For three years, the US has tried to slow China’s advance in artificial intelligence without making it obvious. The tool has been subject to export controls, aimed not at code or talent, but at silicon.
And it all comes down to a single product, NVIDIA’s H200 chip. A single chip has become a tug-of-war between the US and China.
What happened this week shows export controls slowed China’s progress, but also how much leverage Beijing still holds when the rules begin to loosen.
Why an “older” chip is still a big deal
The H200 belongs to Nvidia’s Hopper generation, released in 2024. It sits below the newer Blackwell line, which the US continues to block from China.
That distinction has encouraged the view that the H200 is safe to export, but the data does not support that assumption.
As of late 2025, most of the world’s largest AI computing clusters still rely on Hopper chips. Public disclosures show that 18 of the 20 biggest known clusters run primarily on H100 or H200 systems.
These chips remain capable of training and running frontier AI models well into 2026.
At the system level, the gap between Hopper and Blackwell narrows further. When chips are grouped into large clusters, performance depends on usable compute per dollar rather than headline specifications.
Independent analyses show that H200-based training systems can reach performance close to Blackwell systems at roughly 50% higher cost.
For inference workloads, which are often limited by memory bandwidth rather than raw processing power, the difference can be smaller still.
For China, cost is not the constraint. Access is.
Source: IFP
China’s real bottleneck is not demand
China does not lack AI engineers or companies eager to spend. What it lacks is the ability to produce advanced chips in large numbers.
Estimates from 2025 suggest China’s domestic output of advanced AI processors equals only 1-4% of US production, and is projected to fall further in 2026.
The limits come from restricted access to advanced manufacturing equipment, high-bandwidth memory, and packaging capacity.
This is why US chip sales matter more than they appear. The H200 does not replace Chinese production; it adds to it.
While domestic suppliers like Huawei continue to improve, even optimistic roadmaps do not show a domestically produced chip matching the H200 before late 2027, and not at scale.
In the meantime, each imported Hopper chip expands China’s total compute pool during a phase when compute remains the primary limiting factor.
This is also why the scale now being discussed has alarmed policymakers. According to Reuters, Chinese firms have placed orders for more than 2 million H200 chips, priced around $27,000 each.
That volume exceeds Nvidia’s available inventory and, according to former US national security officials, would roughly match the compute footprint of a typical US frontier AI company.
The dependency argument collapses under scrutiny
Supporters of exports often argue that selling advanced chips keeps China dependent on American technology. Although this sounds reasonable, it does not match observed behavior.
Chinese companies already buy domestic chips and Nvidia chips at the same time. Procurement mandates ensure continued demand for local suppliers regardless of what happens with imports.
Self-sufficiency in semiconductors is not a market outcome in China. It is a political directive.
This is where Jensen Huang’s argument that decoupling is unrealistic collides with reality. Decoupling may indeed be impractical. Leverage through sales is something else entirely.
Selling H200 chips does not delay China’s long-term plans. It overlaps with them.
Chinese firms use Nvidia hardware to improve models now, while continuing to invest in alternatives designed to replace it later.
When those alternatives are good enough, the dependency ends, but the capability does not.
Export controls are working, unevenly
Export controls were never meant to collapse China’s semiconductor industry. They were designed to slow progress at the frontier.
On that measure, they have been effective. Despite repeated announcements of breakthroughs, China has struggled to advance beyond leading nodes with acceptable yields and reliability.
Thousands of semiconductor firms have exited the market in recent years. Promised jumps to more advanced processes have repeatedly failed to materialize on schedule.
These constraints show up clearly in AI outcomes. Today, the primary reason US models outperform Chinese ones is access to more computing.
This week’s US decision adds a new wrinkle. The Trump administration formally approved H200 exports under a complex set of conditions. Shipments must undergo third-party technical reviews.
China-bound sales cannot exceed half the volume sold to US customers. NVIDIA must certify sufficient domestic supply. Chinese buyers must attest that the chips will not be used for military purposes.
A 25% fee will be paid to the US government.
Analysts have already questioned how enforceable this framework will be once chips are routed through cloud providers, with some describing it as a compromise that may be difficult to police in practice.
When approval does not mean access
What happened next matters just as much.
Within days of the US decision, Chinese customs authorities instructed agents that H200 chips were not permitted to enter the country, according to multiple Reuters sources.
Domestic technology firms were summoned and told not to purchase the chips unless absolutely necessary.
The language used was described as severe, amounting to a de facto ban, although officials have not clarified whether the move is temporary or selective.
Some exemptions are reportedly under discussion, particularly for research projects linked to universities.
That mirrors China’s earlier handling of Nvidia’s weaker H20 chip, which the US approved but Beijing effectively blocked, driving Nvidia’s AI market share in China to zero.
The episode reveals a deeper truth. Export approval in Washington does not guarantee access in China.
Beijing retains its own gatekeeping power and appears willing to use it, either to protect domestic champions or to strengthen its hand ahead of higher-level negotiations.
Time is the asset being traded
For investors, the issue is not ideology but compounding advantage.
AI progress responds strongly to scale. More compute enables faster training cycles, more experiments, and deeper inference. Those effects accumulate.
A one or two-year acceleration can echo far beyond the initial hardware sale.
Selling H200 chips compresses the period during which China’s AI sector remains constrained by hardware.
Blocking them extends it. The recent back-and-forth suggests both sides understand this, even if they publicly frame the debate in commercial terms.
The uncomfortable reality is that commercial incentives and strategic interests point in different directions. NVIDIA wants to sell into a market that could generate tens of billions in revenue.
China wants the chips because they remain among the best available. The US gains most by letting time do its work.
The post US approves Nvidia chip sales but China hesitates appeared first on Invezz
