Huawei Moves Up Ascend 960 AI Chip Launch to Q1 2027

Huawei just pulled its next AI chip forward by nine months. At its Huawei Connect conference in Shanghai on September 17, rotating chairman David Wang said the Ascend 960DT, originally slated for the second half of 2027, is now ready to ship in Q1 2027, and its sibling chip, the 960PR, follows in Q3. The company also laid out chips for 2028 and 2029, locking in a one-generation-a-year cadence aimed squarely at Nvidia's grip on AI compute.

• Huawei moved the Ascend 960DT launch up three quarters, from a planned Q3 2027 to Q1 2027.
• A second chip, the Ascend 960PR, ships Q3 2027; both were announced at Huawei Connect on September 17, 2026.
• Huawei confirmed a fixed yearly cadence beyond that: Ascend 970 in 2028, Ascend 980 in 2029.
• The chips lean on Peerium, a new architecture meant to eventually link up to a million processors as one system, and on Huawei's own HBM, HiBL and HiZQ, built in-house because US export controls cut off the usual suppliers.

What exactly did Huawei announce?

Two chips and a schedule change. The Ascend 960DT, a chip aimed at AI training and inference, moves from a planned second-half-2027 launch to Q1 2027, a nine-month pull-forward that Wang attributed to demand outstripping Huawei's own manufacturing capacity. Its counterpart, the Ascend 960PR, arrives on the original Q3 2027 timeline. Both use a new SIMD/SIMT architecture that adds support for a wider set of number formats, including FP8, MXFP8, HiF8, MXFP4 and HiF4, the low-precision formats that make large models cheaper to run. Huawei also confirmed the Ascend 970 for 2028 and the Ascend 980 for 2029, turning what used to be a series of one-off announcements into a standing yearly release schedule.

RelatedGemini 3.8 Flash Model Card Leaks Ahead of Google Launch

Why is Huawei racing to move the date up?

Because the chip alone isn't the product anymore. Huawei introduced Peerium, a computing architecture designed to eventually wire up to a million accelerator cards into one logical system, built on UnifiedBus, the interconnect standard Huawei has now shipped in eleven separate chips. The Ascend 960 supernode is Huawei's first to use near-package optical interconnects, packing 4,096 cards into a single node rated at 8 exaFLOPS of FP8 compute and 1 petabyte of HBM capacity. That's the actual pitch: Nvidia still leads on raw chip performance, so Huawei is trying to win the argument at the system level, wiring together enough of its own chips that the single-chip gap stops mattering as much.

What's the HBM story, and why does it matter this much?

High-bandwidth memory is the part of an AI chip that decides how fast it can actually feed data to its compute cores, and it's also the part US export controls have made hardest for Huawei to buy from SK Hynix, Samsung or Micron. Starting with the Ascend 950PR, Huawei is shipping its own HBM, branded HiBL and HiZQ, developed in-house specifically because the usual suppliers were cut off. The 2028 Ascend 970 is where that bet starts to show results on paper: compute roughly doubles to 4 PFLOPS at FP8 and 8 PFLOPS at FP4, interconnect bandwidth reaches 4 TB/s, and while HBM capacity holds at 288 GB, its bandwidth climbs to 14.4 TB/s. The 980 in 2029 is reported to push capacity further, toward 384 GB. None of this is independently benchmarked yet since the chips don't exist outside Huawei's own labs, but it is a concrete, dated commitment to closing a memory gap that has been Huawei's weakest link against Nvidia's Blackwell and Rubin generations.

1. 2025Ascend 910C ships Huawei's current-generation training and inference chip
2. 2026Ascend 950DT and 950PR first generation on self-developed HiBL/HiZQ HBM
3. Q1 2027Ascend 960DT launches moved up from a planned Q3 2027 on demand outstripping supply upcoming
4. Q3 2027Ascend 960PR launches on its original schedule; 4,096-card supernode, 8 EFLOPS FP8, 1PB HBM
5. 2028Ascend 970 4 PFLOPS FP8 / 8 PFLOPS FP4, 4 TB/s interconnect, HBM bandwidth to 14.4 TB/s
6. 2029Ascend 980 HBM capacity reported to reach 384 GB

Who does this actually affect?

Chinese cloud providers and AI labs first. Baidu, Alibaba, Tencent and the state-backed AI push around models like DeepSeek and Huawei's own Pangu all need training and inference silicon they can buy without a US export license, and right now Ascend is the most credible domestic option at scale. It also affects the handful of countries and companies trying to build AI infrastructure outside the US-Nvidia supply chain, since a system that can plausibly link a million accelerators is a pitch aimed at sovereign compute buyers, not just Chinese customers. Nvidia isn't losing customers in the US or Europe over this news, but the addressable market Nvidia can't legally sell into in China just got a faster-moving domestic alternative.

RelatedDeepSeek Pauses $1.5B Raise After Leaked Compute Remarks

What it means for the market

Huawei is private, so there's no ticker to watch directly, but the ripple effects land on public names. SMIC, Huawei's main domestic foundry partner and Hong Kong-listed, is the clearest proxy for Ascend production volume; a pulled-forward 960DT launch implies SMIC's advanced-node capacity is scaling faster than skeptics expected. On the other side, Nvidia's China-specific revenue has already been squeezed by export licensing, and a credible, system-level Huawei roadmap is the argument bears use for why that share doesn't come back even if licensing eases. The signal worth watching isn't the chip's benchmark numbers, which nobody outside Huawei can verify yet. It's whether SMIC's yields and Huawei's own HBM supply can actually hit these dates, because Huawei has missed its own Ascend timelines before.

Our take

The headline number, nine months pulled forward, is real but it's also the easy part of this story to overstate. Announcing a Q1 2027 date at a Huawei conference and shipping working silicon at volume in Q1 2027 are different achievements, and Huawei's own chip history has both early triumphs and slipped promises in it. What's more interesting than the date is the shape of the strategy: Huawei has stopped selling individual chips as its answer to Nvidia and started selling a systems argument, a million-card fabric plus in-house memory plus a fixed yearly cadence. That's a slower, harder pitch to fake than a single chip's spec sheet, which is exactly why it's worth taking seriously even before the 960DT actually ships.

Original analysis by GenZTech, based on Huawei Connect 2026 announcements and independent industry reporting.