Nvidia, Santa Clara
Designs without a factory. Nearly all AI-model training rests on a single architecture, whose real volume is capped not by design but by packaging in Taiwan.
Nvidia's lock is software
Nvidia's dominance is not defended by the chip. Rival accelerators exist and work: AMD's MI300, Google's TPUs, Amazon's Trainium. What locks the market is CUDA, the proprietary software layer Nvidia has been enriching for nearly twenty years. The libraries (cuDNN, TensorRT, NCCL) and the whole set of AI frameworks (PyTorch foremost) are optimized for CUDA first; migrating a training run to another architecture costs months of re-engineering and the risk of regression, not just dollars. That exit cost, far more than the purchase price, holds customers in place.
Nvidia then extended that lock from the chip to the entire rack. With the Mellanox acquisition it absorbed the interconnect, NVLink internally, InfiniBand between servers, and now sells the GB200 NVL72 as a single system, not as a GPU. The real product is no longer a component but a cabled cluster whose communication fabric is itself a near-monopoly of its own making. The consequence: even a competing chip at the same node is not enough to break the dependency, because one would also have to replicate the software and the network that make tens of thousands of accelerators behave as a single machine.
A company turned instrument of foreign policy
On 9 April 2025, Washington placed exports to China of the H20, the very chip Nvidia had designed to comply with the earlier limits, under license. The immediate result: a 4.5 Md$ charge on inventory that had become unsellable, another 2.5 Md$ unshipped in the quarter, and roughly 8 Md$ of forgone revenue announced for the next. Nvidia acknowledged that it was, in effect, shut out of the Chinese data-center market.
Then, in August 2025, an unprecedented step: to reopen access, Nvidia and AMD agreed to hand 15 % of their Chinese AI-chip sales to the US Treasury in exchange for the licenses (the administration had first demanded 20 %). The product flow of a private company became a lever of state, taxed at the border and renegotiated chip by chip, as early as December 2025, a 25 % rate was floated to authorize the H200. The dependence runs both ways: the West depends on Nvidia for compute, but Nvidia's access to its second-largest market depends on an unstable bilateral political bargain.