Huawei Ascend 960: Q1 2027 Launch, Specs and Nvidia Rivalry

SHANGHAI -- Huawei doesn't do subtle. At its annual Huawei Connect conference on Thursday, the Chinese telecom giant unveiled not one but two next-generation AI chips -- the Ascend 960DT and Ascend 960PR -- and then dropped the detail that actually matters: the first one is now coming in the first quarter of 2027, a full three quarters ahead of schedule.

In a semiconductor war that Washington has spent years trying to win through export controls, Huawei just told the world it's running ahead of its own roadmap. "Our Ascend 960 chip is progressing ahead of expectations, with performance doubling," said David Wang Tao, Huawei's rotating and acting chairman, according to the South China Morning Post. "We have brought the target launch forward by three quarters, and it will be ready in the first quarter of 2027."

For American readers, the context is impossible to miss. The announcements landed days ahead of expected trade talks between President Xi Jinping and President Donald Trump later this month, where AI and semiconductor exports are expected to top the technology agenda. And they arrived just as the United States is holding a separate, first-of-its-kind bilateral dialogue with Beijing devoted exclusively to AI risks -- a signal that chips and AI safety are now inseparable from diplomacy. (See our earlier report on that dialogue: first bilateral AI-safety dialogue.)

Huawei's Ascend 960 AI chips, detailed at Huawei Connect 2026 in Shanghai, are now slated to arrive in the first quarter of 2027 -- three quarters ahead of Huawei's original plan. (Illustration: TechPulse USA (original))

What Huawei actually announced

Strip away the stagecraft and the announcement had a few hard, verifiable facts. Reuters reported Thursday that Huawei will launch two new AI chips in 2027: the 960DT in the first quarter and the Ascend 960PR in the third quarter. Wang said Huawei's UnifiedBus interconnect technology would be key to the next generation of its large AI computing systems, and added that the company has now shipped more than 1,000 such systems to more than 370 customers -- a deployment figure that suggests the Ascend line has moved well past the experimental stage inside China.

Ascend 970 and 980: Huawei's roadmap to 2029

Wang also laid out a longer roadmap: the Ascend 960 family will be followed by the Ascend 970 in 2028 and the Ascend 980 in 2029, with Huawei targeting a doubling of computing performance with each annual generation. "Guided by innovation based on the Tau Law," Wang said, "we aim to continue doubling computing performance, while also significantly increasing bandwidth, memory capacity and interconnect bandwidth."

Ascend 960DT vs 960PR: training chip vs inference chip

Huawei's two-letter suffixes describe what the silicon is built to do. Following Huawei's established naming from the current Ascend 950 generation, the PR variants are tuned for the prefill stage of inference and recommendation workloads, while the DT variants are built for the decode stage of inference and for model training, with more memory and bandwidth for heavier work (per Reuters' September 2025 reporting on the 950 line; there's no indication Huawei is changing the scheme for the 960s).

The sequence is the tell. Huawei is launching the training-focused 960DT first, in Q1 2027, with the inference-focused 960PR following in Q3. Ship the training chip first, and you're saying Chinese labs need more training compute -- and Huawei intends to be the one to provide it. It's the clearest signal yet that Huawei's priority is closing the gap with Nvidia where it hurts most: frontier model training.

Nine months early: the schedule shift is the story

When Huawei laid out its three-year Ascend roadmap last September, the 960 was a Q4 2027 product. The new date -- Q1 2027 -- is three quarters earlier. In semiconductors, pulling a leading-edge accelerator forward by nine months is not a small feat. It either means Huawei's engineering execution is genuinely ahead of plan, or that Beijing's pressure to arm domestic AI labs with homegrown training chips has made "on time" unacceptable.

But roadmap acceleration can also be theater. Announced shipping dates in the chip industry are promises, not deliveries. Bloomberg, reporting ahead of the keynote, flagged the question analysts are actually asking: whether the new component will be powerful enough for training frontier AI models. Everything else -- the dates, the doubling claims, the SuperPoDs -- is preamble to that single unanswered question.

