The price of compute capacity is no longer denominated in dollars per megawatt. It is denominated in months-to-energization.
A 200 MW campus deliverable in Q3 2027 is structurally worth more than a 400 MW campus deliverable in Q1 2030, even at materially higher dollars per megawatt. The frontier-lab capex pipeline through 2028 (Microsoft alone is guiding to roughly $80B in AI infrastructure spend in FY26) is sitting on top of a transformer order book that physically cannot be filled inside the training cycles the spend is meant to feed. Date is the binding constraint. Everything else is paperwork.
The market has not repriced for this yet. It will.
The chokepoint is forged steel, not silicon
The story sold to capital markets is GPU scarcity. The actual bottleneck sits one layer underneath: large power transformers (LPTs), specifically the 345/138 kV and 500/230 kV units that step grid voltage down to the medium-voltage bus a hyperscale campus actually drinks from. As of early 2026, quoted lead times from Hitachi Energy, GE Vernova, Siemens Energy, and Hyundai Electric run 120 to 150-plus weeks for new LPT orders. WoodMac and the U.S. Department of Energy have separately flagged that global LPT manufacturing capacity is concentrated in roughly a dozen plants worldwide, and that U.S. domestic LPT production covers less than 20% of domestic demand. Grain-oriented electrical steel, the input that actually constrains the core, has one meaningful North American producer and a handful of Asian and European mills running near full utilization.
Translate the lead time into deal calendar. An LPT ordered in Q2 2026 lands on a U.S. construction site somewhere between Q4 2028 and Q3 2029. If a developer needs a 2027 energization, the transformer had to be on order in 2024. Most of the speculative AI campuses announced in the last 18 months did not place those orders. They cannot. Their interconnection studies are not far enough along to specify the unit. The order book at the manufacturers is therefore a near-perfect proxy for which projects are real and which projects are press releases. The press releases will quietly disappear.
Interconnection is the second forged-steel chokepoint, except the steel here is queue position. PJM's interconnection backlog crossed 250 GW of pending requests in late 2025, with cluster studies running 24 to 36 months even under the reformed first-ready, first-served process. MISO's DPP cohorts are similarly congested, with the West and Central subregions carrying multi-year study queues. ERCOT has no capacity market and processes large load interconnections faster on paper, but West Texas transmission constraints, the ongoing ERCOT Regional Planning Group reviews of multi-hundred-megawatt loads, and Public Utility Commission of Texas attention on large-load tariff design have turned "fast queue" into "fast queue with a political ceiling." CAISO's cluster study process is 24-plus months from filing to executed Generator Interconnection Agreement, and the operating assumption inside the queue management group is that 2027 cluster filings will largely energize after 2030.
ISO-NE and SPP are the quietest part of this story and the most important. Both queues collapsed under load-driven application volume in 2024 and 2025. SPP paused new generation interconnection requests in parts of 2025 to clear its backlog. ISO-NE's 2024 Forward Capacity Auction cleared at levels signaling structural scarcity, and the region's 2026-2031 transmission planning assumptions explicitly call out the inability to absorb large-load interconnections without targeted upgrades that themselves take five to seven years. The secondary queues are not a relief valve. They are a closed door with a smaller sign on it.
Why the consensus misses this
Equity research, real estate underwriting, and most utility-side planning still price hyperscale campuses as a function of acreage, dollars per megawatt of nameplate capacity, and zoning posture. That framework was correct in 2018. It is structurally wrong in 2026.
The framework misses three things. First, nameplate capacity is a fiction until an executed Generator Interconnection Agreement and an LPT delivery schedule are stapled to it. A 1 GW announcement with a queue position in PJM's 2024 cluster is not 1 GW of anything that finances. It is a lottery ticket. Second, dollars per megawatt is the wrong denominator because the marginal buyer is not optimizing cost of capacity. The marginal buyer is optimizing arrival inside a training-cycle window where each quarter of delay surrenders one full generation of model competitiveness. Anthropic, OpenAI, xAI, Meta, and a growing list of sovereign-backed buyers are running 12 to 18 month model release cadences. A campus that lands in the window is worth a campus that does not, regardless of the per-megawatt spread. Third, the consensus continues to treat greenfield development as the default path. Greenfield is the slowest path. It is becoming the wrong path.
The hyperscalers themselves have figured this out. Look at where the capital is actually moving. Microsoft's reactivation discussion with Constellation around Three Mile Island Unit 1 is not a nuclear story. It is a date-certain capacity story attached to an interconnection point that already exists on the books. Amazon's Talen Energy campus at Susquehanna is the same structure. Google's TVA arrangements at retired coal sites carry the same logic. Each of these transactions priced behind-the-meter or co-located capacity at a premium specifically because the alternative, a new greenfield interconnection request, prices in years of optionality decay.
