while needing powerful chips, has been pushing for months a tech substitution agenda and strengthening domestic hardware. Reuters clearly describes the dilemma: allowing H200 entry could slow the growth of the local industry if clients flood NVIDIA, because domestic chips still don’t reach that level.
Within this context, an idea has emerged that reflects the regulatory balance: in urgent meetings, there was a proposal to require each purchase of H200 to be “paired” with a certain ratio of domestic chips. It would be a kind of “combined purchase” that enables China to absorb foreign power for critical loads while still pursuing the goal of developing local suppliers.
A Reuters-cited analyst summarizes it bluntly: the H200 would perform about 2–3 times better than the most advanced national accelerators, and Chinese demand exceeds local production capacity. This gap explains why cloud providers and large companies are pushing to relax restrictions “with conditions.”
What we know (and don’t) as of today
| Element | What has been reported |
|---|---|
| US authorization possibility | Allow export of H200 to China with a 25% tariff on sales. |
| Chinese status | The Chinese government has not yet approved purchases; urgent meetings are underway to decide. |
| Demand | Chinese companies are trying to place large orders. |
| Production | Quantities of H200 are limited; NVIDIA prioritizes Blackwell and Rubin. |
| Industrial dependency | H200 is manufactured in TSMC on a 4 nm process; advanced capacity is scarce and highly contested. |
| Chinese “balancing plan” | Proposal to link H200 purchases to a ratio of domestic chips. |
Why this matters beyond NVIDIA
This episode leaves several important lessons for the tech sector, even for those not glued to GPUs:
- The AI economy rules: if a “legally accessible” GPU multiplies capacity over alternatives, markets will try to buy it even if the window is narrow.
- The supply chain is geopolitical: it’s not just “selling chips”; it’s who makes them, under what licenses, with what tariffs, and under what usage conditions.
- TSMC’s capacity is the “single pressure point”: when everyone competes for the same bottleneck, the winner is not always the one who wants it most, but the one whose priorities align best with industrial and political goals.
Most importantly, it signals a broader trend: AI is no longer a product, it’s a strategic infrastructure. Every capacity decision at a foundry, every restriction or exception, and every “conditional purchase” formula moves pieces that affect prices, timelines, and competitiveness.
Frequently Asked Questions
What’s the difference between NVIDIA’s H200 and the H20 designed for China?
According to Reuters, the H200 has about six times the compute power of the H20, which is a scaled-down version created to meet restrictions.
Is it guaranteed that Chinese companies will be able to buy H200 in 2026?
No. Reuters indicates that China has not yet approved the purchases, and discussions include conditions like linking them to domestic chips.
Could this demand affect GPU availability in Europe or the US?
NVIDIA states it is managing supply so that licensed sales to China do not impact U.S. customers. Still, Reuters warns that advanced capacity at TSMC is limited and highly competitive.
Why does NVIDIA need TSMC to increase production if the chip already exists?
Because the H200 is manufactured at TSMC (4 nm), and adding capacity involves rebalancing lines in an environment where NVIDIA competes with giants like Google for limited advanced capacity.