At its annual Connect summit in Shanghai on September 17, Huawei gave its next-generation AI chip an accelerated timeline. The Ascend 960DT, originally planned for Q3 2027, will now ship in Q1 2027 — a three-quarter pull-forward. A companion chip, the Ascend 960PR, moved up one quarter to Q3 2027.

David Wang, Huawei’s rotating chairman, announced the revised roadmap and described the Ascend 960 family as delivering roughly double the performance of its predecessor. The specs back that up: the 960DT is rated at 2 FP8 PFLOPS and 4 FP4 PFLOPS, with 288 GB of HiZQ memory running at 9.6 TB/s bandwidth and a 2.2 TB/s interconnect. That puts it in the same performance tier as high-end Nvidia hardware — on paper, at least.

For context on how fast this market is moving: Huawei’s existing Ascend 910C already powers a significant share of AI training and inference inside China. The 960 series represents the next generational jump, and the accelerated schedule signals that Huawei considers the competitive window with Nvidia to be right now, not later.

Why This Matters Beyond China

The obvious read is that this is a China story — Huawei building chips China can actually buy, since Nvidia’s H100 and Blackwell chips remain export-restricted under US trade policy. That is true. But the ripple effects reach further.

First, the pricing pressure. Nvidia’s ability to charge premium prices for AI accelerators has depended partly on there being no credible alternative at scale. Huawei closing the performance gap — whether or not it fully closes it — creates a negotiating lever for any enterprise operating in markets where Huawei hardware is available.

Second, the supply question. The global demand for AI compute far outstrips what Nvidia can currently build. Any credible second supplier increases the ceiling on how fast companies can actually deploy AI at scale. That matters for every business planning an agentic AI rollout in 2026 and 2027.

Third, this announcement lands one week before US President Trump and Chinese President Xi are scheduled to meet in Washington on September 24. Semiconductor policy is certain to be on the agenda. The timing of Huawei’s Connect summit is probably not a coincidence.

The Broader AI Chip Race

Huawei’s roadmap extends beyond the 960 series. The company outlined the Ascend 970 for 2028 and the Ascend 980 for 2029, targeting 288 GB and 384 GB of memory respectively. The consistent doubling of performance year over year, if it holds, means enterprises will have access to fundamentally more capable hardware every 12 to 18 months — regardless of which country their vendor is headquartered in.

For now, the immediate question is whether the 960DT delivers on the paper specs once it ships. Huawei’s prior Ascend generations had a gap between announced and real-world performance on enterprise workloads. The 960 family is designed for both training and inference, and the inference story — running large agentic workflows at scale — is where most enterprise buyers are focusing in 2026.

What This Means for Business

If you are planning significant AI infrastructure investment in 2026 or 2027, the main takeaway is this: do not lock in hardware assumptions based on today’s pricing or availability. The competitive dynamics in AI compute are shifting faster than most enterprise procurement cycles account for.

Specific decisions worth revisiting:

Cloud vs. on-premise. The cost per inference token on cloud platforms is partly a function of the underlying hardware. As Nvidia faces real competition, cloud providers will have more room to compete on price. Running your own hardware may become more viable in markets where Ascend chips are available.

Vendor lock-in. Enterprise AI platforms that are tightly coupled to Nvidia’s CUDA ecosystem become more of a risk as the hardware landscape diversifies. It is worth asking your AI vendors which compute back-ends they support.

Timing. Deploying large-scale agentic AI on hardware that will be superseded in six to nine months is not automatically a mistake — ROI can accrue quickly — but it is worth modeling the refresh cycle into your business case.

The companies that will get the most out of the AI compute explosion are the ones that understand these dynamics now, rather than discovering them when their CFO asks why the GPU bill doubled.

If you want to explore how AI agents can work within your existing infrastructure, the Omni by Enterprise DNA discovery call is a good starting point. We help businesses cut through the hardware hype and focus on what actually drives results.