Power Is the New Competitive Advantage
Why AI is already a grid-access race, not a compute race.
Microsoft has already bought some of the chips it cannot yet use. Satya Nadella said publicly that GPUs are sitting idle in inventory because Azure cannot get electricity online fast enough. Investors keep calling this an AI capex race. It is already a grid-access race.
Big tech is on track to spend north of $650 billion on AI infrastructure in 2026. The market is pricing the spend. It is not yet pricing the constraint that decides whether that spend becomes product.
The misread
The public story is straightforward: AI is a compute race, decided by silicon access and capital scale. The evidence appears to support it. Nvidia is the story, allocations are gated, capex is the number everyone quotes.
All true. All beside the point. Compute is an input the hyperscalers can buy. Power is an input they have to negotiate for over years, through utilities, state commissions, federal regulators, and equipment manufacturers whose lead times are quarters at best and years more often. Compute can be bought. Grid access has to be secured.
Why this is a moat
A gigawatt of delivered power to a specific data center in 2028 is the output of a grid-interconnection queue measured in years, a transformer order with an 18-to-30-month lead time, a generation source that takes three to ten years to site, and a substation permit mediated by a state regulator. Each step moves at the speed of the slowest regulated actor. More capital does not make any of them move faster.
That is what makes it a moat rather than a sector-wide expense. A hyperscaler that started securing power in 2023 is two or three years ahead of an entrant starting today, and no amount of money buys forward position in the queue.
The proof: Microsoft and Three Mile Island
The clearest live case is Microsoft’s twenty-year PPA with Constellation Energy to restart Three Mile Island Unit 1, roughly 835 MW, renamed Crane Clean Energy Center, targeting return to service in 2028.
Look at what the deal actually is. It is not Microsoft buying electricity. It is Microsoft underwriting the economic risk of a reactor restart. Constellation gets a 20-year revenue floor that makes a billion-plus-dollar refurbishment bankable. Microsoft gets carbon-free power for the useful life of the data centers being designed around it. The arrangement required a willing nuclear operator, a state PUC that would entertain it, an NRC pathway, and twenty years of offtake Microsoft was willing to put on its balance sheet.
The asset Microsoft acquired is not electrons. It is certainty. Twenty years of priced, scheduled, regulated delivery into a specific grid node. A new entrant with $50 billion in 2026 cannot acquire that certainty in 2028. The restart is under way. It does not come back faster because more money shows up.
The counterexample: why announced is not secured
Amazon’s co-location arrangement at Susquehanna looked, on paper, like a cleaner version of the Microsoft deal. It ran directly into FERC. The expanded interconnection agreement was rejected. Grid-reliability concerns and cost-shifting questions became the central issue. The asset was there. The demand was there. The willing counterparty was there. The regulatory framework wasn’t.
Announced isn’t the same as secured. Google’s SMR agreement with Kairos and Meta’s multi-site nuclear RFP sit in the same category: publicly committed, not yet dispatching electrons. Secured power is a higher bar than committed capital, and the market has not yet learned to tell them apart.
The concession
The strongest objection is not trivial. It has three parts.
First, inference efficiency. If performance per watt improves faster than demand grows, the constraint softens. I do not think that happens on this cycle. But the possibility is real.
Second, regulation. FERC, state PUCs, and Congress can shift the cost of data-center power onto the data centers themselves: contribution fees, special rate classes, mandatory on-site generation. Any of those compresses the upside.
Third, and this matters most: the hyperscalers may win operationally and still not be the best equity expression of the thesis. If all four secure power, they spend comparable amounts doing it. The margin flows to the enablers: merchant generators, regulated utilities in the right territories, transformer and switchgear suppliers. The hyperscalers can be protagonists in the story without being winners in the market.
The predictive claim
Over the next 24 to 36 months, equity markets will learn to separate AI-infrastructure spend that has secured power from AI-infrastructure spend that has not. Two hyperscalers will announce otherwise identical clusters, and the market will price one as a cash-generating asset and the other as a stranded commitment. That discrimination isn’t happening yet. It will.
The repricing does not stop with the hyperscalers. Merchant generators and regulated utilities in the right territories begin to trade less like yield vehicles and more like strategic infrastructure. After that, the bottleneck moves again. Not to more generation, but to the hardware that lets generation matter: transformers, switchgear, substations. That layer gets priced last.
* * *
In household finance, the topline hides the constraint. In AI, capex hides electricity.
Related reading: The Regulatory Moat and Leverage Advantage.
If this essay landed, two next steps.
Run the same process on your business. The Structural Audit is a diagnostic of the structure underneath your revenue: personnel, financial systems, software stack, AI readiness, and operating cadence.
Find your tightest constraint in four minutes. The Structural Advantage Diagnostic is 18 questions across the seven pillars: income, capital, time, health, network, geography. No email required. It returns your weakest pillar and what to do about it.
What is the bottleneck in your business or household that money alone cannot solve faster? Hit reply or leave a comment. I read every one.