Anthropic's AI ported a model to AMD chips in 3 days, a job that used to take 6-12 months
aakashgupta · x · 2026-11-12
An Anthropic engineer plugged in AMD MI355X chips, had the company's AI port the model over a weekend, and skipped AMD CEO's offer to help. Anthropic's IPO materials show 3 days vs the 6-12 months CUDA porting traditionally takes. Caveats: only prompt processing was ported, not generation or training. But the business math matters: Nvidia's $96B quarter at 75% gross margin rests on switching pain, and Anthropic runs compute as a single fungible pool across Nvidia/AMD/Google/Amazon, paying SpaceX $3.1B/month — every 5% price cut is worth $1.9B. Nvidia committed up to $10B to Anthropic last November. A moat measured in engineer-months is only as deep as your cheapest engineer, and Anthropic's cheapest is now the model.
Related event: Anthropic IPO Filing Reveals Research Acceleration Threshold Reached(5 posts)→
More from Venture
- Claim: 45% of Nvidia chip sales went to Neo-Clouds and China, now cut off — SumitGup · 2026-11-14
- Chart shows network effects in action: 0 to 33M users in six months — jacob_posel · 2026-11-14
- The Bank of Nvidia: ZeroHedge maps $100BN+ in circular AI deals after Huang exit — SumitGup · 2026-11-14
- Conviction wraps 7th Embed batch, lists 8 AI startups including Instrumental and Flow — saranormous · 2026-11-14
- Mnemonic Labs (YC F26) launches to train a foundation model for quant trading — ycombinator · 2026-11-14
- Tech bond yields hit highest since 2009 as AI funding rush shows strain — SumitGup · 2026-11-14