# The AI Data Centre Is a Timed Infrastructure Contract: Utilisation, Power and the Price of Delay

> An evidence-led October 2026 analysis connects new power contracts and hyperscaler cash flows to a reproducible model of utilisation, electricity cost, commissioning delay and social licence.

- **Category:** Energy Markets & Procurement Economics
- **Author:** Voltformer Research Desk
- **Publication date:** 2026-10-22
- **Reading time:** 19 min read
- **Key topics:** AI Data Centres, Electricity Contracts, Capacity Utilisation, Infrastructure Economics, Capital Expenditure
- **Body language:** en
- **URL:** https://voltformer.com/articles/data-center-economics-power-utilisation-delay-october-2026

### 1. Thesis: a data centre sells time, not a connection announcement

A large AI campus is often described with one spectacular number: a capital budget, a gigawatt commitment or an accelerator count. None establishes that a customer can obtain a useful computation at a competitive cost on the required date. The scarce product is a chain of simultaneous capabilities: an authorised site, deliverable electricity, commissioned cooling and power distribution, functioning servers, a usable network, suitable software and paying demand. The weakest link determines the date at which the other investments begin to earn their keep.

Our argument is that the October 2026 investment debate should move from announced scale to **productive capacity over time**. This is more demanding than asking whether AI demand is real. Demand can be strong while an individual campus is uneconomic; a power project can be valuable while the customer computing platform disappoints; consolidated cash flow can remain healthy while a particular generation of equipment earns little. These propositions are compatible. Treating them as opposing positions obscures the actual underwriting problem.

This article evaluates primary disclosures available on 22 October 2026. Company announcements establish what an issuer says or has contracted, not independently verified operating performance. Our numerical examples are explicitly assumed analytical models, not prices offered by a vendor, measured campus performance or forecasts of investment returns. They are designed so that a reader can replace every input and reproduce every result.

### 2. What changed in October—and what did not

The most informative recent announcements concern the allocation of cost and timing risk. AEP's 14 October announcement describes a planned 702 MW fuel-cell generation project, approximately USD 4 billion of investment and a 20-year Google power agreement, with full operation expected by the end of 2028. It is a future generation project, not electricity already delivered to an operating 702 MW IT load. [AEP announcement](https://www.aep.com/news/stories/view/12314/).

Black Hills' 6 October disclosure separates 564 MW of planned company-owned generation from up to 590 MW of grid-connected service and approximately 2.1 GW of managed third-party resources. Its 2.7 GW resource description includes reserve margins. Those quantities must not be added again or relabelled as IT demand. Its announcement also describes cost recovery, early termination and credit protections; the complete contractual and regulatory position still requires the actual agreements and approvals. [Black Hills disclosure](https://ir.blackhillscorp.com/news-releases/news-release-details/black-hills-corp-executes-definitive-agreements-serve-googles).

Oracle's 13 October New Mexico announcement is explicitly a proposal. It describes annual minimum power spending of USD 10 million for ten years and a separate proposal for as much as USD 55 million of annual customer credits for up to three years. Construction, commissioning, emergency and life-safety service are distinguished from the site's principal fuel-cell microgrid. A proposal and its headline do not establish approved tariffs, actual credits paid, or a universal total project subsidy. [Oracle proposal](https://www.oracle.com/news/announcement/oracle-proposes-plan-to-offset-customers-utility-costs-2026-10-13/).

| Primary evidence | What is established | What remains unproven |
|---|---|---|
| AEP, 14 October | Planned generation investment and contracted PPA framework | Completion date, realised output and final delivered cost |
| Black Hills, 6 October | Distinct service, generation and resource-management boundaries | Full future operating performance and counterparty survival |
| Oracle, 13 October | Proposed minimum spending and customer-credit structure | Final approval, actual payments and complete system-cost allocation |
| Meta, 21 October | Reported consolidated quarterly cash flows and updated capex guidance | A standalone AI campus's cash return |

The common feature is an attempt to make the large load financially accountable for its supporting infrastructure. That can improve incentives. It does not eliminate construction delays, common-mode failures, environmental constraints or changes in the demand for compute. A contract moves an exposure to a counterparty; it does not remove the underlying physical event.

