$10B OpenAI & Microsoft Cluster in 2025 – Dylan Patel & @Asianometry
Watch on YouTube ↗ Summary based on the YouTube transcript and episode description.
Dylan Patel and Jon (Asianometry) detail Microsoft and OpenAI’s $10B+ fiber-connected multi-datacenter supercluster and predict OpenAI raises $50-100B imminently.
- Microsoft signed $10B+ deals with fiber companies (Lumen, Zayo) to connect five multi-datacenter regions into one logical training cluster.
- Total connected capacity will exceed 1 GW of power, approaching ~1 million GPUs.
- Next-gen Nvidia GPUs draw ~2,000W all-in vs ~1,200–1,400W for H100, compressing per-GPU cluster economics.
- Largest single cluster in 2025: ~100K GPUs (OpenAI Arizona, xAI Memphis already exist); 2026 upper bound: 300K–700K GPUs, ~1M H100-equivalents.
- Microsoft’s full five-region buildout reaches 5 GW by 2026; a single 1 GW site comes online in 2026.
- OpenAI cannot finance its planned 2025–2026 clusters without raising $50–100B, which Patel expects by early 2025.
- Anthropic has analogous large-scale infrastructure plans with Amazon.
- Multi-datacenter training efficiency loss is an open question; Patel believes it is solvable but admits the exact penalty is unknown.
2024-10-09 · Watch on YouTube
