Training AI without a single NVIDIA GPU: how China closed its own loop
19 set 2026
⏱️ Reading time: ~5 minutes
Introduction
For years, the answer to "what does it take to train a frontier model?" was short and uncomfortable: NVIDIA GPUs, lots of them. It was less a technical choice than a fact of life — the entire ecosystem, from framework down to kernel, was built around CUDA.
In 2026, that stopped being the only answer. A handful of Chinese labs completed full training cycles — from data preprocessing to final scale-up — without touching a single Western chip. This was not a lab experiment: these are production models, with open weights and APIs charging per use. And the trigger wasn't technical. It was regulatory.
The concrete cases
GLM-Image, by Zhipu AI (January 2026). The opening milestone. The first state-of-the-art multimodal model to complete its entire training run on Chinese chips, on Ascend Atlas 800T A2 hardware with the MindSpore framework. Hybrid architecture: a 9-billion-parameter autoregressive model handling instruction understanding and overall image composition, plus a 7-billion-parameter diffusion decoder responsible for fine detail and text rendering. It led the CVTG-2K benchmark (word accuracy of 0.9116) and LongText-Bench (0.952 in English, 0.979 in Chinese). Weights published on GitHub, Hugging Face, and ModelScope; API priced at 0.1 yuan — roughly US$ 0.014 — per generated image.
GLM-5, by Zhipu (February 2026). The leap in scale: a language model of roughly 744 billion parameters, trained entirely on Ascend, with reports of around 100,000 chips in play. Its benchmarks landed within a few points of GPT-5.2 and Claude Opus 4.5.
openPangu 2.0, by Huawei itself (July 31, 2026). A mixture of experts with 505 billion parameters, trained exclusively on Ascend NPUs and released as open source.
DeepSeek V4. Moved its training to the Ascend 950, dropping NVIDIA in the process.
The trigger was a list, not a roadmap
It's worth recording how this started. In January 2025, Zhipu was added to the US Commerce Department's Entity List, which cut off its access to H100 and A100 — the cards that had become synonymous with advanced training. The company had no plan B; it had to build one, in partnership with Huawei.
That's the central irony here: the restriction meant to stall development ended up funding the alternative, the hard way.
What the engineering actually required
None of this was plug and play. Swapping CUDA for Ascend meant rewriting the layer most teams never touch.
The stack has three floors: the Ascend NPUs in hardware; CANN, the layer that translates PyTorch and TensorFlow instructions into operations the chip understands; and MindSpore as the training framework. CANN and the Mind-series tooling were open-sourced in August 2025 — a deliberate move to attract developers.
On top of that, Zhipu needed bespoke optimizations: dynamic-graph multi-level pipelined deployment, to run training stages concurrently and cut bottlenecks; custom fusion operators compatible with the Ascend architecture; and multi-stream parallelism, overlapping communication and computation during distributed training. All of this exists because the frameworks assume, by default, that there's an NVIDIA GPU on the other end.
The hardware, and where it's headed
At Huawei Connect 2026, held in September, the company pulled the Ascend 960 forward by three quarters: the 960DT arrives in the first quarter of 2027 and the 960PR in the third. The annual cadence is now the rule — Ascend 970 in 2028, 980 in 2029. Alongside it came a portfolio of 11 UnifiedBus chips for SuperPoD and SuperCluster systems, covering compute, interconnect, storage, and management, plus a scale-up plan in the range of a million NPUs.
The caveats — and they matter
Three points block any triumphalist reading.
The individual chip still loses. The 960DT is estimated at around 288 GB of memory and 4 petaFLOPS in FP4 — roughly half the FP8 and a third of the FP4 of NVIDIA's B300. The Chinese bet isn't to win on silicon, but to compensate through scale and interconnect. Anyone who runs a data center knows that trade isn't free: more nodes means more network, more power, more failure domain.
The real cost wasn't disclosed. Zhipu didn't say how many processors it used, how long training took, or how any of it compares to an equivalent NVIDIA cluster. The engineering-time cost — the most expensive of all — remains unknown.
"Without NVIDIA" describes the training, not the whole chain. There is reporting questioning the upstream dependency — lithography, HBM memory — behind that independence. Autonomy at the training layer is not the same as end-to-end autonomy.
Summary
What 2026 proved isn't that China caught up with NVIDIA. It's something more modest and, at the same time, more consequential: there is a second viable path to training frontier models, and it is already in production.
For anyone running infrastructure, the practical implication is that CUDA's de facto monopoly now has a working competitor — immature, more expensive in engineering, with weaker per-unit hardware, but working. And historically, when a technical monopoly gains a real alternative, the first thing that changes isn't the technology. It's the price.