Huawei Says It Has Mass-Produced 381 Chips Under Its 'Tau Scaling Law' — and Kirin 9050 Pro Is the Proof

Richard Yu, Huawei's executive director and chairman of its consumer business group, said this week that Huawei's semiconductor unit has designed and mass-produced 381 chips under the company's Tau Scaling Law, spanning smartphones, AI computing, general-purpose compute, networking and smart vehicles. The reveal came in a video posted on October 4, and the Mate 90 series — which starts at 5,999 yuan — is the first mass-market product built on the new generation.

The Tau (τ) Scaling Law is Huawei's answer to a problem every chipmaker now faces: transistors are close to their physical limits, and the decades-old trick of shrinking geometry to gain performance is getting harder and more expensive. Huawei's formulation, first laid out by He Tingbo, the head of its semiconductor business, at the IEEE International Symposium on Circuits and Systems in May, is to trade geometric scaling for what the company calls time scaling — cutting how far data travels and how long compute waits, rather than how small the transistors are.

The Kirin 9050 Pro, inside the Mate 90 Pro Max, is the first flagship chip to use the accompanying LogicFolding technology in full: logic layers are stacked vertically and rewired through high-density inter-layer interconnects that shorten critical paths. Huawei's figures put transistor density at 238 million per square millimeter, 28% above the previous Kirin 9030 Pro, with 5 million signal bonds, 125 TB/s of cross-die bandwidth, 55% fewer clock buffers and 30% lower critical-path latency. The Lingxi CPU runs 9 cores and 16 threads, and Huawei credits LogicFolding with a 23% multi-core gain.

The generation-on-generation numbers, all vendor-reported and not independently verified: Kirin 9030 gains 13% CPU, 54% GPU and 13% NPU over the Kirin 9020; Kirin 9035 adds 11% CPU, 10% GPU and 51% NPU over the 9030. At the device level Huawei claims the Mate 90 is 27% faster than the Mate 80, the Pro 20% faster than its predecessor, and the Pro Max 31% faster. On the AI side, the company says the Mate 90 series can run a 30-billion-parameter mixture-of-experts model entirely on device.

Capacity, not design, is the bottleneck. Yu acknowledged that China's advanced semiconductor production still cannot meet fast-growing demand, and that Huawei's Ascend AI chips draw on the same constrained capacity as its phone silicon. He also noted that sharp memory price increases over the past two years have limited how many smartphones Huawei can ship. Internally, he said, a requirement plan now balances phone and data center demand against what the domestic supply chain can actually produce.

The chain behind the 381 chips is largely domestic: SMIC handles foundry work, Naura supplies equipment, JCET-backed OSATs handle advanced packaging, and CXMT's high-bandwidth memory has begun sampling to domestic AI firms. The market backdrop supports the story — IDC counted Huawei's China smartphone shipments up 19.4% year over year in the second quarter, a 22.6% share that returned the company to first place — and Huawei says HarmonyOS 6 and 7 now run on more than 90 million devices, a figure it expects to reach 100 million by year-end.

What it means: Huawei is betting that architecture, packaging and interconnect can substitute for leading-edge lithography it cannot buy, and the 2031 target — transistor density equivalent to a 1.4nm-class process on Tau-based chips — makes that bet explicit. The claims are compelling but self-reported; the open question is whether independent teardowns and benchmarks confirm that LogicFolding delivers near-flagship efficiency on mature nodes, or whether the density numbers arrive with thermal and yield costs the press materials omit.