Updates
The recency layer: material register changes and their decision consequence — each update deep-links into the chapters it updates. Register current through .
The first newsletter hasn't been sent yet — the guide is brand new. Subscribe below to get the first issue; meanwhile, the raw change feed follows.
What changed in the register
- 2026-08-26 power as a share of AI data center operating cost; the largest single opex line at scale:
25-60%→ ~25–65% · Ch 3.3 - 2026-08-26 cost-per-token / price-performance advantage cited by custom-silicon programs vs merchant GPU:
~30–50%+→ ~30–40%+ · Ch 7.5 - 2026-08-26 projected industry AI-asset D&A — exceeds combined hyperscaler 2025 profits (CONTESTED):
~$400B/yr→ ~$500B/yr · Ch 16.4 - 2026-08-26 spot/preemptible discount vs on-demand; the price of transferring interruption risk to the customer:
~60-80% off→ ~60–91% off · Ch 10.9 - 2026-08-26 Sequoia's annual revenue gap to justify capex (Cahn); widening, not closing, in 2026:
$840B→ ~$1.5T lifetime per capex-year · Ch 16.4 - 2026-08-26 data-center share of PJM's $16.4B capacity-auction cost — the cost-shift driver:
~$6.5B (40%)→ ~$6.3B (38%) · Ch 3.11 - 2026-08-26 time-to-power: greenfield self-build vs wholesale colo (live 50k+ GPU cluster) vs neocloud:
Path- and project-specific→ ~24–36 mo once power is secured (queue adds 4–7 yr in dense hubs) / ~6–12 mo fit-out / days–weeks · Ch 1.6 - 2026-08-26 reciprocating-engine lead time (needs no boost compression) vs 18–36 mo+ for aeroderivative turbines:
12–18 mo→ 15–24 mo · Ch 4.9 - 2026-08-24 GPU utilization in a well-orchestrated AI cluster vs ~50-60% traditional enterprise:
85-90%→ 55.2% MFU (named MegaScale run) · Ch 16.3 - 2026-08-24 QCT GB200 NVL72 QoolRack reference maxima: 45 °C liquid inlet and 65 °C liquid return; separate limits, not a selected operating pair:
45 °C maximum supply; 65 °C maximum return (separate limits)→ 45 °C maximum liquid inlet; 65 °C maximum liquid return (separate limits) · Ch F - 2026-08-24 GPU racks deployed at Colossus Phase 1 — implies a sustained ~12 racks/day set-and-cable rate over 122 days:
~1,500 racks→ ~1,563 rack-equivalents; ~12.8/day · Ch 7.15 - 2026-08-24 HBM as share of DRAM revenue (~30%) vs share of DRAM bits (~8%) — why fabs prioritize it:
~30% / ~8%→ >30% value / >10% bits (2025 forecast) · Ch 7.6