Summary by TwinMind · 5m 14s
Summary by TwinMind · 5m 14s

$10B OpenAI & Microsoft Cluster in 2025 – Dylan Patel & @Asianometry

$10B OpenAI & Microsoft Cluster in 2025 – Dylan Patel & @Asianometry

Dwarkesh Patel

Dwarkesh Patel

Training Regime Parallelization Evolution

  • Training compute shifting toward synthetic data generation and search processes
  • Post-training techniques emerging to utilize available compute capacity
  • Multi-data center training becoming viable solution for scale

Microsoft-OpenAI Infrastructure Expansion

  • Microsoft signed $10+ billion deals with fiber companies (Lumen, Zayo, others)
  • Permits filed for inter-data center connections
  • 5 regions being connected comprising multiple data centers each
  • Total power consumption: north of 1 gigawatt (close to 1 million GPUs)

GPU Power Consumption Scaling

  • H100 GPU: 700W base, 1,200-1,400W total system power
  • Next-gen Nvidia GPUs: 1,200W base, 2,000W total system power
  • Power per GPU increasing with each generation

Current and Planned Cluster Deployments

  • Existing 100K H100 clusters:
    • OpenAI in Arizona
    • xAI in Memphis
    • Multiple others under construction
  • Planned GB200 clusters:
    • 5+ planned 100K GB200 clusters by Microsoft/OpenAI partners
    • Potential 500K GB200 cluster (1 gigawatt capacity)
    • Timeline: online next year

2025-2026 Scaling Projections

  • 2025: 100K GPUs largest current cluster
  • Next year: 300K-500K GPUs per cluster (up to 700K upper bound)
  • GB200s are 2-3x faster than H100s
  • H100-equivalent basis: 1 million chips next year in single cluster
  • Multi-site efficiency loss: 20-50% (exact figure unknown)

Long-term Infrastructure Plans

  • 2026: Single 1-gigawatt site planned
  • Microsoft 5-gigawatt total capacity across multiple locations
  • Each location: multiple sites, some single sites
  • Total capacity: easily north of 2-3 gigawatts
  • Integration of old chips with new chips under consideration

Financial Requirements

  • OpenAI needs $50-100 billion funding for planned cluster scale
  • Expected fundraising: end of 2024/early 2025
  • Sam Altman's strengths: recruiting and fundraising capabilities

Competitive Landscape

  • Anthropic has similar plans with Amazon
  • Multiple providers involved: QTS, Cooper, others
  • Self-built data centers also being constructed
  • Old crypto mining sites being repurposed (CoreWeave Texas, Oracle Crusoe Texas)
  • Locations: Wisconsin, Arizona, Texas, others