Key Takeaways
- Chinese AI developers like DeepSeek and Zhipu AI are co-designing hardware directly with Huawei, embedding staff inside each other's offices to fix software bugs.
- Huawei Ascend chips already run production inference workloads, but historically failed on training runs for 2- to 3-trillion parameter frontier models.
- Rather than waiting to match Nvidia on single-die compute, Huawei uses its telecommunications networking background to bind thousands of lower-yield chips together.
- Industry reports indicate the next flagship DeepSeek model is being pre-trained entirely on Huawei Ascend clusters.
The Co-Location Play
For years, the wall protecting Nvidia was software. Every AI engineer learned on CUDA. Every major open-source library ran fastest on Nvidia drivers. If a rival chip had better raw FLOPS on paper, it still gathered dust on server racks because developers refused to debug missing kernels.
Kyle Chan explained the core trap: “Up until now, the main problem was getting Huawei's AI chips, the Ascend series, to be reliable enough and functional enough to be used by Chinese AI companies who had grown up building on Nvidia and the Nvidia CUDA eco- CUDA ecosystem.” (play from 22:02)
Huawei did not break this inertia with open-source documentation or developer grants. They broke it with direct labor. Chinese frontier labs like Zhipu AI and DeepSeek entered tightly coordinated alliances with Huawei. As Chan observed, “And in particular, you have a really tight partnership between DeepSeek and Huawei, where they literally have personnel situated at each other's companies.” (play from 23:06)
Engineers from DeepSeek sit side by side with Huawei silicon architects. When a training kernel crashes on Ascend hardware, the model developer does not file a bug ticket into a void; the engineer who wrote the chip firmware sits two desks away. That tight loop turns software fixes from six-month release cycles into daily patches.
Networking Over Raw Die Performance
The biggest hurdle has always been pre-training. Inference is comparatively simple: you load weights into memory and compute forward passes. Training 2- to 3-trillion parameter models across thousands of nodes requires constant, high-speed synchronization.
“And then more recently, we have this push for the training side, right?” (play from 23:44) Chan pointed out. “So, up until now, the Huawei Ascend chips were good enough for inference, but were not quite good enough for training, especially for these really massive 2-, 3-trillion parameter models.” (play from 23:48)
Huawei cannot access the newest lithography nodes abroad due to trade restrictions. Their individual processors lag Nvidia's top dies in energy efficiency and raw floating-point calculations. So Huawei bypassed the single-die race entirely, falling back on their decades as a telecom equipment builder.
By treating high-bandwidth interconnects and custom networking protocols as primary features, Huawei strings together large clusters that absorb individual chip defects. The strategy is showing immediate returns. As Chan noted, “And there's rumors that the next DeepSeek model, for example, will be fully pre-trained on Huawei chips.” (play from 24:40)
What to Do With This
Audit your infrastructure dependencies this week. List your three biggest vendor dependencies, and identify where you are paying a 40 percent markup simply to avoid rewriting internal tools for an alternative provider. Run a 48-hour internal hackathon to port one core data pipeline to open tooling, proving whether switching costs are structural or merely habitual.