Huawei unveils Ascend 960 SuperPod with NPO at Connect 2026

At Huawei Connect 2026 in Shanghai, Huawei’s China newsroom said rotating chairman David Wang unveiled the Ascend 960 SuperPod — framed as the industry’s first SuperPod using near-packaged optics (NPO / Hi-ONE) — and same-day wires report the Ascend 960DT training chip is now targeted for Q1 2027 and the 960PR inference chip for Q3 2027.

HARDWARE desk — China’s largest telecom and AI-infrastructure vendor put a dated Connect keynote behind an NPO Ascend 960 SuperPod and a pulled-forward 2027 Ascend 960 chip calendar, the clearest same-day public roadmap signal in the China–NVIDIA accelerator race.

Company framing: the world’s first SuperPod to use NPO — near-packaged optics, meaning the optical engine sits next to the chip package instead of in a pluggable module on the front of the rack — through Huawei’s Hi-ONE optical engine. Huawei says the box uses an orthogonal layout and full liquid cooling, and that UnifiedBus (灵衢) lets the SuperPod treat memory as one address space. The company says the SuperPod can scale to 4,096 accelerator cards and claims up to 8 EFLOPS of FP8 math and 16 EFLOPS of FP4 math. EFLOPS means exaFLOPS — a billion billion floating-point operations per second, a vendor peak-throughput number, not a measured job. FP8 and FP4 are thinner number formats that use 8 or 4 bits. File that first-NPO / Hi-ONE / 4,096-card / EFLOPS picture as Huawei’s. “World’s first” is company wording — not a desk ranking. This desk did not bench a rack.

Company claim: about 5,500 Hi-ONE units replace about 48,000 800G pluggable optical modules. Huawei says that swap drops power by more than 550 kilowatts, doubles mean time between failures — how long the system runs before a fault — and takes availability to about 99.8%. File those swap / power / uptime lines as Huawei’s. This desk did not meter a hall or time a failure.

Same keynote, still Huawei, and secondary to the Ascend 960 lead: the company also described Kunpeng SuperPod upgrades on UnifiedBus all-optical networking, said to scale to 4,096 nodes, and an OceanStor M900 KV-cache cluster on UnifiedBus. Kunpeng is Huawei’s general-purpose CPU line. KV-cache is the working memory a large model keeps while it answers. File those as Huawei’s side announcements, not as this card’s lead. Extra cluster-scale talk — up to 512,000 cards, or 1 million with multi-track topology — stays in Sources.

Same-day Reuters: rotating chairman David Wang said Huawei will have the Ascend 960DT training chip ready in the first quarter of 2027 — Reuters says three quarters earlier than previously planned — and the 960PR inference chip ready in the third quarter of 2027, one quarter ahead. Training means teaching a new model. Inference means running a trained model to answer a query. Reuters also said UnifiedBus is how Huawei ties processors, memory, storage, and networking inside those next systems, and that company materials say Huawei has already deployed more than 1,000 systems using earlier Ascend 910C processors, with Ascend 950 systems in commercial use. File the 960DT / 960PR calendar, the UnifiedBus line, and those deployed-system figures as Reuters’. This desk did not see a ship manifest. Those 2027 dates are targets, not a box you can buy today.

Same-day TechCrunch: a Huawei spokesperson said the 960DT had been planned for the third quarter of 2027 and is now expected in the first quarter of 2027, and that “Ascend 960 chips are launching ahead of schedule, doubling performance and advancing year by year,” with Wang announcing the updated timeline at Connect. File that pull-forward and the year-by-year doubling line as TechCrunch / the spokesperson via TechCrunch. Reuters’ “three quarters earlier” and TechCrunch’s prior-Q3 wording are the two wires’ own calendars — this desk is not inventing a single older plan. This desk did not interview the spokesperson.

Plain English for the rest of the card: SuperPod = Huawei’s name for a rack-scale AI training and inference cluster (超节点). NPO = near-packaged optics, lights packaged next to the chip instead of pluggable modules. Hi-ONE = Huawei’s NPO optical engine. Ascend = Huawei’s AI accelerator line, the chips that compete with NVIDIA GPUs inside China. UnifiedBus (灵衢) = Huawei’s interconnect that lets memory, storage, and networking share one fabric. EFLOPS = exaFLOPS, a billion billion floating-point operations per second. FP8 / FP4 = 8-bit and 4-bit number formats used for peak-throughput claims. 960DT = the training-chip target. 960PR = the inference-chip target. Those 2027 chip dates are targets, not a retail SKU.

PRIMARY here: Huawei’s 17 Sep 2026 China newsroom release — Tier A PRIMARY company source, the original record. Reuters and TechCrunch are same-day independent Tier B/C confirm of Wang’s 960DT Q1 2027 / 960PR Q3 2027 calendar and of the pull-forward — not a substitute primary, and not a source for invented FLOPS comparisons versus NVIDIA or U.S. export-license outcomes. The Connect keynote, the Ascend 960 SuperPod unveil, the first-NPO / Hi-ONE framing, the orthogonal / full-liquid-cooling / UnifiedBus / 4,096-card picture, the 8 EFLOPS FP8 / 16 EFLOPS FP4 claims, the ~5,500 Hi-ONE / ~48,000 800G-module swap, the >550 kW / doubled MTBF / ~99.8% availability lines, and the Kunpeng SuperPod / OceanStor M900 side announcements are Huawei-attributed. The 960DT Q1 2027 / 960PR Q3 2027 targets, the UnifiedBus next-system line, and the >1,000 Ascend 910C / 950-in-commercial-use figures are Reuters-attributed. The prior-Q3-2027 pull-forward and the year-by-year doubling line are TechCrunch / spokesperson-attributed. Peak-throughput, power, and uptime numbers stay company-attributed — not independently remeasured here. NOT claimed: an independent FLOPS audit versus NVIDIA, a U.S. BIS license outcome, that 960DT or 960PR is available to buy today, a customer-win count this desk verified, that this desk sat in the keynote hall, a stock tip, or investment advice. Distinct from the already-filed nvidia-vera-rubin-nvl72-mlperf, polaris-electro-optics-50m-series-b, dmatrix-nvlink-fusion-raptor, and euclyd-200m-series-a.