The argument is about velocity, scope, and optionality, and it ends with a pattern the AI buildout has already repeated three times. Power is simply the last layer waiting its turn.
The velocity problem
Start with the aggregate numbers, because they reframe what the constraint actually is. The four largest hyperscalers are guiding to roughly $725 billion in combined 2026 capital expenditure, nearly doubling a 2025 that was itself a record, with Goldman Sachs projecting on the order of $5 trillion across the group through the end of the decade. Capital intensity has reached 45 to 57 percent of revenue, territory these businesses have never operated in, funded in part by more than $100 billion of debt raised in 2025 alone with projections of well over a trillion in issuance to come. At that spend rate the binding constraint is not the cost of a dollar. It is how many dollars the organization can deploy well simultaneously: how many programs, supply chains, and workstreams the institution can run at quality at once. Call it capital velocity. Every category of spend now competes for organizational throughput, not just budget.
Generation ownership is a heavy consumer of exactly that throughput. Owning power means standing up a second business inside the first: procurement against the multi-year OEM backlogs documented in The Turbine Gap, EPC management, fuel and water contracting, long-term service agreements, spares, operators, compliance. The question was never whether a hyperscaler can do this; obviously it can. The question is whether standing up a power company inside a compute company is the fastest path to compute, when every workstream on that list sits between the capital decision and the first rack.
The market is repricing capital discipline in real time
The equity market has started charging for undisciplined capex. When Meta raised its 2026 guidance toward the $125 to $145 billion range citing component and data center costs, the stock fell more than nine percent in a session, the first genuine investor rebellion against the spending curve, and institutional coverage since has openly questioned whether historically lean businesses are becoming permanently capital-hungry. The mechanism behind the anxiety has a name in every equity research note now circulating: the depreciation cliff. Capex at this scale converts into tens of billions of dollars of incremental annual depreciation landing on income statements from 2027 onward, and the market is pricing which management teams have a plan for that and which are hoping.
In that environment, examine which dollar on the balance sheet is hardest to defend on an earnings call. Not the GPU dollar; that one has a revenue narrative attached. The generation dollar: a 25-year, single-site, single-technology commitment with no compute revenue line of its own, sitting in the depreciation base for decades. Structures that deliver the megawatts while keeping that commitment off the owned-asset line are not accounting cosmetics. They are how a capacity program stays defensible to the people who price your equity. And the debt market has already voted on which side of this it wants to fund: Meta's roughly $30 billion bond offering in late 2025, one of the largest investment-grade corporate deals on record, drew an order book reported north of $100 billion. Capital is not scarce for this buildout. Structure is.
The sequence: every layer above power has already financialized
Here is the pattern, and once you see it the conclusion writes itself. The AI stack has been financializing layer by layer, top down, for three years.
The compute layer went first: chips moved off simple ownership economics into vendor financing, neocloud structures, and capacity contracts, with GPU fleets now routinely financed against offtake rather than held as plain capex, and the largest single buyer of compute, OpenAI, carrying aggregate commitments across its vendor ecosystem reported at a peak of around $1.4 trillion before being reset toward roughly $600 billion by 2030 in early 2026. Read the reset as the proof rather than the retraction: commitments of that scale exist only as structured capacity contracts, which is precisely why they could be resized in a quarter. Owned assets do not resize. The shell layer went next, and publicly: Meta financed its largest campus, the Hyperion project in Louisiana, through a roughly $27 billion joint-venture structure with Blue Owl's private credit platform, keeping the majority of the asset off Meta's consolidated ownership while preserving full operational control. Read that transaction the way debt markets did: a hyperscaler with one of the strongest balance sheets in corporate history chose not to own its own building, because velocity and balance-sheet capacity beat ownership even at their cost of capital.
The precedent behind the precedent is telecom towers. Carriers owned their towers until they realized towers were a separate business consuming capital and management attention that belonged in the network; they sold them to specialists, leased back capacity, and the tower companies became some of the best infrastructure businesses ever built, trading at roughly triple the EBITDA multiples of the carriers that sold them the assets. Same cash flow, triple the valuation, purely from moving the asset into the hands whose entire business it was.
Compute companies will not own generation for the same reason telcos do not own towers.
The multiple arbitrage says the market will pay them to make the same discovery faster. So run the sequence: chips financialized, shells financialized, and the layer below the shell is power. It is the last layer still sitting on the default assumption of ownership, and it is the layer with the longest asset life, the least connection to any competitive differentiation, and the most volatile technology and regulatory environment, which is to say it is the layer least suited to that assumption. The forward call, with high confidence: within roughly 24 months, dedicated generation for AI campuses becomes a recognized asset class held by infrastructure and private credit capital, with hyperscalers as offtakers under capacity and consumption structures. The instruments are already visible at the edges: capacity rights, slot reservations, tolling frameworks. The market will produce the equivalent of reserved instances for megawatts because the buyers who invented reserved instances are now the buyers of megawatts. The only open question is whether your organization is early to that market or provides the case study in why it formed.
The optionality math
A generation asset on your balance sheet is a quarter-century bet on a site, a technology, and a demand forecast, placed in the most volatile infrastructure market in living memory. Reserved capacity is a right, not an obligation. When model architectures shift the density assumptions, when a site's water math breaks, when the grid interconnection finally energizes on the timeline mapped in Bridge to Permanent, the buyer holding capacity rights redirects in a quarter. The buyer holding turbines opens a negotiation with themselves, and that negotiation has a name in every post-mortem: sunk cost.
The tradeoff, stated plainly: at sufficient scale and duration, ownership wins. A permanent flagship campus with a thirty-year horizon and a settled technology stack should internalize its power over time, and Bridge to Permanent maps exactly that lifecycle. Capex-to-opex is not a religion. It is the correct instrument for the phase of the market we are in, where speed and reversibility are worth more than the final increment of levelized cost.
The discipline compresses to one sentence: separate “do we need the power” from “do we need to own the machine.” The first answer is yes everywhere. The second answer is yes far less often than current balance sheets suggest, and the tower industry's entire existence is the proof.