Monday Jul 20
HARDWAREOther

Huawei Links 1,024 Chips Into One Machine

20JUL
8,192 CHIPS AT SCALEHUAWEI6.7X CLAIM

China's clearest answer to Nvidia's chip lead just got bigger. Huawei linked 1,024 Ascend chips into one supercomputing unit. Huawei claims 6.7 times the compute of Nvidia's next-gen cluster.

The design scales up to 8,192 chips linked together. Huawei calls it the industry's largest AI computing supernode.

That 6.7 times number is Huawei's own benchmark, not an independent one. Huawei's rotating chairman admits a real gap remains at the single-chip level.

This is China's system-level answer to US export controls on chips. Stack enough smaller chips together, and single-chip performance matters less.

full brief & sources

⚡ Why this matters

  • Shows China's chip strategy: win at the system level even while behind on single chips.
  • A credible system-level challenge to Nvidia's supply-constrained dominance changes buyer leverage.
  • Fits the long arc of the AI infrastructure race, not just a product launch.

🔍 What happened

  • Huawei unveiled the Atlas 950 SuperPoD, linking 1,024 Ascend chips into one unit.
  • The architecture scales up to 8,192 linked Ascend chips per deployment.
  • Huawei claims 6.7 times the compute and 15 times the memory of Nvidia's NVL144.
  • Huawei's rotating chairman acknowledged a short-term gap in single-chip performance versus Nvidia.
  • A Korea launch is reportedly planned for the fourth quarter of 2026.

💬 Smart takes

  • Huawei rotating chairman: there is a short-term gap in single-chip performance compared with Nvidia.
  • Skeptic: Huawei's own benchmarks against Nvidia's next-gen cluster aren't independently verified, and system-level claims are easy to inflate.

🧭 Where this goes

  1. LikelyHuawei pushes the SuperPoD hardest into markets cut off from Nvidia by export rules.
  2. Possibleindependent benchmarks narrow the gap between Huawei's claims and real-world performance.
  3. PossibleSouth Korea becomes a real test market once the Q4 2026 launch lands.
  4. Wild Carda major non-Chinese cloud provider adopts Ascend clusters at scale within 18 months.

🥄 The Spoon Take

Huawei can't out-build Nvidia chip for chip yet, so it's out-linking it instead, wiring thousands of weaker chips into one machine. If that math holds up outside Huawei's own slides, export controls just got a lot less effective.

🤔 Pushback

Every number in this story comes from Huawei, and nobody outside the company has run these chips against real workloads yet.

·Huawei Central - Atlas 950 spec claims·Seoul Economic Daily - SuperPoD unveiling