Huawei’s New AI Chip — A Bold Challenge to Nvidia - #32
Series “Laboratory”
Introduction
Welcome back to Laboratory, our deep-dive series charting the forces that will shape the next chapter of AI. Today we zero in on Chip Wars: Ascend vs. CUDA, the high-stakes drama playing out inside the data-centre racks of Beijing, where Huawei’s new Ascend 910D is squaring off against Nvidia’s crown-jewel GPUs. What began as a sanctions story now reads like a manufacturing thriller: think chiplets lashed together like Lego, power budgets that would shame a steel mill, and an entire ecosystem racing to prove that sovereignty can be carved out of silicon.
Here’s what we’ll be unpacking:
Silicon Under Siege – why U.S. export controls turned China’s AI appetite into a hardware gold rush, and how Huawei slipped past the pickets.
Lego-Chiplets – the multi-die packaging trick that lets Ascend jump a lithography generation without EUV tools.
Cluster Calculus – inside CloudMatrix 384, the monster rack that strings 384 Ascends to out-muscle Nvidia’s NVL72 on sheer scale.
Energy Reckoning – the real cost of compute sovereignty when every training run lights up the grid like a small town.
Upside Plays – fabs, optics, immersion-cooling vendors, and other second-order winners most portfolios are still ignoring.
Two-Stack Future – how parallel ecosystems built on CUDA and MindSpore could polarise global AI development for a decade.
Strap in as we follow the electrons—from SMIC’s clean rooms to hyperscale clusters on the outskirts of Beijing—to see how mastering this next generation of silicon will redraw the balance of power in artificial intelligence.