Capability&Consequence

Essay

The Building Built for a Three-Year Machine

Long-term data centre commitments depend on short hardware cycles. Each party needs to price the exposure it actually holds.

Document
CAC-007
Issue
1.0
Published
July 30, 2026
Reading time
10 minutes
Review due
July 29, 2027

TL;DR

Buildings, leases, power agreements and debt have much longer lives than the compute supporting their economics. Operator useful-life estimates reveal material disagreement about the refresh cycle. Each party should state its implied demand duration, identify hardware transitions within the commitment and assign retrofit and residual-value exposure to a named owner.

In The Three Clocks Behind the AI Business Case (CAC-006), I argued that AI investment cases collapse three timelines into one. The capability clock governs when the technology changes. The operating clock governs when organisations absorb it. The economic clock governs when value is realised.

Within an enterprise, the mismatch can lead to a business case that overstates how quickly a technical improvement becomes a financial return. The data centre buildout presents a larger version of the same issue: buildings, power agreements and debt commitments measured in decades depend on demand from compute whose competitive life follows a chip cycle measured in a year or two.

That difference in duration can be managed, just as financial institutions manage mismatches between their assets and liabilities. But management requires someone to identify the exposure, price it and accept responsibility for it. Borrowing an assumption from a party with different incentives does not establish that the risk has been assessed.

For boards, investors and infrastructure leaders, the practical question is therefore: at each layer of the capital stack, who bears the consequences of hardware refresh, and what evidence shows that they have priced those consequences? The answer matters independently of whether aggregate AI demand continues to grow.

The clock itself is accelerating

The hardware roadmap creates two important pressures. First, GPU platforms at the centre of the buildout now turn over on roughly an annual rhythm, changing the economics against which an existing investment competes. Each generation resets the economics of the one before it. The old hardware still works. What changes is how much a unit of capital deployed on it can earn against the new alternative.

Second, successive platforms impose different physical requirements, which means the next generation may require more than a component replacement. An industry-average rack draws under ten kilowatts. Current rack-scale flagship systems draw about 120 to 130 kilowatts. They require direct liquid cooling as a condition of deployment, not an option.

The successor platform announced in March 2026 is specified at roughly 190 to 230 kilowatts per rack. It is entirely liquid-cooled and uses a new power-delivery architecture. Each step is more than a faster component in the same slot. It imposes different demands on the building's power distribution, cooling plant and floor loading.

This is the pattern from The Model Will Be Gone Before the Product Is (CAC-004): the interface refuses to stay fixed. For a data centre, that interface is the building's electrical and thermal envelope. Its fitness must be assessed against the hardware it will need to accommodate during its economic life.

A space engineered around one generation's densities is not neutral toward the next. Can it host the successor economically? The answer has a capital cost, and that cost lands on whoever holds the shell when the boundary arrives.

Walk the stack by who controls the clock

The hyperscaler building for itself holds both sides of the mismatch. It sets its refresh cadence and designs the next space around a roadmap it can see. It can absorb transition costs inside businesses large enough to hide them.

The four largest hyperscalers reported hundreds of billions of dollars in spending, with further commitments planned. Their visibility into workloads and hardware requirements gives them a stronger basis for managing refresh than parties whose exposure depends on a tenant's decisions. That advantage does not eliminate the mismatch, but it helps explain why the same assumption may carry a different risk for an operator and a lessor.

The lessor holds the shell but not the clock. A developer or REIT signing a build-to-suit lease with a hyperscaler holds an asset with a building's useful life. Its tenant's demand rides the chip cadence.

During the lease, the tenant's credit carries the structure. The exposed question sits at the boundary. When the lease ends, the shell's value depends on what the next tenant's hardware requires. Will its power density, cooling architecture and grid position still fit? Current generation-to-generation transitions suggest that nobody can assume the answer.

A long lease can protect contracted cash flow while leaving residual-value exposure at expiry. Underwriting a twenty-five-year shell against a fifteen-year lease therefore includes a judgement about its economic fitness after the tenant's commitment ends. The model should state that judgement explicitly.

The financing layer holds paper against the leases. Data center securitization has moved from a niche to one of the fastest-growing parts of the US structured-credit market.

