Huawei Speeds Up Ascend 960DT AI Chip to Early 2027

What happened

Huawei used its Huawei Connect conference in Shanghai on September 17 to announce that its next training-focused AI processor, the Ascend 960DT, will now arrive in the first quarter of 2027, three quarters ahead of the late-2027 target the company had previously set 13812. Rotating chairman David Wang told the audience that the chip's development was running ahead of expectations, with performance roughly doubling versus plan, and that a companion inference-oriented chip, the Ascend 960PR, would follow in the third quarter of 2027, also a quarter earlier than previously scheduled 11121415. Wang added that Huawei intends to keep to an annual cadence after that, with the Ascend 970 due in 2028 and the Ascend 980 in 2029 71114.

The chip announcements arrived alongside a broader systems pitch. Huawei introduced a new interconnect and clustering approach — described in various reports as UnifiedBus 2.0 and a new "Peerium" architecture — meant to let very large numbers of processors, memory and storage act as a single machine 213. Huawei said its Atlas 960 SuperPoD would connect up to 4,096 Ascend accelerators, while a separate research paper posted the same day described a 256,000-node computing system built on the same interconnect philosophy 1113. Huawei also touted the Atlas 950 SuperPoD's commercial rollout past 1,000 units and detailed the Atlas 960's scale-up to as many as 15,488 Ascend cards spread across 220 cabinets 11.

The inference chip already on the market

While the 960-series chips are still more than a year away, Huawei's most commercially relevant inference product today is the Ascend 950PR, launched earlier in 2026 in the Atlas 350 accelerator card 161720. Huawei and multiple outlets covering the launch say the 950PR delivers somewhere between 2.8 and 2.87 times the FP4 inference throughput of Nvidia's H20 — the China-market chip Nvidia built specifically to comply with U.S. export rules — while reportedly costing around a quarter as much 16171920. Specs commonly cited include 1.56 petaflops of FP4 compute, 112GB of HBM, 1.4TB/s of memory bandwidth, and 600W of power draw, with street pricing put at roughly 111,000 yuan, or about $16,000 17. That power draw runs about 200W above the H20's typical envelope, a tradeoff Huawei argues is offset by higher throughput per workload 17.

Huawei has paired that chip with an expansion push, reportedly targeting South Korea with Atlas SuperPods built around 8,192 Ascend 950 accelerators per deployment, pitched explicitly as a cheaper, ecosystem-compatible alternative to Nvidia for buyers wary of tight GPU supply 1619.

Why it matters for inference costs

Training and inference impose different demands on hardware, and Huawei's roadmap increasingly reflects that split: the 960DT is being positioned for training and memory-heavy workloads, while the 960PR is aimed at inference and prefill, where low-precision throughput and cost per token matter more than raw capacity. That distinction is central to why the coverage treats this as an inference-economics story as much as a chip-design one. A chip's headline price or FLOPS figure is not the same as delivered cost per token — power, cooling, networking, software porting and utilization all factor into total cost of ownership, and several reports flag that Huawei's software stack (CANN) still trails Nvidia's CUDA ecosystem in maturity even as compatibility work continues.

Supply is the wrinkle undercutting Huawei's cost pitch. One report details a 60% price increase Huawei imposed on the earlier Ascend 950DT, tied to component shortages, alongside word that DeepSeek plans to deploy at least 160,000 of those chips at a data center in Inner Mongolia 13. That same reporting quotes Huawei executives saying domestic demand already exceeds what the company can produce, constraining its ability to sell aggressively overseas 13.

Where the reporting agrees

Across outlets, the core facts are consistent: Huawei announced at Huawei Connect on September 17 that the Ascend 960DT is moving to Q1 2027 and the Ascend 960PR to Q3 2027, delivered by rotating chairman David Wang, with the 970 and 980 slated for 2028 and 2029 123678911121415. Every account frames this as part of Huawei's effort to reduce China's dependence on Nvidia amid U.S. export restrictions 13691112. There is also broad agreement that Huawei is competing as much on system-level architecture — SuperPoDs, UnifiedBus, massive clustering — as on individual chip specifications 2111415. And separately, coverage of the already-shipping Ascend 950PR converges on the same headline comparison: roughly 2.8 to 2.87 times the H20's inference performance at about a quarter of the cost 16171920.

Where it doesn't

The reporting diverges on scale claims and framing. TechCrunch's coverage flagged an unresolved discrepancy in supernode size: Huawei's newer statements reference a 4,096-chip Ascend 960 supernode, while earlier disclosures described a 15,488-chip Atlas 960 system — a gap an analyst quoted in that reporting called out directly rather than one Huawei has reconciled publicly 111215. Outlets also differ in how they weigh the announcement's significance. TechRepublic, Android Headlines and New Atlas describe the accelerated timeline as a potentially decisive move that could free China from U.S. restrictions 136, while other reporting is more measured, treating it as a signal of ambition rather than proof of near-term volume relief, given that domestic demand already outstrips Huawei's production capacity 13. Pricing details for the 950PR also aren't fully uniform: most figures cluster around $16,000 and 111,000 yuan 17, but one account notes a reported alternative price point near 70,000 yuan for a higher-memory variant, without full reconciliation of why the figures differ 17. Attribution also varies — some technical specifications are presented as confirmed Huawei disclosures 1114, while others are explicitly sourced to TrendForce, Mydrivers or unnamed industry contacts relayed through Reuters and other wire coverage 131819.

The reading the evidence supports

Taken together, the coverage supports treating this as a genuine acceleration of Huawei's roadmap rather than a symbolic gesture, but not as evidence that Huawei has closed the gap with Nvidia's global lineup. The comparison points Huawei promotes — 2.8 to 2.87 times the H20 — are real but narrow: the H20 is Nvidia's deliberately hobbled, export-compliant chip, not its flagship Blackwell-class hardware, a distinction several of the more analytical pieces make explicit while the more celebratory headlines tend to elide. At the same time, the supply and pricing reporting on the 950DT's 60% price hike and Huawei's own admission that domestic demand outstrips capacity are hard to square with any narrative of imminent, low-cost mass availability. The most defensible synthesis is that Huawei is building a increasingly capable, increasingly self-contained Chinese AI-computing stack aimed squarely at inference workloads, where availability and political insulation from U.S. policy now matter as much as raw chip performance — but the timeline moving up by three quarters is a roadmap commitment, not yet a manufacturing or cost-per-token result that outside parties have verified.