由 admin 
The AI infrastructure investment debate is becoming a grid-timing debate. Electricity supply, regulatory decisions and connection schedules can dictate whether a GPU-heavy project earns revenue on time. Using the IEA’s 2025 energy outlook and Reuters’ October 2026 report on PJM, this original finance analysis outlines how to underwrite data-center power risk without confusing projections, proposals and approved rules.
Electricity access is a financial constraint, not just an engineering input
AI infrastructure forecasts often begin with GPUs, capital costs and rental rates. That model can fail when a fully funded data-center campus cannot obtain power on the schedule assumed in the loan documents. Connection queues, grid upgrades, backup generation and utility arrangements influence when revenue-producing compute capacity is actually delivered. Investors should model energy availability alongside hardware delivery instead of assuming that installed accelerators automatically become billable capacity.
The International Energy Agency’s 2025 Energy and AI report estimates that electricity generation serving data centers could rise from 460 terawatt-hours in 2024 to more than 1,000 terawatt-hours by 2030 in its base case. That is a global scenario, not a forecast for any single project or company. Its financial relevance lies in the competition for grid access and the uncertainty surrounding when local capacity upgrades will arrive. Power access deserves its own probability-weighted launch schedule.
Recent PJM decisions show why policy cannot be modeled as settled
Reuters reported on October 12, 2026 that US regulators allowed PJM Interconnection to develop a registry of data centers but rejected core elements of a temporary power-curtailment proposal. The registry would gather information such as in-service dates, power ramp schedules and fuel sources. Reuters also reported PJM’s forecast that data-center capacity demand in the region could triple to approximately 29 gigawatts by 2030. Those statements reflect the cited report and forecast, not guaranteed realized outcomes.
The decision illustrates a project-finance dilemma. A grid operator can identify future load without having settled all the rules for prioritizing or funding the infrastructure required to serve it. A developer that treats one proposed policy as certain may misprice contractual obligations. Finance teams should maintain separate scenarios for approved rules, disputed proposals and potential revisions, with an explicit date when assumptions were last checked.
Build a three-clock model for each compute campus
A robust forecast separates the hardware clock, the building clock and the power clock. The hardware clock covers accelerator deposits, shipping and deployment. The building clock covers permits, construction, cooling and inspection. The power clock covers interconnection, transmission work, testing and contracted supply. Cash expenses can begin on all three clocks long before customers are able to use the GPUs, so treating these events as one launch date distorts the runway picture.
Model each as an independent milestone with a most likely and downside date. When one clock slips, recalculate idle hardware depreciation, additional interest, leasing costs and the effect on customer commitments. Consider a hypothetical project that has purchased equipment but cannot energize it for six additional months: the result is not simply six months of lost sales. It can also involve storage, financing expense, warranties, refresh risk and changed chip pricing. Stress-test the cash-flow effects rather than relying on a single delay percentage.
Evaluate power contracts alongside customer concentration
Data-center operators may sign long-term power arrangements while selling compute to a smaller number of AI customers. These tenors do not necessarily match. A fixed volume of electricity, minimum purchase obligation or construction contribution may remain payable even if an anchor tenant reduces its consumption. Finance models should explicitly identify take-or-pay commitments, indexation, transmission pass-throughs, curtailment rights and termination remedies in the actual contracts.
On the demand side, an apparent high occupancy rate can hide customer concentration risk. If a major tenant has optional expansion phases while the data center has already committed to building permanent power capacity, the landlord carries asymmetrical downside. Link customer deposit schedules to power milestones where feasible. Separate contracted recurring revenue from projected utilization, and do not recognize a forecasted expansion as an executed order. Underwriting should describe which parties bear each delay and energy-price shock.
Stress-test debt service with a grid-delay scenario
A useful credit committee scenario shifts the grid connection date later while holding debt draws and hardware purchase commitments constant. The resulting model should show monthly unrestricted cash, interest coverage, committed capex, collateral values and refinancing timing. Another scenario adds a higher power-price assumption and lower customer utilization. A combined scenario asks what happens when ramp-up is late at the same time that the cost of delivered power rises.
Avoid false precision in headline ratios. A debt-service coverage ratio based on stabilized annual earnings may look healthy even though a project cannot meet interest payments during the commissioning period. Include debt service during construction and liquidity available before steady-state operations. Lenders should review the rights of the power provider to interrupt service and whether the borrower has the operational flexibility to move workloads to another site. These variables are project-specific, not solved by a national energy forecast.
What diligence should demand before approving funding
Request the latest interconnection correspondence, utility study milestones, engineering procurement schedules, power purchase terms, backup generation assumptions and evidence that projected cooling capacity can operate at the expected load. Cross-check every delivery date against third-party documents rather than marketing slides. Ask management to explain who pays for grid upgrades and who can terminate a power agreement after a schedule failure. Model regulatory proposals only as scenarios until formally adopted.
Finally, reconcile site-by-site power constraints with the consolidated investor narrative. An aggregate megawatt target can mix operating sites, facilities under construction and projects still seeking interconnection. Present those categories separately, with the corresponding customer contract status. The opportunity in AI infrastructure may be substantial, but disciplined underwriting prices the time and cost of making compute truly available—not only the price of GPUs on an invoice.
The most valuable new spreadsheet line may be the date a data center can reliably draw its contracted power. Without that line, even a sophisticated GPU valuation model can overstate the speed of cash generation.
Sources
- IEA, Energy and AI: Energy supply for AI (2025)
- Reuters, US panel approves PJM data-center registry (October 12, 2026)
Cover image: Unsplash photo-1702446736200-6b9a345dea6f. This report provides financial analysis, not an investment recommendation.