Issuance surpassed tens of billions of dollars in 2025. Sell-side projections put annual issuance at US$30–40 billion, with some estimates taking outstanding balances toward US$180 billion by 2028. These structures are usually sized against contracted lease cash flows. Five-year weighted average lives often sit inside longer lease terms, keeping the paper itself short.

The clock exposure is second-order but real. Refinancing at each maturity assumes that the collateral's value story survives every refresh boundary along the way. That story includes the lease, the tenant and the shell's continued fitness.

A market growing this quickly is also a market where most of the collateral has never lived through a refresh boundary.

The power layer holds the longest clock of all. Microsoft's agreement with Constellation to restart Three Mile Island Unit 1 is a twenty-year power purchase commitment attached to an 835-megawatt plant.

Utilities are planning generation and transmission against data center load forecasts. Municipalities are granting multi-decade concessions and abatements. These are the least reversible commitments in the stack. They are made against its most revisable element: the hyperscaler's commitment to a particular site, workload and hardware generation.

A twenty-year power position against AI demand includes assumptions about the persistence and location of that demand. Utilities and municipalities should make those assumptions explicit, including what happens if the customer's workload or preferred site changes before the commitment expires.

Useful-life estimates reveal a material disagreement

The most instructive evidence about the refresh clock comes from the clock-controllers' own accounting, because useful-life estimates are where a company writes down, in auditable form, how long it believes its hardware earns. The public record runs as follows.

Microsoft extended the useful life of server and networking equipment from four years to six, effective fiscal 2023. It disclosed an operating-income benefit of roughly US$3.7 billion in the first year. Alphabet made an equivalent move to six years. It disclosed a US$3.9 billion reduction in depreciation and a US$3.0 billion increase in net income for 2023.

Meta reached 5.5 years through successive extensions, the latest effective January 2025. It disclosed a US$2.9 billion depreciation reduction for the year. Oracle moved from four years to five in 2023. Amazon subsequently revised its estimate in the opposite direction, making the uncertainty particularly visible.

Effective January 2025, it shortened the useful life of some servers and networking equipment from six years back to five. It took roughly US$920 million of accelerated depreciation on early retirements. It also guided to about US$0.7 billion lower 2025 operating income.

Amazon attributed the change to the increased pace of technology development, particularly in the area of artificial intelligence and machine learning. Subsequent quarterly filings recorded an ongoing income effect in the hundreds of millions. An extension and a reduction can each be defensible for a particular fleet. What matters here is that the estimates differ materially among companies with unusually strong visibility into the technology.

Companies operating substantially similar fleets in the same technology cycle reached varying conclusions about how long the hardware earns. The difference is worth billions of dollars of reported income per company per year. Short-seller Michael Burry claimed in late 2025 that the group's schedules would understate depreciation by about US$176 billion across 2026–2028. That is one contested estimate of the gap's size. It does not need to be right for the underlying point to stand.

Lessors underwriting residual value, lenders sizing collateral and analysts capitalising earnings should therefore avoid treating an operator's useful-life estimate as a settled industry fact. It is an owned judgement about a particular fleet, subject to revision, with consequences large enough to affect reported income and valuation.

This is the argument of Two Numbers You Will Be Asked For (CAC-004-F) at the scale of the market. A useful-life estimate is a judgement that requires a documented rationale. An estimate extended during the buildout flatters reported income while capital spending compounds. It is a claim that deserves evidence proportionate to what now rests on it.

What each party should model

Each party should document the assumptions within its own commitment and assign the resulting exposure to a named owner. Five considerations belong in that assessment.

State the implied demand duration. Every commitment in the stack implies a period over which AI compute demand must persist in roughly its current physical form. That applies to a shell, lease, tranche, PPA or abatement.

Write the number down. A twenty-year power position implies twenty years. A five-year ABS tranche implies that the lease survives its refinancing dates. Many of these numbers have never been stated because they were never modelled as assumptions.

Locate the refresh boundaries inside the commitment. At the current cadence, a fifteen-year lease spans many hardware generations.

For each boundary, name who pays if the transition changes the building's power distribution, cooling plant or structural loading. Then state what happens to the commitment if the economic answer is a new building rather than a retrofit.

Underwrite the residual, not just the lease. A party holding the asset past its contracted cash flows is betting on the shell's fitness for hardware that does not yet exist.

