AI Infrastructure and the Return of Industrial Economics

The comparison with oil is useful up to a point. Both systems require physical infrastructure before customers receive the output. Both can reward control over an input that is difficult to deliver quickly. Both invite large investments when prices signal scarcity.

The analogy becomes misleading when it suggests that compute is a uniform commodity or that every data center is a durable toll road. A GPU-hour describes access to equipment; it does not guarantee the same completed work across chips, networks and software. Chips improve, workloads change and customers acquire alternatives. I use the industrial analogy to understand the capital required, then assess the actual business on its own terms.

The physical layer matters

Delivering compute requires power, cooling, buildings, networking and operations to function together. The constraint is often in the connection between those layers. Available chips are not productive capacity if the building cannot supply the power or remove the heat they require.

The IEA’s 2025 analysis of electricity supply for AI highlights the role of generation and grid infrastructure. That supports a physical view of the industry, but it does not establish that every announced project will be needed or delivered. Location, timing and workload requirements still determine the useful supply. IEA: energy supply for AI.

There are also several businesses inside the broad label “AI infrastructure.” A colocation provider supplies space, power and supporting services. A compute provider supplies usable computing capacity. A cloud platform may add managed software, storage and other services. An owner operating all those layers takes on different risks from a landlord leasing a facility to a tenant.

I would identify the product being sold before comparing margins or valuations. Is the customer paying for a building, a reserved electrical load, a hardware allocation or a managed outcome? Who owns the equipment, funds its replacement and absorbs idle time? A shared exposure to AI demand does not make the business models interchangeable.

Different assets deserve different assumptions

Land, grid access, buildings, electrical systems and accelerators have different lives. A suitable site may support several hardware generations. Its reuse can still require investment if future racks need more power, different cooling or altered floor layouts. “Long-lived” should describe a plausible reuse case, not an exemption from future spending.

Equipment has both a physical life and a competitive life. An older accelerator may continue working and earn revenue on less demanding workloads. Its rental price or resale value may fall if newer equipment delivers more useful work per dollar or per watt. The decline need not occur at the same speed for every fleet, and a new product launch does not make all existing equipment worthless.

Accounting depreciation spreads an asset’s recorded cost over an estimated useful life. It is not a direct observation of resale value or a promise that pricing will hold. Microsoft’s fiscal 2025 report, for example, assigns different useful-life ranges to computer equipment and buildings. Those are dated accounting estimates for that company, not a universal forecast for an AI fleet. Microsoft 2025 Annual Report, property and equipment policy.

For underwriting, I would model the installed equipment by generation: acquisition cost, expected use, customer pricing, operating cost, refresh spending and residual value. I would also test a shorter competitive life. Extending an accounting life can reduce the annual depreciation expense while leaving the original cash outlay unchanged; it cannot independently improve the economics of the hardware.

Backlog needs to become cash

A large contract is evidence of a commercial commitment, but its value depends on what has actually been promised. A take-or-pay arrangement requires payment for agreed capacity even if the customer does not fully use it, subject to the contract’s terms. A usage-based arrangement leaves more demand risk with the provider. Availability obligations, acceptance tests and termination rights matter in both.

CoreWeave’s March 2025 registration statement provides a concrete, dated example. It describes committed capacity contracts, customer prepayments and billing that can begin on a fixed date or when specified capacity is delivered. The lesson is the sequence: signing, financing and installation are distinct from providing the contracted service. Its terms should not be assumed for every provider. CoreWeave March 2025 S-1, business model.

Cash and revenue can arrive at different times. A prepayment helps fund construction but generally creates an obligation to provide future service; it is not automatically earned revenue. Conversely, revenue can be recognized before the related cash is collected. Remaining performance obligations describe revenue associated with contracted work still to be delivered under the company’s accounting policy. They are not cash in the bank or guaranteed profit. CoreWeave’s filing explains these distinctions and adjustments for potential service credits and delivery delays.

The counterparty matters as much as the headline amount. I would examine its ability to pay, any guarantees or collateral, and the provider’s concentration in that customer. A contract that helps finance a project can still leave significant risk if the customer depends on another financing round or disputes whether service requirements have been met.

I would trace each major commitment from required upfront capital to delivery, invoicing and collection. That makes visible the period in which the provider has committed money but has not yet earned recurring cash. It also shows whether customer prepayments fund a healthy deployment cycle or merely shift cash from later years into the present.

Scarcity is an input to the model

The cash bridge should continue beyond revenue. Electricity, staffing, networking, maintenance, rent, equipment replacement and financing all have claims on the receipts. EBITDA excludes interest, taxes, depreciation and amortization; it does not tell us how much cash remains after maintaining a competitive fleet.

Hypothetical example: a provider collects $40 million in a year, pays $15 million of cash operating costs including rent, reinvests $12 million in equipment and pays $3 million of interest. The remaining $10 million is before taxes and debt principal repayments. That arithmetic is an illustrative cash bridge, not a valuation or an estimate for a real operator. A year with heavier replacement spending could look very different.

Leasing changes the timing and structure of obligations rather than eliminating them. Microsoft’s fiscal 2025 lease note separately reports operating and finance lease cash flows and obligations. I would reconcile those with cash capital expenditure and future commitments before comparing two companies’ spending. Otherwise, one operator can look less capital intensive simply because equivalent obligations appear elsewhere in its accounts. Microsoft 2025 Annual Report, Note 13.

Scarcity can support pricing while demand exceeds deliverable supply. It also encourages customers and competitors to solve the shortage. The thesis is stronger when a company can deliver reliably, finance commitments prudently and retain customers through a hardware transition. It weakens when the return depends mainly on today’s capacity price lasting indefinitely.

The industrial analogy explains the capital required. The contracts and asset lives determine who earns a return on it.