China's Huawei and America's Nvidia are locked in an AI chip race that U.S. export controls were supposed to end. This week's announcement suggests otherwise. (Illustration: TechPulse USA (original))

The real strategy isn't the chip -- it's 15,488 of them

Huawei has been remarkably honest about its own weakness, which makes its strategy unusually legible. "Gaps still exist between Huawei and Nvidia in terms of individual chips and ecosystems, and we are working to close these gaps," rotating chairman Eric Xu told Chinese media last year. "But we have the capability to build SuperPoDs and SuperClusters -- which are the source of our confidence."

Thursday's centerpiece was the Atlas 960 SuperPoD: up to 15,488 Ascend cards across 220 cabinets and 2,200 square metres, which Huawei says will deliver three times the training performance and four times the inference performance of the current Atlas 950 SuperPoD. Two technical additions stood out: near-packaged optics (NPO), making their SuperPoD debut to push higher speeds at lower power, and an upgraded UnifiedBus interconnect with Huawei's Hi-ONE optical system, which Huawei says can link as many as 4,000 processors so separate chips function as one machine. Wang claimed Huawei is "set to begin mass production of the world's first NPO product with an integrated light source."

A same-day Huawei paper on arXiv added scale to the claim: the company says it is already deploying a 256,000-node-class single AI computing system -- described by Huawei researcher Liao Heng not as a network of machines, but as "one machine," with UnifiedBus spanning it end to end -- the latest entry in the global race to build hyperscale AI clusters. If even close to accurate, it's the physical embodiment of Huawei's core bet: that interconnect and packaging innovation can substitute for the advanced foundry access U.S. sanctions deny it.

The American angle: sanctions, talks, and Nvidia's China problem

None of this happens in a vacuum. The U.S. Commerce Department's Bureau of Industry and Security has spent the Trump era tightening the screws -- including 2025 guidance warning that using Huawei's Ascend chips "anywhere in the world" risks violating U.S. export controls -- while Beijing has moved the other way, with Chinese regulators reportedly directing large domestic tech firms to cancel orders for Nvidia's China-specific RTX Pro 6000D chips. The result is a Chinese market increasingly reserved for Chinese silicon, and Huawei is positioning the Ascend line as its only credible occupant. And memory is the industry's other choke point in the AI hardware stack.

For Nvidia, the timing is awkward. Last year, Huawei's Eric Xu claimed the Atlas 950 cluster would offer 6.7 times more compute power than Nvidia's forthcoming Vera Rubin NVL144 rack-scale system -- a comparison Nvidia has never publicly engaged with. Analysts, per Bloomberg, are withholding judgment on whether Huawei's silicon can actually train frontier-class models, but even the credible threat of a domestic alternative accelerates China's decoupling from Nvidia's China business -- a revenue stream Nvidia investors can no longer take for granted.

And then there's the diplomacy. SCMP noted the announcements came ahead of expected Xi-Trump trade talks later this month, where AI and semiconductor exports are expected to top the technology agenda. In trade negotiations, nothing beats a good prop -- and "we no longer need your chips" is a powerful one.

Analysis: what the Ascend 960 really tells us about the chip race

The following is analysis, not reporting -- our read on what Thursday's announcements mean, grounded in the verified facts above.

First, the export-control regime has bent, not broken, Huawei's trajectory. U.S. policy assumed that cutting China off from TSMC's advanced nodes would freeze Chinese AI compute in place. Instead, Huawei moved the bottleneck: denied leading-edge transistor density, it poured engineering into interconnect (UnifiedBus, NPO, Hi-ONE optics), in-house high-bandwidth memory (its HiBL and HiZQ lines), and brute-force clustering at a scale Western vendors haven't attempted commercially. Yields and bandwidth efficiency on domestic fabs remain the great unknown -- but the roadmap keeps moving right, not left.

