Updates
A dated record of what changed in the guide and why. Each update names the figures that moved, the decisions they affect and the chapters it revised. Below the updates, the register feed lists every figure change as it is recorded.
Register current through .
No updates are published on this page yet. The register feed below lists recent figure changes. Subscribe to get each update by email.
Register feed
- 2026-10-19 Illustrative 3.0 MW power package, 2N versus 4N/3: ten-year cost at equal usable load:
Cost tie: about $3.6m each; required operating states decide→ Cost tie: about $11.9m each; operating states decide · Ch 12.1 - 2026-10-19 Derived keys enumerated in Gu et al.’s tested Hopper GPU-CC implementation:
44+ keys→ 44 derived keys - 2026-10-19 Company-reported hyperscaler fleet PUE, mixed cooling: AWS 1.14 (2025), Google 1.09 TTM (Q2 2026), Meta 1.08 (2024), Microsoft 1.17 (FY2025):
1.08–1.14→ 1.08–1.17 · Ch 15.2 - 2026-10-19 Platform voltage-droop improvement from backside power on Intel's PowerVia test chip, which also gained 6% in frequency:
~30%→ >30% · Ch 7.12 - 2026-10-19 NCCL all_reduce bandwidth Together AI expected on a 400 GB/s fabric (2024 rule of thumb); a golden-node baseline decides:
≈370 of 400 GB/s (~92%) in Together AI's published cluster-testing rule of thumb; the project baseline and declared tolerance set the gate→ ≈370 of 400 GB/s (~92%) · Ch 10.5 - 2026-10-19 Size a host pool from CPU service demand — result and flip threshold:
29 required cores; 100 GB host state. The 32-core pool passes the screen; its CPU ceiling is about 110 requests/s.→ 29 required cores; 100 GB host state. The 32-core pool passes the screen; its CPU ceiling is 112 requests/s. · Ch 7.8 - 2026-10-19 HPE GB200 NVL72 secondary-loop flow from a guide heat balance (115 kW liquid, OCP 7.5–12 °C rise):
~165–236 L/min→ ~140–220 L/min (water); ~144–230 L/min PG25 · Ch 7.13 - 2026-10-19 Worked example result: throughput passes, recovery misses its limit by 30 s:
HOLD: 1,050 tokens/s passes; 150 s recovery exceeds the 120 s limit by 30 s. Retest the full release after repair.→ Not accepted: 1,050 tokens/s passes; recovery takes 150 s against a 120 s limit (30 s over). Retest the full release after repair. · Ch 7.14 - 2026-10-19 DC-SCM puts BMC, RoT and TPM on a pluggable module; Rev 2.2 (September 2025) is the current specification:
DC-SCM 2.0→ DC-SCM Rev 2.2 · Ch 11.4 - 2026-10-19 Eclypsium reported roughly 1,000 internet-exposed instances in a Shodan search associated with its MegaRAC Redfish investigation; exposure is not confirmed exploitability of every instance:
1,000+→ ~1,000 exposed instances - 2026-10-19 Most a prewarmed split pool can cost per minute, against $1 for aggregated, when it passes 97 vs 95 requests:
Assumed: $1 per provisioned minute per candidate, including reserve→ ≈$1.021 per minute ($97/95; ≈2.1% premium) · Ch 10.11 - 2026-10-19 Worked case inputs: fixed program cash, saving per matched device, committed volume and service window (guide teaching budgets):
Design $300M + software $60M + integration $40M; sustaining $20M/year; delay reserve $40M; 3-year service window; 40,000 service-equivalent devices (30,000 downside); merchant $50,000/device; custom $35,000/device; delayed window 2 years.→ $500M fixed program cash; $15,000 saving per device; 40,000 devices (30,000 downside) over a 3-year window · Ch 7.5