TL;DR
Huawei's Atlas SuperPoD cluster hits 120 FP4 exaflops across 15,488 Ascend 960 chips, but domestic demand so far outstrips supply that Huawei's rotating chairman has ruled out any near-term international expansion.
What happened
- September 17, 2026, at Huawei Connect in Shanghai, rotating chairman Eric Xu confirmed the company cannot satisfy Chinese demand for its AI accelerators, let alone export them broadly.
- Xu's quote was direct: "Since we don't have enough capacity to even satisfy the demand in China, we don't have a plan to expand into the international market in a fully-fledged way."
- September 21-22, Tom's Hardware detailed the Atlas SuperPoD: a single cluster of 15,488 Ascend 960 NPUs linked by optical networking, rated at up to 30 FP8 exaflops and 120 FP4 exaflops by Huawei.
- The Ascend lineup now spans three generations: the 910C (dual-die, up to 780 TFLOPS BF16) as the current cluster backbone, the 910D as the next-gen entrant, and the 960 as the SuperPoD workhorse.
- Optical interconnects, not electrical, hold the 15,488-chip system together, a deliberate choice to avoid the signal-degradation penalty at that node count.
Why it matters
- China's AI compute market is now supply-constrained, not demand-constrained: Huawei is rationing its most advanced silicon domestically rather than chasing overseas revenue, a structural shift from three years of scrambling to prove Ascend could replace Nvidia.
- The 15,488-chip cluster is more than 200 times larger in raw chip count than a single Nvidia NVL72 rack, but Tom's Hardware notes Huawei's performance-per-watt is expected to trail Nvidia by a wide margin, making the exaflop headline less decisive than it appears.
- Scale-first, efficiency-second works when customers are state-directed enterprises with subsidized power and no alternative supplier; it is a much harder sell to commercial cloud operators paying market electricity rates.
- The dual-die 910C packaging compounds yield risk across die, interposer, and final package stages, consuming a disproportionate share of China's scarce advanced packaging capacity and directly explaining the supply crunch.
- Export-control friction already limits where advanced Chinese AI silicon can legally ship, so the capacity constraint and the geopolitical constraint are reinforcing, not competing, bottlenecks.
What to watch next
- Foundry and packaging capacity signals: any announcement from SMIC, CXMT, or Chinese OSAT firms expanding advanced packaging lines would be the earliest indicator that Huawei's supply ceiling is rising.
- Ascend 910D deployment timeline: Huawei has not detailed the 910D's full specifications publicly; a formal launch with yield data would reveal whether the next generation escapes the dual-die packaging bottleneck.
- Nvidia's response at rack scale: if Huawei's optical-linked SuperPoD architecture gains traction as a training platform inside Chinese hyperscalers, watch for Nvidia to accelerate NVLink and optical interconnect roadmap disclosures aimed at the same customers it can still reach.
Originally published on Present of AI, a daily source-linked AI news timeline. Read the full timeline or browse the open dataset.