### 3. Three boundaries that every investment memo needs

The first boundary is physical. A 100 MW IT design capacity, a 120 MW facility import capability and 100 MW of renewable generation are different quantities. The first concerns computing equipment; the second includes supporting systems at a stated operating point; the third is a generation rating whose annual output and availability depend on technology and location. A contractual energy volume is different again. Conversion between these boundaries requires explicit assumptions rather than a convenient label.

The second boundary is temporal. Land, buildings, substations, servers and models have different lead times and useful lives. A long power commitment can outlive several computing generations. A server delivery can precede the energisation date. A model's commercial advantage can disappear before its hardware is fully utilised. There is no single economically meaningful life for “the data centre”. The investment should be decomposed into replaceable cohorts and long-lived enabling assets.

The third boundary is contractual. The campus owner, electricity supplier, equipment financier, cloud platform and end customer may be separate entities with different cancellation rights. A signed reservation is not necessarily a cash receipt; an order backlog is not necessarily a take-or-pay revenue stream; a minimum power payment does not automatically cover all network reinforcement. The cash-flow model must identify who pays, under what conditions, when collateral is available and what happens after default.

### 4. Cash spending, lease additions and depreciation tell different stories

Meta's 21 October release reports, for the quarter ended 30 September, operating cash flow of USD 32,786 million, property-and-equipment purchases of USD 33,942 million and finance-lease principal payments of USD 964 million. The resulting reported free cash flow is negative USD 2,120 million. Its full-year capex guidance, including finance-lease principal payments, is USD 135–145 billion. These are consolidated figures and guidance, respectively. [Meta Q3 results](https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-Third-Quarter-2026-Results/default.aspx).

| Meta quarterly cash bridge, USD million | Amount | Interpretation |
|---|---:|---|
| Operating cash flow | 32,786 | Consolidated cash generation |
| Property-and-equipment purchases | −33,942 | Cash capital outflow |
| Finance-lease principal payments | −964 | Additional financing-related capital outflow in this definition |
| Free cash flow | −2,120 | 32,786 − 33,942 − 964 |

Negative quarterly free cash flow is a financing fact, not by itself proof that AI investments destroy value. Conversely, growing revenue is not proof that a new campus covers its cost of capital. A defensible assessment needs incremental cash generation attributable to the investment, replacement requirements, timing and risk. It must avoid assigning all existing advertising or software cash flow to the newest infrastructure and then declaring the infrastructure profitable.

Microsoft's July fiscal-fourth-quarter call illustrates the presentation problem. It distinguishes USD 41 billion of capex, USD 5.6 billion of finance-lease additions and USD 35.8 billion of cash PP&E payments. It also describes extending estimated useful lives of data centres and office buildings from 15 to 25 years, affecting depreciation and future lease classification. That is not a statement that GPUs now last 25 years. [Microsoft earnings call](https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q4).

Comparisons should reconcile lease additions with principal repayments, fiscal with calendar periods, and group totals with project boundaries. An accounting classification can change reported capex while leaving physical resource commitments substantially unchanged. Depreciation is an allocation rule; it does not refill a depleted cash account or demonstrate that customers will continue to pay for an older accelerator generation.