The arbitrage is not subtle anymore. It is being priced into transactions in real time. The market has just not generalized the lesson.
The forward view
Our view at Autonomous Industries runs as follows.
By 2027, "date-certified" capacity will be a distinct asset class with its own bid stack. Capacity inside an executed GIA, with an LPT order placed and a CUP either approved or on a defined approval path, will trade against a published forward curve, in the same way distillate inventory and pipeline capacity trade against forward curves today. Newmark, JLL, and CBRE will publish indices. Insurance markets will write delay coverage. The product will be financeable.
By 2028, date-certified capacity will clear at a 30 to 60% premium per megawatt over queued-but-uncertain capacity of equal nameplate. The premium is not speculative. It is the present value of the foregone training cycle a buyer protects by paying it. A 200 MW campus delivering revenue or strategic value in 2028 versus 2030 is worth, at frontier-lab compute economics, somewhere between $400M and $900M in net present value differential. The premium is rational and will widen, not compress, as model cadences accelerate.
By 2029, capacity interconnection rights inherited from retired coal plants will be the single most contested asset in U.S. power markets. CIR transfer mechanics vary by ISO, but the principle is consistent: when a thermal generator retires, the interconnection capacity it occupied does not always dissolve back into the queue. In PJM, MISO, and parts of SPP, CIRs attached to a retiring unit can be transferred or repurposed under defined rules to a new generator or, in increasingly creative structures, to a co-located load. The coal retirement schedule through 2030 implies roughly 80 to 100 GW of U.S. nameplate capacity coming offline. The CIRs attached to those retirements, at sites already permitted for industrial use, with switchyards already energized and transmission already studied, are the closest thing in the American grid to "instant" capacity. Owners of those sites will not be selling fee simple. They will be selling option premia.
By 2030, the greenfield development model collapses into a niche. The economic logic that justified speculative land assembly, queue filing, and multi-year permitting will not survive a market in which CIR-anchored brownfield, behind-the-meter gas adjacency, and pre-permitted parcels carry a structural date advantage of 24 to 48 months. New greenfield will continue to get built, but only where it can be packaged with date certainty at origination, which means only where the developer either owns the LPT order book exposure or has secured it contractually before site work begins.
By 2032, sovereign and frontier-lab buyers will be transacting directly with date-controlling parties, bypassing traditional developer intermediation. The disintermediation is already starting. The "developer" function survives only where it owns a piece of the date chain, transformer slots, CIR positions, executed GIAs, pre-permitted parcels, behind-the-meter generation, that the buyer cannot replicate on its own timeline.
What compounds from here
Three second-order effects compound off the date-certainty repricing.
Insurance markets reprice delay. Liquidated damages and delay-in-startup coverage on AI infrastructure projects are currently underwritten on commercial real estate analogs that no longer apply. As frontier-lab buyers begin attaching nine-figure liquidated damages to commissioning dates, the insurance product itself becomes a constraint on which projects can even be offered. Few markets will write meaningful capacity against a date the underwriter cannot diligence to the LPT serial number. This narrows the field of credible developers further.
Sovereign capital enters the bid stack. UAE, Saudi, Singaporean, and Norwegian capital have already begun deploying into U.S. and European AI infrastructure with explicit date-priority mandates. The G42 and Stargate-adjacent vehicles are not optimizing cost of capacity. They are optimizing strategic arrival inside a national AI posture window. A bidder optimizing for date, not price, sets the marginal clearing price for date-certified product. Domestic institutional capital that continues to underwrite on traditional cap-rate frameworks will be repeatedly outbid for the assets that actually matter.
Utility commissions begin to ration. The political backlash already visible in Virginia, Ohio, Georgia, and Texas around large-load tariffs, cost allocation, and ratepayer impact will harden into formal rationing regimes by 2028. Capacity will be allocated, not auctioned, in at least three U.S. jurisdictions before the decade closes. The developers who hold pre-rationing interconnection rights will hold the equivalent of grandfathered spectrum.
The compounding effect is straightforward. Each layer of the stack, transformers, CIRs, GIAs, permits, insurance, capital, regulation, ratchets in the same direction: scarcity of date, not scarcity of megawatts. A market that priced megawatts will spend the next five years repricing dates.
The infrastructure industry sells nameplate. The buyers are paying for arrival. The spread between those two truths is the entire investable opportunity of this cycle.
Published by Autonomous Industries