TL;DR
Huawei has unveiled the Ascend 960 accelerator family and a 4096-chip "Superpod" server reaching 16 Exaflops, signaling China's most aggressive push yet to close the AI compute gap despite crippling export restrictions.
What happened
- Ascend 960DT (training) targets Q1 2027 delivery; Ascend 960PR (inference) follows in Q3 2027, ahead of earlier expectations.
- The 960DT is Huawei's first publicly known chip with HBM: 288 GB across eight stacks, 9.6 TB/s bandwidth, 4 Petaflops FP4 peak.
- The 960PR doubles FP4 compute to 8 Petaflops but cuts memory to 192 GB at 2.4 TB/s, likely using cheaper LPDDR RAM for cost-sensitive inference workloads.
- The Atlas 960E Superpod clusters 4096 Ascend 960 chips, hitting 16 Exaflops FP4 and over 1 Petabyte of HBM in the all-DT configuration.
- Huawei uses on-chip optical engines to interconnect accelerators and has stated a long-term goal of one million Ascend chips in a single system.
Why it matters
- Single-chip performance gap remains severe: one Nvidia Rubin GPU delivers 50 FP4 Petaflops and one AMD Instinct MI455X delivers 40, meaning it takes more than ten Ascend 960DTs to match either rival chip.
- Huawei's answer is horizontal scale, not silicon parity: the Superpod strategy trades per-chip efficiency for sheer accelerator count, a viable path for Chinese hyperscalers locked out of Western silicon.
- HBM sourcing is an open question: CXMT's HBM3e is a candidate but limited in production volume, making memory supply a potential bottleneck for any large deployment.
- SMIC's 7 nm process versus Nvidia's 3 nm and AMD's 2 nm means Huawei carries a structural power and density disadvantage that scale alone cannot fully offset.
- A published annual roadmap through 2029 (Ascend 970 at 14 Petaflops in 2028, Ascend 980 at 28 Petaflops with 384 GB at 38.4 TB/s in 2029) signals institutional commitment and gives Chinese cloud buyers a credible domestic alternative to plan around.
What to watch next
- HBM supply confirmation: whether CXMT can ramp HBM3e at scale will determine if Huawei can actually ship Superpods in volume by Q1 2027.
- Customer adoption signals: watch Chinese hyperscalers (Alibaba, Baidu, Tencent) for procurement announcements that would validate the Superpod's real-world competitiveness.
- Export control escalation: any tightening of restrictions on SMIC's tooling or CXMT's HBM materials could disrupt the entire roadmap before the Ascend 970 generation arrives.
Originally published on Present of AI, a daily source-linked AI news timeline. Read the full timeline or browse the open dataset.