The Definitive Guide toAI Data Centers
Ask the GuideAboutAccount

Dispatches

The recency layer: what changed and why it matters — each dispatch deep-links into the chapters it updates. Published when register figures actually move, not on a calendar; recent cadence has been roughly weekly.

  1. 2026-07-07

    Transformer queues hold; interconnection stretches past 8 years

    What moved in the register: HV transformer lead times hold at 128–160+ weeks, PJM application-to-COD passes 8 years, the GPU:CPU ratio keeps falling, and Stockholm's waste-heat target slips to 2035.

  2. 2026-07-03

    Register corrections: MTTI at 16k-GPU scale revised down to ~186 min; MI355X FP4 up; NVL72 cooling flows revised down

    What moved in the register: the Llama-3 405B mean-time-to-interrupt was corrected down to ~186 min (interruptions more frequent, not fewer), MI355X FP4 to ~10.1 PFLOPS, and GB200 NVL72 secondary-loop flow to ~170–235 L/min.

What changed in the register

  • 2026-08-13 revenue per energized GW of AI capacity; ~$1B from 200 MW arriving 6 months early (contested — single-source): ~$10–13B/GW/yr → ~$12–13B/GW/yr · Ch 16.1
  • 2026-08-13 Microsoft server-and-component reuse-and-recycle rate (FY2025, second consecutive year above the 90% target, hit 2025 target a year early): 90.9% → 92% · Ch 15.6
  • 2026-08-13 post-FEC BER pass floor for AI fabric links (tightening toward 1e-13 at the highest lane rates): ~1e-12 → ≤1e-13 · Ch 13.7
  • 2026-08-13 single-program lifecycle capacity (CAR + DSU + operational PD/BI) assembled because single carriers can't cover a hyperscale campus: $3.5B → up to $5B · Ch 2.6
  • 2026-08-01 L2L heat-rejection capacity vs L2A, at under ~2x the cost at scale: 8–10x → 6.6x · Ch 5.6
  • 2026-08-01 Marvell total FY2027 revenue guide, AI-driven (custom AI silicon ~$1.5–2B now; >$10B targeted by FY2029); Broadcom AI-revenue growth (Q2 FY2026 YoY): ~$11B / +143% → ~$11.5B / +143% · Ch 7.5
  • 2026-08-01 CoreWeave backlog vs revenue — single-anchor concentration; FY25 rev $5.13B, interest ~40% of adj. EBITDA: ~$66.8B (~13x) → ~$99.4B (~19x) · Ch 2.5
  • 2026-07-27 network share of Llama 3 interruptions — switch/cable 8.4%; 10.0% including NIC: 8.4% → 8.4% (10.0% incl. NIC) · Ch 10.2
  • 2026-07-27 incremental revenue from bringing 200 MW online ~6 months early (~$10–13B/GW/yr) (contested — single-source): ~$1–1.2B → ~$1.0–1.3B gross (NPV ~$400–500M) · Ch 6.4
  • 2026-07-27 KV-cache memory reduction from attention architecture and quantization; DeepSeek-V2 reported a 93.3% reduction (~15×) from MLA versus its DeepSeek-67B baseline: 4–40x → 93.3% (~15x) · Ch 9.7
  • 2026-07-27 construction-waste reduction from factory prefab vs on-site (one study: ~79% across 59 projects): up to ~90% → 78.8% · Ch 6.4
  • 2026-07-27 global data center capex 2026, approaching — volatile market figure; analyst estimates differ by capex-scope definition: ~$1T → >$1T · Ch 0.2