The US Treasury approved shipments of Nvidia's H200 to China. The official statement called it a measured step. But measured steps rarely involve handing a rival a scalpel. The H200 is not H100, not B200. It is a highly specific cut of silicon designed to fit inside a policy box.
Over the past six months, the narrative around AI chips has shifted from 'total ban' to 'selective allowance.' The H200 approval is the clearest signal yet that the US has abandoned the delusion of a complete technological blockade. Instead, it has adopted a regulated leak: let China have just enough compute to keep buying, not enough to eclipse.
H200 is built on Nvidia's Hopper architecture, using TSMC's 4N process (a custom 5nm-class node). The GPU die itself is nearly identical to H100. The differentiator is memory: 141 GB of HBM3e, stacked via TSMC CoWoS advanced packaging. This gives it 4.8 TB/s memory bandwidth—1.7x more than H100. For inference tasks, this bandwidth advantage translates directly into faster token generation. For training, the raw compute (Tensor Core count) remains the same as H100. H200 is an inference butterfly, a training caterpillar.
The permission was granted under the existing export control framework. BIS (Bureau of Industry and Security) evaluated the chip's Total Processing Performance (TPP) and performance density. H200's TPP falls just under the threshold set in the October 2022 rules. It is legal to export. But legality and strategy are not the same.
The core of this story is not a chip. It is a control mechanism. The H200 allows the US to achieve three objectives simultaneously: drain Chinese AI companies of capital by selling them high-margin hardware, limit their ability to train frontier models by capping raw compute, and maintain the appearance of engagement. It is a tariff on ambition.
From a supply chain perspective, the H200 reinforces two critical dependencies. First, TSMC's CoWoS packaging capacity—already stretched—must now serve both Western hyperscalers and Chinese buyers. CoWoS is the single bottleneck for all high-end AI accelerators. Any disruption in Hsinchu stops the world's AI supply chain. Second, HBM3e memory comes predominantly from SK Hynix and Samsung, both Korean. Geopolitical risk clusters around two companies and one island. This is not diversification. It is a house of cards on a ledger of trust.
Centralization Risk Score: 9/10. The H200 supply chain concentrates at TSMC (fabrication and packaging) and SK Hynix (HBM). TSMC itself faces a single point of failure in the Taiwan Strait. No redundancies exist. If the strait freezes, every H200 becomes a paperweight.
Let us be precise about what the H200 cannot do. It cannot train a GPT-4 class model from scratch at competitive speed. The parameter count China's labs aim for—1 trillion plus—demands the full H100/B200 compute density. H200 bandwidth reduces training time for memory-bound operations, but the heavy lifting still requires sheer FLOPs. The US kept the engine, sold the transmission.
The bulls will point to the obvious: China gets a better chip than before. Baidu, Alibaba, ByteDance will upgrade their inference clusters. AI applications—chatbots, image generators, code assistants—will see latency improvements. This is true. But it is a tactical gain, not a strategic one. The H200 creates a ceiling: every model trained on Chinese soil will eventually hit the memory bandwidth limit. Scaling beyond requires a new architecture, which the US controls.
Moreover, the H300 (projected) and Blackwell B200 will not be exported. The US will sell only what it can afford to lose. This is not partnership. It is a leash.
From a financial perspective, Nvidia wins. The H200 carries a premium over the H100 in the Chinese gray market. But Nvidia's largest customer segment (hyperscalers) will still demand H100/B200 for frontier work. The H200 is a second-tier product sold at first-tier prices. Gross margins on these chips will exceed 75% because China is a price-inelastic market. Code does not lie, but the auditors often do.
Security is not a badge you wear. It is a process. The H200 reintroduction highlights a procedural failure: the US export control regime treats hardware performance as a static metric, but AI progress is dynamic. By the time the next batch of controls adapts, China will have absorbed the H200's architectural lessons and optimized its own chip designs. The Huawei Ascend 910C is already targeting H200-level bandwidth. The gap will shrink faster than regulators can update TPP formulas.
What the market has not priced is the asymmetry of dependency. China buys H200 and becomes reliant on TSMC's CoWoS timeline. If TSMC raises prices or delays, Chinese AI companies absorb the cost. Meanwhile, US-based labs run on H100 and B200, not subject to the same scheduling risk. This is not a level playing field. It is a ladder with rungs removed at different heights.
The contrarian argument that resonates: the H200 approval proves that containment is impossible. If the US cannot prevent a modified chip from entering China, it cannot prevent reverse engineering of packaging techniques or memory controller logic. The knowledge transfer embedded in each H200—how to manage high-bandwidth memory, how to cool a 700W TDP package—is more valuable than the chip itself. China's Fabless companies will study the H200's thermal and electrical characteristics. They will copy what they can. The US just provided a reference design.
Takeaway: The H200 is not a solution. It is a hostage negotiation with Moore's Law. Every chip sold to China funds the next generation of Chinese countermeasures. The ledger remembers every exploit. When the next export control revision arrives, the calculus will have shifted. And the industry will once again discover that security is a process, not a badge you wear.