Huawei's AI accelerator roadmap, claims that it makes Earth's mightiest clusters (www.theregister.com)

At its Connect conference Huawei unveiled a multi‑year Ascend AI-accelerator roadmap and new system form factors aimed at replacing reliance on Western GPUs. Key chips: the Ascend 950PR (Q1 2026) promises ~1 petaflop FP8 with 2 TB/s interconnect and 128 GB memory at 1.6 TB/s; the 950DT (Q4 2026) targets ~2 petaflops of FP4 with 144 GB at 4 TB/s; the 2027 Ascend 960 moves to 288 GB at 9.6 TB/s; and the 2028 Ascend 970 pushes memory bandwidth to 14.4 TB/s. Huawei also announced “SuperPoD” racks supporting 8,192 and 15,488 Ascend units and modular “superclusters” that scale to ~500,000 and over one million accelerators, implying custom high‑bandwidth memory (HBM) or domestically sourced equivalents and very high on‑chip/interconnect throughput optimized for large model training and inference. The roadmap is significant because China has barred local firms from buying U.S. accelerators, creating strong demand for domestic chips and giving cloud providers (notably Tencent Cloud, which is expanding overseas) an incentive to deploy Huawei hardware internationally. If delivered, these parts and hyperscale systems could undercut Nvidia/AMD on price and availability, particularly in developing markets, and change competitive dynamics for AI infrastructure. Caveats remain: independent users have reported that current Ascend parts don’t always meet claimed parity with Western GPUs, so execution and real‑world benchmarks will determine whether Huawei’s bold specs translate into credible, deployable alternatives.