What this is

This technical-validation view is an illustrative decision model of physical-infrastructure-induced AI compute continuity for a sample 5 GW AI data-center campus with 7.7 GW of dedicated generation. It exposes workload targets and input provenance so a technical lead can identify what must be replaced, reconciled, and approved. It is not operational software, project engineering, an OEM guarantee, an actuarial model, a tariff or interconnection analysis, or a representation of any company’s actual design, plans, or endorsement.

This decision model does not connect to or control a data center. A future operational implementation would integrate into an operator’s DCIM environment through governed APIs only after site engineering, controls architecture, cybersecurity, data quality, commissioning, and operating-authority validation.

Sample campus setting [ILLUSTRATIVE]

• Full-build example: 5,000 MW thermal AI cooling load and 7,700 MW of dedicated generation; phase buttons show an illustrative ramp, not a disclosed construction schedule.
• The “current design path” is a generic comparison baseline without a multi-hour stored-cooling layer. It is not a characterization of any named operator or project.
• The SEGI path combines stored cooling, turbine inlet-air chilling, controls, and operating doctrine. The thermal battery stores cooling—not electricity—and grid support is demand reduction rather than power export.

Key assumptions [EST / DOCTRINE / DESK — every one adjustable or documented]

• Cooling load ≈ IT electrical load at first order; campus phases (1,000 / 2,500 / 5,000 MW thermal) are an illustrative ramp, not a disclosed schedule.
• GPU density 375–545 GPUs per MW (GB200-class rack basis); rack value $2–3M; GPU-hour value $3–10.50 (default $7).
• Chiller displacement 0.171–0.227 MW per MW of cooling (284.3 tons/MW at 0.6–0.8 kW/ton), less a 7.5% auxiliary allowance for pumps, CDUs, and controls (auxiliary power remains required during discharge).
• Hot-weather output at risk 8–20% of gas-fleet capacity (ASHRAE / NREL / Texas A&M ranges); combined-cycle specifics need project modeling.
• Inlet-air chilling cooling duty ≈ 1.5 MW thermal per MW of generation recovered (EST, humid-climate latent-load basis), drawn from the same finite stored-cooling reserve — so a deep derate or a short autonomy window only partly averts the loss; flagged for engineering.
• Recovery tail anchored to the Google London July 2022 incident (4h08m cooling impact; 18h23m service restoration; 35h15m long tail).
• Installed-cost anchor $425–450M per 750 MW-thermal module, scaled linearly — a proof-of-concept assumption flagged for engineering.
• Event probability is deliberately user-owned; no public actuarial basis exists.
• Selected-scenario annual ppm = annual event probability × equivalent full-load compute-loss hours ÷ 8,760 × 1,000,000. ACRI then compares that result with the binding workload’s permitted annual downtime. Use the training-share and inference-target controls to test the operator’s workload policy; any non-zero training share binds at five nines, or 10 ppm (about 5.3 minutes/year).
• The 1–10 score is a transparent presentation band based on the share of the binding workload budget consumed. Scores 1–5 are within the selected-scenario budget; 6–10 exceed it. ACRI is not an industry standard, actuarial rating, or all-hazards certification.
• Hot-weather derate is not ACRI-scored until project load-flow, dispatch, OEM performance, and workload evidence establish a compute-interruption curve. Grid emergency is a flexibility case, not a compute-interruption case, and is not ACRI-scored.
• Each input is marked by provenance: doctrine, public range, incident, proxy, statutory reference, or required site/utility data. Red site-data tags identify values that must be replaced before production use.

Mechanism honesty

• The thermal battery stores and discharges cooling, not electricity. “Demand handed back” is displaced chiller load—demand reduction, not export. Project-specific grid-service quantity and duration require utility and engineering confirmation.
• The reserve provides cooling autonomy, not compute runtime; compute continuity follows from avoided forced shutdown. "Minutes" of ride-through is the operational response window, not a universal damage clock.
• Stored cooling is finite, and the model enforces it. If an event moderately outlasts the window, the protected campus sheds load and paces the discharge so the reserve lasts to restoration; if the event far outlasts it, the campus goes through an orderly shutdown — offline, but checkpointed and without thermal excursion, so restart is a short managed ramp rather than the multi-hour recovery tail that follows an emergency power-down. The reserve never carries load it has no stored cooling to support. Window length is a sizing decision (2.5–4 h doctrine range; a ~2 h bridge is engineer-affirmed feasible).
• Value lenses are never summed. The live loss meters show only accrued AI compute interruption cost from the capacity curve; hardware, rollback, recovery-tail, grid, and capital lenses remain separate.
• ACRI is annualized modeled risk, not a meter that rises because an incident happened. During playback, the dashboard reveals how each additional hour of modeled shortfall increases the selected event’s contribution under the chosen annual probability.
• Scenario values are never summed. The dashboard does not claim an all-hazards composite because event dependence, common-cause failure, and site-specific frequency require a governed risk register and defensible correlation treatment.

Deliberately excluded

Fuel and dispatch economics, LCOE, OEM performance curves, hourly weather data, tariff/DR program design, O&M, financing structure, water, CO₂ accounting, construction schedule. A production version would add project-specific engineering in each of these lanes.

Reproducibility

The model is self-contained. The methodology dialog and technical-lead evidence register document the equation, score bands, workload budget, scenario boundaries, input provenance, and production source requirements so a reviewer can reproduce or replace every assumption.