Ex-Google expert: data center capacity and electricity, not GPUs, are the real AI bottleneck
RihardJarc · x · 2026-10-20
An AlphaSense expert call with a former $GOOGL employee (apparently tied to Fireworks) argues the real AI compute bottleneck is US data center capacity and electricity, not GPU production.
Key points:
- Open-source models closing in fast: since Q2, customers have been replacing Anthropic/OpenAI models in multi-agent architectures with OSS alternatives (starting with Qwen2-7B/9B for tool calling and MCP), moving up to low-level orchestration; the transcript cites Cursor's similar shift, Composer 2.5's strength, and GLM 5.2 / Kimi K3 as the next wave.
- Capacity is the top constraint: some companies' traffic grew 5x while US GPU availability lags; overseas capacity in Asia/Australia adds unacceptable latency.
- Three-tier supply: hyperscalers are reliable but expensive with 3+ year commitments and scarce B300s; neoclouds (Nebius, Lambda, CoreWeave, etc.) host 80-90% of capacity but suffer reliability issues — GPUs silently dropping at 1-2 AM require round-the-clock ops; the third tier of "neo-neo clouds" (4,000 GPUs per site) lacks real cloud expertise and is essentially unusable.
- Commercial terms: US neoclouds typically require 20-30% prepayment with month/quarterly payments and end-of-term rebates for reliability incidents; APAC providers require no prepayment.
- Outlook: GPU availability rises meaningfully in 1-2 years, but US power is the hard limit; Meta and xAI are beginning to lease out spare GPU fleets, with more to follow.
More from Infra
- Victor Taelin publicly begs to buy an 8-GPU Rubin machine — 'nobody is selling it' — fforres · 2026-10-21
- Nscale joins PyTorch Foundation as Platinum member after Anyscale acquisition move — PyTorch · 2026-10-21
- Decision Models launches on Antseed with an open-source auto model router — airesearch12 · 2026-10-21
- Gradium open-sources Phonon: 31M/40M edge TTS models with Rust inference stack for 10+ backends — l-m-z · 2026-10-21
- PostgreSQL 19 slips to November as LLM-assisted code review floods committers with edge cases — sanjaykalra · 2026-10-21
- Eon cuts engineer onboarding from 2 days to 2 minutes with an MCP Gateway on Bedrock AgentCore — AWS ML Blog · 2026-10-20