AINEWS
At Huawei Connect, Wang Tao framed AI training as a systems problem. The proof: a new Ascend 960DT chip, 4,096 of them wired into a single supernode training unit via 5,500 custom optical engines.
Huawei's rotating chairman Wang Tao used his Huawei Connect keynote this week to frame AI infrastructure as a systems problem. "One good chip is far from enough," Wang Tao said, according to trade-press coverage of the event (QbitAI). Huawei is now shipping a stack rather than just a chip: the new Ascend 960DT accelerator, a 4,096-card "supernode" wired with roughly 5,500 custom near-package optical engines, a software platform the company says has crossed developer escape velocity, and a yearly roadmap tied to China's domestic silicon and packaging supply chain.
The headline chip is the Ascend 960DT. Huawei is shipping it three quarters ahead of plan, with company-disclosed specs of 2 PFLOPS of FP8 throughput, 4 PFLOPS of FP4, up to 288 GB of HBM (high-bandwidth memory), and 9.6 TB/s of memory bandwidth. The 960 supernode stitches 4,096 Ascend NPUs into a single coordinated training unit with up to 8 EFLOPS of FP8 throughput and more than 1 PB of total HBM. That is a fourfold jump in card count over the prior 950 generation, which topped out at 1,024 cards, and an order-of-magnitude jump over the 384-card 910C.
The new Hi-ONE engine is a near-package optical (NPO) design with 36 lanes at 200G, 7.2 Tbps per engine, and an integrated light source. Wang Tao framed it as a commercial first. In Huawei's system-level accounting, about 5,500 Hi-ONE engines replace roughly 48,000 conventional 800G optical modules, cut power draw by more than 550 kW, double mean time between failures, and push system availability to 99.8%. These figures are Huawei's own. They are also the load-bearing claim behind Wang Tao's thesis that the bottleneck has moved from silicon to integration.
Wang Tao declined to declare co-packaged optics (CPO), the next step on the optical roadmap, ready for prime time. He put the total cost of ownership advantage of NPO over CPO at "at least 40%," contingent on CPO's reliability and yield maturing. He also pointed at a Model FLOPs Utilization (MFU) number: in Huawei's simulation, at the same 100,000-card cluster scale, a 4,096-card supernode cluster raises MFU 2.75x over a traditional 8-card server cluster. MFU is the share of theoretical compute a cluster actually delivers during training, and the comparison point matters. A 100,000-card cluster running on 8-card servers is wire-bound; running it on 4,096-card supernodes is a topology change, not a chip change.
Huawei's CANN compute platform now operates under routine open-source community management, with more than 5,200 monthly active developers and 61% of them external. Ascend has also landed on PyTorch's official support track. The historical complaint about non-Nvidia AI silicon is that the software stack is the real moat, and Huawei is now claiming, on the record, that the developer ratio has flipped. The cadence Wang Tao laid out (Ascend 950 now, 960 now, 970 in 2028, 980 in 2029) is the fourth leg of the argument, and the one he tied most explicitly to industrial policy. "China's domestic semiconductor manufacturing and packaging supply chain is now fully connected," he said, framing that as the precondition for a one-generation-per-year roadmap.
Every load-bearing number in the announcement comes from Huawei's own simulation or executive statements. No independent benchmark of Hi-ONE optics against CPO or traditional pluggable optics is in the source bundle. No third party has confirmed the 2.75x MFU uplift or the 40%-plus TCO gap. There is no independent shipment, customer adoption, or revenue data for the Ascend 960 series. The through-line is real and worth reading. The numbers behind it are the company's own.
Until a non-Huawei MFU number on a 4,096-card Ascend topology lands, the integration thesis is Huawei's to defend. Three things would change the read. An outside benchmark of the Hi-ONE engine against CPO and pluggable optics. A third-party audit of the 5,200-MAU CANN developer count. Any reported customer deployment of a 960 supernode at scale. The argument Huawei is making is that wiring, not silicon, is now the contest. The data the rest of the industry will need to evaluate that claim has not yet been published.