Huawei just hit the accelerator on its anti-Nvidia road map. At Huawei Connect, rotating and acting chairman David Wang said the Ascend 960DT AI chip is now expected in Q1 2027—months earlier than the previous Q3 2027 target a company spokesperson confirmed to TechCrunch.
“Ascend 960 chips are launching ahead of schedule, doubling performance and advancing year by year,” the spokesperson said. In chip-land, “ahead of schedule” usually means either the engineers got lucky or the geopolitics got loud. With Trump and Xi lined up to meet days later in Washington, the timing is not subtle.
Peerium, UnifiedBus, and the giant-computer fantasy
Huawei’s bigger story is not one die—it is stitching hundreds of thousands (eventually millions) of accelerators into what it calls a Peerium Computing Architecture. The glue is UnifiedBus, Huawei’s way to link processors with memory, storage, and networking without borrowing Nvidia’s proprietary playbook.
Rotating chairman Eric Xu framed Peerium as the path to larger training and inference systems. First systems on that architecture—the Atlas 950 SuperPoD and SuperCluster—already claim the ability to wire up to 256,000 accelerator cards in a SuperCluster configuration. That is “one giant computer” cosplay with a very real facilities bill.
Atlas 960 SuperPoD and near-packaged optics
The Connect stage also pushed the Atlas 960 SuperPoD. Coverage from SiliconReport and SCMP describes denser clusters, UnifiedBus paired with a Hi-ONE optical system, and near-packaged optics—optics sitting much closer to the silicon so the interconnect does not become the bottleneck the moment you leave the package.
Huawei’s pitch: training and inference gains on the order of roughly 2.3× / 2.5× versus Atlas 950 for gigantic (~10-trillion-parameter) model regimes, plus a SuperPoD that can scale toward 15,488 Ascend cards across hundreds of cabinets. Analyst Rui Ma flagged a tension worth keeping: earlier roadmaps talked about that huge SuperPoD, while some Connect messaging referenced systems with 4,096 chips. Chip early, cluster smaller—or different SKUs for different stages. Either way, the marketing and the rack drawings need a careful read.
Huawei also nudged the Ascend 960 PR inference-oriented variant forward to Q3 2027. Training silicon gets the headlines; inference silicon pays the electricity bill.
Sanctions did not end the story
U.S. export controls still try to starve China of the most advanced process tech and Nvidia’s top GPUs. Huawei’s answer remains vertical integration: design the NPU, own the interconnect, invent the optical tricks, and ship clusters into a domestic market that cannot casually order Blackwells. Ma’s blunt take: trying to freeze China’s semiconductor climb is increasingly futile because the self-sufficiency stakes are existential.
On the policy theater side, Xu argued China needs to push AI harder to catch the U.S. before it can fully wrestle with the risks of stronger systems—an inversion of Western “slow down for safety” rhetoric that Trump has also rejected when it conflicts with beating Beijing.
What “doubling performance” actually buys
Wang’s line about performance doubling year over year is marketing with a purpose. Domestic customers comparing Ascend generations care less about FLOPs theater and more about whether a SuperPoD can finish a training run before the product window closes. Bringing 960DT into Q1 2027 compresses that window for Chinese labs that cannot casually rent Nvidia clouds abroad.
Near-packaged optics and UnifiedBus are the unsexy heroes. If the copper/optical hop off the package stays fat, denser cabinets stop being science projects. If optics yields or firmware wobble, you get beautiful keynotes and lonely racks. Huawei’s Hi-ONE pitch is basically: treat the optical path like part of the chip product, not an afterthought cable.
Geeknewz take
Do not read Q1 2027 as “Huawei beat Nvidia.” Read it as “Huawei refuses to wait.” Moving Ascend 960DT up three quarters, talking Peerium scale-out, and putting near-packaged optics next to UnifiedBus is how you build a parallel AI stack when the global one is gated. Benchmarks, yield, and actual customer deployments will decide whether Connect bravado becomes cluster reality.
Until then, the scoreboard is simple: Nvidia still owns most of the world’s AI FLOPs mindshare. Huawei owns a domestic mandate, a ruthless interconnect roadmap, and a calendar that just got more aggressive.
Source: TechCrunch — Huawei plans Q1 2027 launch of new AI chip as it takes on Nvidia (Kate Park, Sep 17, 2026); additional SuperPoD/optics detail via SiliconReport / SCMP reporting from Huawei Connect.