An honest residual-value range is wide. A narrow one is the anomaly that requires explanation.

Treat borrowed useful lives as borrowed. If a model inherits a depreciation or refresh assumption from an operator's filings, record whose assumption it is. The operators themselves disagree by enough to move valuations.

Amazon's reversal is the documented proof that these estimates revise downward as well as up.

Separate the demand question from the duration question. Demand for AI compute can be everything its proponents claim and the mismatch can still bind.

The question is not only how much demand exists. It is where that demand appears and in what physical form at each boundary. This is why the analysis survives any view on the boom itself.

The practical test

For any long-dated commitment in the data centre stack, one question:

If the hardware generation turns over twice before this commitment matures, who pays for what changes — and did that party price it?

The answer should identify the responsible party, the expected cost and the evidence supporting the estimate. Demand growth alone does not answer a question about retrofit cost, location or residual value. Those assumptions belong in the approval of the commitment, with responsibilities allocated before the hardware transition tests them.

The decision this should change

For each material data centre commitment, document demand duration, refresh boundaries and the party responsible for retrofit or residual-value loss. Support useful-life and replacement assumptions with evidence appropriate to the specific asset and obligation.

What this adds

Prevailing consensus

Data centre infrastructure is being underwritten as long-duration digital real estate: contracted cash flows from investment-grade tenants, secular demand growth, and residual values supported by scarcity of power and sites.

What this challenges

The demand behind those cash flows is generated by hardware on an annual cadence, and successive generations are no longer physically interchangeable. The parties closest to that clock are publicly divided about how long the hardware earns. Their estimates differ by years and move reported income by billions. The layers that do not control the clock have largely inherited their duration assumptions from the layer that does.

New contribution

The essay applies the three-clocks framework (CAC-006) to the buildout's capital stack. It maps clock control across the operator, lessor, lender, power provider and municipality. It then uses the 2022–2025 useful-life record, including Amazon's reversal, as auditable evidence that the refresh clock is a live dispute rather than a settled input.

What would weaken the argument

The thesis would weaken under three conditions. First, successive hardware platforms become deployable inside existing power and cooling envelopes without material retrofit. Second, a meaningful sample of shells passes refresh boundaries and is re-leased at or near prior economics. Third, operator useful-life estimates converge and observed replacement behaviour validates them rather than forcing another revision.

Sources and references

  1. Bloomberg (via Yahoo Finance) — Meta's extension of server useful lives to 5.5 years, prior extensions by Microsoft and Oracle, and 2025 hyperscaler capex of approximately US$300 billion versus US$217 billion in 2024
  2. Deep Quarry — Amazon's 2025 useful-life reduction from six years to five, the US$920 million of accelerated depreciation, and disclosed quarterly income effects from the 10-Q record
  3. Deep Quarry — Depreciation of GPUs: useful-life changes across the group and their treatment under ASC 360
  4. Levelheaded Investing — Michael Burry's estimate of understated depreciation across the hyperscalers, 2026–2028
  5. IntuitionLabs — Physical requirements of current GPU platforms: approximately 120–140 kW liquid-cooled racks versus conventional 10–15 kW enterprise racks
  6. NVIDIA — GB200 and GB300 NVL72 rack-scale liquid-cooled systems
  7. NVIDIA — Vera Rubin platform and the liquid-cooled, DC busbar-powered MGX rack architecture
  8. Schneider Electric — Reference designs for approximately 142 kW-per-rack AI infrastructure
  9. Constellation — 20-year power purchase agreement with Microsoft and the restart of Three Mile Island Unit 1 (835 MW)
  10. L&G Asset Management — Data centre ABS and CMBS issuance surpassing US$25 billion in 2025; typical five-year weighted average lives
  11. CREFC — Data Center e-Primer: approximately US$57 billion of US issuance since 2021 and projected US$30–40 billion annual supply
  12. Structured Finance Association — How data centre ABS and CMBS fit the financing ecosystem; Barclays projection of approximately US$180 billion outstanding by 2028
  13. Capability & Consequence — The Three Clocks Behind the AI Business Case (CAC-006)
  14. Capability & Consequence — The Model Will Be Gone Before the Product Is (CAC-004)
  15. Capability & Consequence — Two Numbers You Will Be Asked For (CAC-004-F)