DeepSeek compute projection: >1GW Blackwell-equivalent by end of 2027 on Huawei Ascend
teortaxesTex · x · 2026-10-20
A projection thread on Huawei Ascend supply and DeepSeek's compute roadmap. If "most" early 960DT supply means 2/3, that implies >1M 960DTs and well over 2GW. The accompanying analysis reconciles sources (Reuters: 750k 950PR as 2026 shipment plan; Epoch AI: 320k 950DT modeled; Bloomberg: DeepSeek's planned 160k-card Inner Mongolia deployment ≈ half a year's 950DT output), stressing that 950PR vs 950DT memory configs differ hugely and total Ascend output is a poor proxy for frontier-training supply; the 960DT is a memory-intensive upgrade (up to 288GB, 9.6TB/s, 2 PFLOPS FP8, Q1 2027). Author's optimistic take: DeepSeek will have >1GW Blackwell-equivalent compute by end-2027.
Related event: Report: Most Huawei 960DT Supply in 2027 to Fuel DeepSeek Superclusters(3 posts)→
More from Infra
- doubleAI, co-founded by Mobileye CEO, claims its model cuts compute costs by 67% — pstAsiatech · 2026-10-20
- US sanctions 5 Chinese entities tied to ex-Google engineer who stole TPU cluster secrets — rohanpaul_ai · 2026-10-20
- AI chip startup Tenstorrent raises $1.5B at $6.2B valuation after takeover interest — pstAsiatech · 2026-10-20
- Fresher Lemonade: CLI tips for pulling, configuring and benchmarking local models — admcpr · 2026-10-20
- BigMoeOnEdge streams 91B+ MoE models on 12GB phones and 16GB Macs, hitting 25-30 tok/s — dai_app · 2026-10-20
- NVIDIA Brev lets you add your old gaming PC to a unified GPU console alongside cloud instances — victorialslocum · 2026-10-20