Second, Huawei is now competing on cadence, not just silicon. An annual doubling cycle with named products stretching to the Ascend 980 in 2029 is a deliberate mirror of Nvidia's own rhythm -- a message aimed at Chinese AI labs and investors as much as at Washington: the domestic supply chain now has a predictable upgrade path, which is what actually drives purchasing decisions. Nvidia's moat was never just transistor physics -- it was certainty. Huawei is trying to manufacture certainty.

Third, the individual-chip gap still matters, and Huawei knows it. No one outside Huawei has independently benchmarked an Ascend 960 -- the chips don't exist in customer hands yet. Individual Ascend chips have historically trailed Nvidia's H100 and B200-class parts in raw power, and analysts will be watching whether the 960 generation genuinely closes that gap or papers over it with cluster scale. The honest assessment: Huawei's systems are increasingly formidable, while its chips remain the open question. U.S. policy should be calibrated to that distinction -- punishing the chips you can verify, not the press releases you can't.

Fourth, and most important for American readers: the chip race is now a standards race. UnifiedBus, proprietary memory formats like HiF4, Huawei's in-house HBM -- every one of these is a fork in the road away from the Nvidia/CUDA ecosystem. If Huawei's domestic customers standardize on its interconnects and memory, China's AI stack decouples from American technology not just commercially but architecturally, making future reconciliation -- and future American leverage -- much harder. That's the long-term cost of a sanctions regime that works too slowly to stop the alternative from being built.

Can the Ascend 960 train frontier AI models?

This is the question analysts keep coming back to. Bloomberg, reporting ahead of the Huawei Connect keynote, flagged it as the central unknown: whether the new component will be powerful enough for training frontier AI models. Everything else -- the Q1 2027 date, the performance-doubling claims, the 15,488-card SuperPoDs -- is preamble to that single unanswered question.

The honest answer is that no one outside Huawei knows. The chips do not exist in customer hands yet, and Huawei has published no detailed per-chip specifications for the 960DT or 960PR. Whether the Ascend 960 can train a frontier-class AI model has no public evidence either way, and Bloomberg reports analysts are watching precisely this.

What is clear from Huawei's own framing is where the company thinks the answer matters less: the system around the chip. The SuperPoD strategy -- clustering thousands of Ascend cards with UnifiedBus interconnect and in-house high-bandwidth memory -- is designed to compensate for any single-chip deficit with scale. Whether that compensation is enough for frontier training remains the industry's most consequential open question.

What we still don't know

For all the Thursday fireworks, several critical facts remain undisclosed or unverifiable. Huawei has not published detailed per-chip specifications for the 960DT or 960PR -- the compute, memory, and bandwidth figures circulating in the trade press (including 2 PFLOPS of FP8 compute) trace back to Huawei's 2025 roadmap disclosures, not to new silicon. Manufacturing details -- process node, foundry partner, yields -- were not disclosed; analysts generally assume domestic production via SMIC, but Huawei won't say. No pricing has been announced. And the big one -- whether the Ascend 960 can train a frontier-class AI model -- has no public evidence either way; Bloomberg reports analysts are watching precisely this.

The honest version of this story is that Huawei has announced a plan, not a product. Plans can slip, yields can disappoint, and "performance doubling" is a claim made on a Shanghai stage, not a benchmark run by an independent lab. But dismiss it at your peril: a year ago, skeptics said the same about the 950 roadmap, and Huawei just shipped it early.

Sources: Reuters (September 17, 2026), Bloomberg via The Japan Times (September 17, 2026), South China Morning Post (September 17, 2026), Reuters technology roadmap explainer (September 2025), RCR Wireless, Tom's Hardware, Wccftech, SDxCentral. All Huawei performance claims are attributed as company statements unless otherwise noted. This is a developing story; TechPulse USA will update this post as independent benchmarks and shipping details emerge.