Resurfaced Liang Wenfeng Q&A: DeepSeek built TileLang to escape CUDA, Huawei 4:1 vs GB300
teortaxesTex · x · 2026-11-15
teortaxesTex notes his May translation of DeepSeek founder Liang Wenfeng's closed-door investor Q&A is being recirculated as breaking news. Key claims: DeepSeek stopped relying on Nvidia's software stack during V3 via its own TileLang compiler ("Nvidia is digging its own grave"); Huawei's 950 Supernode matches GB300 latency but needs four Huawei cards per Nvidia card, lagging two years.
More from Companies & People
- Jacob Coxon, Anthropic's viral resignee, says he might return to an AI lab — rohanpaul_ai · 2026-11-15
- Jensen Huang's NVIDIA return: Kalshi says 9%, FutureSearch AI puts it at 7% — NathanpmYoung · 2026-11-15
- Musk confirms Ship 42 looks great after tower catch — elonmusk · 2026-11-15
- Musk posts then-and-now: Falcon 1 debris to Starship stacked on the tower — beffjezos · 2026-11-15
- Nvidia was already all-in on agents: internal AI factory serving all 40k employees — AccBalanced · 2026-11-15
- Jensen Huang hits Taipei's Raohe night market, poses with vendors now displayed as stall signage — minilek · 2026-11-15