DeepSeek bets next model on Huawei chips after Ascend 910C training run failed
soumitrashukla9 · x · 2026-09-22
Per leaked meeting details, DeepSeek is betting its next model on Huawei silicon after a failed attempt: Liang Wenfeng told investors Huawei should start delivering training chips as early as Q4 and "it has to work" — an earlier run on Ascend 910C failed, and DeepSeek has trained on Nvidia since.
Key points:
- Investors had to hand over phones; the last meeting leaked, including Liang saying DeepSeek could buy "noncompliant cards"
- A $7.5B round is being finalized at a $75B valuation, $4.5B going to compute
- Liang estimates a frontier-scale run needs 50,000 Nvidia GB300s or 200,000 Huawei Ascend 950s
- Huawei's bottleneck is memory: only Samsung, SK Hynix, and Micron make the HBM in the most advanced chips
More from Venture
- Strategy 101: incumbents tie complements, entrants break them — enter Muse — Afinetheorem · 2026-09-22
- Why Shopify says yes to Muse and Amazon says no: complements economics — Afinetheorem · 2026-09-22
- Deel launches Akai: 8,000 agents doing work of ~600 employees, added $140M ARR in 90 days — AIwithGhotai · 2026-09-22
- YouTube Premium's weak uptake is the strongest proof consumer AI agents can't be subscription-based — ivan_bezdomny · 2026-09-22
- APIHub Launches Pay-Per-Call API Marketplace for AI Agents With USDC — modelcontextprotocol · 2026-09-22
- SB Energy's $50B-Valuation IPO for World's Largest Data Center Project Delayed on Weak Investor Demand — firstadopter · 2026-09-22