U.S. export restrictions on Nvidia's H100 and H200 accelerators were designed to hold back Chinese AI by denying access to the compute needed to train frontier models. The outcome has been more complicated. Chinese labs, unable to scale by simply buying more chips, shifted engineering resources toward doing more with less—and the efficiency gains that followed are now a structural cost advantage.
DeepSeek's R1 reasoning model is the clearest data point. The lab trained R1 for $294,000. Comparable training runs at American frontier labs cost hundreds of millions of dollars. That gap did not emerge because Chinese engineers had better chips. It emerged because they had worse ones and had to find a different path.
The hardware constraints pushed Chinese development in two directions simultaneously. Labs worked to improve chip utilization during training runs—extracting more compute from each pass—and redesigned model architectures to reduce the total memory and compute required to run inference. Post-training techniques, including reinforcement learning from human feedback and chain-of-thought distillation, pulled additional reasoning capability from base models without requiring larger training runs. The result is a software stack optimized for efficiency rather than raw scale.
On the hardware side, Huawei's HiSilicon division has made measurable progress with its Ascend AI accelerator line. The current production chip, the Ascend 910C, delivers 256 teraflops of FP16 compute and 1.2 terabytes per second of memory bandwidth. It carries 16 built-in Arm-compatible CPU cores, allowing it to operate without a separate host processor—a design choice that cuts cluster deployment costs. The Ascend 910C's performance is roughly 77 percent of the Nvidia H100. The H100 costs three to five times more per unit.
For enterprise workloads where training a new frontier model from scratch is not the goal—inference, fine-tuning, deployment at scale—the cost math on the Ascend 910C is competitive. A buyer who can accept 77 percent of H100 performance at 20 to 33 percent of the price is making a rational decision, not a compromise. That customer segment is large: most production AI deployments are inference-heavy, not training-heavy.
China's third domestic layer is high-bandwidth memory. CXMT, the Chinese memory manufacturer, has been scaling high-bandwidth memory production domestically. High-bandwidth memory is the component that feeds data to AI accelerators fast enough to keep them running at full utilization—it is as critical to cluster performance as the accelerator itself, and until recently it was almost entirely sourced from Samsung and SK Hynix in South Korea. Domestic production reduces that import dependency and, combined with the Ascend 910C, moves China closer to a full domestic AI cluster stack.
Huawei is pairing the hardware push with a software ecosystem play. The company is working with DeepSeek directly—DeepSeek's open-weight models have been engineered to run on both Nvidia and Huawei processors. That dual compatibility gives enterprises a path to adopt DeepSeek models on Ascend hardware without rewriting their inference stack from scratch. It mirrors the approach China used to build domestic developer ecosystems in other hardware markets: open what incumbents keep closed, subsidize adoption, and create a generation of developers building on the alternative.
Frontier training—building the largest base models from scratch—still depends on advanced hardware China cannot fully replace domestically. The performance gap at the absolute cutting edge of training compute remains real. Where Chinese labs are genuinely close to parity is in running competitive models on domestic chips. The software efficiency stack they built under constraint is already mature enough to narrow the capability gap on older-generation silicon.
Many enterprise workloads do not require the best-performing model available—they require a model that performs well enough for the task at the lowest cost per token. Chinese models that combine near-competitive performance with substantially lower inference costs win that segment on price, not on cutting-edge capability. That is a large and growing segment of enterprise AI spending.
