NASA и ERCOT требуют доказательств спроса перед инвестициями в инфраструктуру для ИИ

How NASA’s Artemis Lessons Apply to AI Infrastructure Planning NASA’s Artemis audit and recent ERCOT planning changes point to a new discipline for AI infrastructure: proving demand before billions of dollars are committed. NASA’s Office of Inspector General recently concluded that the agency likely avoided billions of dollars in additional spending by terminating or repurposing four Artemis campaign systems. The audit underscores a question now confronting utilities, regulators, and AI data center developers: How much evidence should exist before forecasts become billion-dollar infrastructure commitments? Earlier this year, Data Center Knowledge reported that ERCOT questioned whether every announced AI data center and large industrial project would become operating demand. FERC has since directed six regional transmission organizations and independent system operators to explain or revise how they evaluate large data center loads, allocate infrastructure costs, and protect existing customers. For data center developers, the answer carries real consequences. The standards that grid operators use to evaluate proposed demand influence how quickly projects move through interconnection studies, when transmission upgrades are planned, who pays for new infrastructure, and how much financial evidence developers may ultimately need before utilities build around their projects. Recent ERCOT planning documents suggest one answer is beginning to emerge. Rather than treating every proposed megawatt as a planning assumption, ERCOT told Texas regulators it has “concerns with using the preliminary load forecast values” for reliability and transmission planning. Instead, the grid operator said planners should incorporate historical realization rates and “other objective, credible, independent information.” Transmission providers project Texas summer peak demand reaching 220.9 GW by 2032, but ERCOT projects 145.6 GW. The roughly 75 GW difference reflects the grid operator’s conclusion that not every proposed large-load project will become operating demand on the timeline developers anticipate. Recent changes to ERCOT’s large-load process reinforce that approach by requiring additional documentation, project milestones, and supporting information before developers advance through portions of the interconnection process. When Assumptions Harden NASA’s Artemis program – an ambitious campaign to return humans to the Moon – shows how quickly early assumptions can become expensive technology commitments. The inspector general found Boeing’s Exploration Upper Stage grew from a $962 million contract to an estimated $3.7 billion before NASA ended the program. Bechtel’s Mobile Launcher 2 expanded from $383 million to roughly $2 billion before the agency redirected its plans. Neil Osnato, founder of Persistence Analytics Group, told Data Center Knowledge the parallels between NASA’s Artemis audit and ERCOT’s recent planning changes became clear after he read the inspector general’s report. “Submitted load is not the same as realizable load,” he said. “Forecast demand is not the same as durable demand. Queue volume is not the same as site, financing, equipment, water, construction, and customer readiness.” Osnato said the NASA audit and ERCOT’s recent planning changes ultimately point to the same question: “What evidence should exist before projected demand or mission assumptions become long-term infrastructure commitments?” “The issue is not whether the mission is important. The issue is whether the assumptions behind hardware, contractors, schedule, architecture, cost, and mission dependency were verified early enough before billions of dollars hardened around them.” A year before the inspector general’s review, the Government Accountability Office reported that Artemis accounted for nearly half of the historical cost growth across NASA’s major projects, underscoring how early assumptions can influence long-term capital programs. Shared vs. Dedicated Infrastructure Ihab Osman, whose experience spans airport, transportation, and digital infrastructure, told Data Center Knowledge utilities can determine whether projects are progressing toward construction without being able to validate the AI business assumptions behind them. “They can verify site control, financing, permitting, interconnection progress, and construction schedules,” Osman said. “Those indicators do not prove the underlying AI demand. They show whether the project is executable, not whether the customer’s technology forecast is correct.” Osman said planners should distinguish between infrastructure that benefits an entire region and infrastructure built primarily for a single customer. A regional transmission upgrade may strengthen reliability and support multiple customers regardless of whether an individual AI campus proceeds, while a dedicated substation or radial transmission line built primarily for one customer carries a different financial risk if that project never materializes. “Plausible demand must become a decision-grade investment case before irreversible capital is committed,” Osman added. Prithpal Khajuria, segment leader for energy and utilities at Intel, said the industry may need stronger financial commitments before large-load requests begin driving transmission expansion plans. “ERCOT alone is tracking large-load requests that are several times larger than its current peak demand, highlighting the scale of speculative activity in the queue,” Khajuria said. He said escrow-backed financial guarantees or similar milestone-based commitments could distinguish projects with financing, site control, and realistic deployment schedules from speculative proposals while reducing the risk that customers pay for infrastructure built around projects that never materialize. “The guiding principle is simple,” Khajuria said. “The party creating the infrastructure need should bear the associated financial risk.” Osnato said the lesson for grid planners is straightforward. “If you want utilities and planners to treat your project as credible long-term demand, bring evidence before asking the system to build around you.”