{"resourceId":"arxiv-2606-25098-power-flexible-ai-dc","versions":[{"version":"external-47677170757a3e9d1f674382b7d41fc924adb86ebc567db1e01896797e228fcc","resource":{"id":"arxiv-2606-25098-power-flexible-ai-dc","title":"arXiv:2606.25098 — Power-flexible AI data centers: measured UK compliance and US geo-shifting","organization":"Emerald AI, EPRI, National Grid, NVIDIA, Oracle (multi-affiliate authors)","sector":"Academic / industry research on grid-interactive AI compute","geography":"United Kingdom and United States experimental deployments","publishedAt":"June 23, 2026 (arXiv 2606.25098v1)","publicationDate":"2026-06-23","eventDate":"2026-06-23","sourceName":"arXiv","sourceLabel":"Preprint (IEEE submission noted)","sourceUrl":"https://arxiv.org/html/2606.25098v1","evidenceClass":"academic-research","outcomeClass":"mixed","topics":["infrastructure","operating-model","data-security"],"finding":"Preprint argues GPU AI data centers can act as grid-interactive assets via integrated grid signals, workload orchestration, and power telemetry. Experimental claims include: 100% compliance across 200+ distinct National Grid power events on a 130 kW NVIDIA Blackwell Ultra cluster; approximately 30% power reduction within 40 seconds in emergency load-reduction testing; and US geo-shifting that migrated 10% of live inference traffic from an Oracle H100 cluster in Virginia to Illinois during a Dominion Energy winter peak—while preserving priority service levels in the described trials.","sledRelevance":"Why now: supplies the academic primary behind Sep Flex/AEMA partner narratives with quantified response times and compliance rates SLED facilities/utilities can cite—while making clear trials are 130 kW-class / partial-traffic, not campus gigawatt proof. Transferability limited by utility protocols and workload mix.","evidence":"Abstract and intro summarize 130 kW deployment flexibility modes; contributions list UK 200+ event compliance and 10% Virginia→Illinois inference shift; body reports ~30% cut in ~40 seconds for emergency scenarios.","architectureImplications":"Needs seconds-level GPU power telemetry, dispatcher integration, and geo-routing for inference when spatial flex is used.","governanceImplications":"IEEE-submitted preprint—not yet a final peer-reviewed version here; treat numbers as author-reported pending venue acceptance. Utility partnerships shape what signals are legal/operational.","securityPrivacyImplications":"Geo-shifting inference may cross jurisdictional data boundaries—paper focuses on power, not privacy transfer limits.","caveats":"Author set includes Emerald AI and NVIDIA; 130 kW vs commercial AI factories; preprint status; carbon-aware and multi-day claims need careful reading of methods sections.","streamIds":["nvidia"],"roles":{"sales":"Use as the measured-methods companion to Flex marketing: cite 200+ event compliance and 30%-in-40s only with scale caveats. Do not sell geo-shift without data-residency analysis. Ideal for utility-facing discovery, not as a campus ROI calculator.","engineering":"Fit for research/AI-factory teams designing flex control loops. Integrate grid event APIs with GPU power caps and scheduler preemption; prerequisites include telemetry cadence and priority classes. Proof: reproduce compliance % and response latency on a local pilot cluster.","delivery":"Run protocol-aligned drills with utility partners; log every event. Ownership: research computing / AI factory ops with utility co-ownership. Risks: preprint revisions; privacy blocks on geo-shift; over-generalizing 130 kW results."},"roleDetails":{"sales":{"customerProblem":"Utilities and boards demand evidence that AI loads can flex, not just densify.","stakeholders":["facilities","utility","research computing","risk/compliance"],"discoveryQuestions":["What dispatch protocols does your utility use?","What fraction of load is delay-tolerant?","Any data-residency limits on geo-shift?"],"valueHypothesis":"Documented flex compliance can unlock interconnection options if scaled carefully beyond 130 kW demos.","potentialEngagement":"Methods-aligned flex measurement study with local utility on a pilot cluster.","unsupportedClaims":["Gigawatt-scale proof","Final IEEE peer-review status","Privacy-safe geo-shift everywhere"]},"engineering":{"fit":"Instrumented GPU clusters with mixed job classes; weak without telemetry or utility signals.","architectureIntegration":"Grid interface → orchestrator → power caps/scheduling → optional geo inference router.","prerequisites":["seconds-level power telemetry","job priority metadata","utility signal path"],"constraints":["130 kW trial scale","preprint","affiliate authorship"],"security":"Protect dispatch interfaces; assess cross-region data movement separately.","proofOfValue":["Event compliance rate","Response time to target kW","Priority SLO hold"]},"delivery":{"work":"Implement control loop, execute scripted National-Grid-like events, publish traces.","dependencies":["utility/EPRI-like partner","cluster scheduler hooks","networking for geo-shift if used"],"ownership":"AI infrastructure lead with facilities and utility liaison.","skillsAdoption":"Power-systems literacy for AI ops teams.","governanceCheckpoints":["Preprint vs final citation policy","Data-residency gate before geo-shift","Scale-up review"],"acceptanceCriteria":["≥target compliance on agreed event set","Documented 40s-class response if claimed","No silent priority breaches"],"risks":["Methods not reproducible on site stack","Legal limits on load shedding","Author bias"]}},"retrievedAt":"2026-09-21T03:16:40Z","enrichedAt":"2026-09-21T03:24:09Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Flex events may delay batch research jobs; communicate schedules to academic users.","procurementImplications":"Supports utility negotiation packages; does not replace interconnection studies.","operatingModelImplications":"Treats AI factories as dispatchable grid resources with shared runbooks.","sourceVerification":{"openedUrl":"https://arxiv.org/html/2606.25098v1","referenceExcerpt":"Achieved 100% compliance across 200+ distinct National Grid power events on a 130 kW AI cluster","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}