### 5. A reproducible campus model: assumptions before conclusions

Consider a synthetic campus with 100,000 identical compute units and 100 MW peak IT power. A compute unit is a normalised accounting unit for this model, not a named GPU, a token or a benchmark result. Its service quality and throughput are held constant across cases. The model therefore cannot compare chip vendors or estimate a language model's cost per token.

| Assumed input | Value | Boundary |
|---|---:|---|
| Installed normalised compute units | 100,000 | Constant across cases |
| Peak IT power | 100 MW | Not facility import power |
| Idle IT power fraction | 25% of peak | Assumed linear active/idle approximation |
| Facility PUE | 1.20 | Constant analytical approximation |
| Delivered energy price | 70 USD/MWh | Assumption; demand charges excluded |
| Computing investment and recovery life | USD 1.2 billion; 5 years | No residual value |
| Facility investment and recovery life | USD 400 million; 15 years | Separate from compute equipment |
| Capital recovery rate | 8% annually | Assumed required recovery rate |
| Fixed non-energy operating expense | USD 20 million/year | Same in every case |
| Annual operating period | 8,760 hours | No leap-year adjustment |

Useful utilisation, u, is the fraction of installed compute-unit hours delivering the stipulated useful service. It is not GPU occupancy, electrical load factor, reservation rate or billed utilisation unless those definitions coincide in a specific contract. Data processing, retries and idle capacity can consume electricity without producing the service counted in the denominator. This distinction prevents an apparently efficient electrical metric from concealing an economically idle installation.

With idle fraction `a = 0.25`, average IT power is `100 × [a + (1 − a)u] MW`. Facility energy is that power multiplied by PUE and 8,760 hours. Annual productive service is `100,000 × u × 8,760` compute-unit hours. These are separate physical and economic quantities. The idle fraction is a simplification to be replaced by measured workload-dependent power curves.

### 6. Capital recovery and the utilisation denominator

Annualised capital is calculated with the capital recovery factor, `CRF = r(1+r)^n / [(1+r)^n − 1]`. At 8%, CRF is 0.250456 for five years and 0.116830 for fifteen years. Annualised compute capital is approximately USD 300.548 million; facility capital is USD 46.732 million. Adding fixed operating expense gives USD 367.280 million before electricity. This is an annualised economic cost, not GAAP depreciation, actual debt service or a claim about financing availability.

The unit cost equals annualised capital plus fixed operations plus facility electricity spending, divided by productive compute-unit hours. Rounded outputs appear below; calculation uses unrounded intermediate values.

| Useful utilisation | Average IT MW | Facility MWh/year | Annual cost, USD million | Productive hours, million | USD/productive unit-hour |
|---|---:|---:|---:|---:|---:|
| 40% | 55 | 578,160 | 407.751 | 350.4 | 1.1637 |
| 60% | 70 | 735,840 | 418.788 | 525.6 | 0.7968 |
| 80% | 85 | 893,520 | 429.826 | 700.8 | 0.6133 |

Electricity use increases as utilisation rises, but cost per useful service falls because the fixed capital burden is spread across more productive hours. This is the central distinction between reducing consumption and improving the economics of a fixed installation. It does not justify wasteful demand creation. The additional hours must produce outcomes that somebody values enough to pay for; otherwise the denominator is merely being redefined to make a ratio look better.

![Illustrative cost per productive compute unit-hour across assumed useful utilisation and electricity prices. All inputs are hypothetical; see the complete model in the article.](https://voltformer.com/content/media/97892fad095a7b130c645c0265ade428a933f725cbd93a81ca09293d223d027a.png)

[Open figure at full resolution](https://voltformer.com/content/media/97892fad095a7b130c645c0265ade428a933f725cbd93a81ca09293d223d027a.png)

Figure 1 varies useful utilisation and the assumed delivered energy price while holding hardware, quality, PUE and capital assumptions fixed. It is a model sensitivity map, not observed market pricing. The table above is the 70 USD/MWh slice. Every colour therefore has an explicit unit and an independently reproducible calculation.

### 7. The price of electricity versus the price of unused equipment

At 60% utilisation, a 10 USD/MWh price increase adds USD 7.3584 million annually, or exactly USD 0.014 per productive unit-hour in this model. A reduction from 60% to 40% utilisation at unchanged electricity price increases unit cost by approximately USD 0.367. This does not mean that power price is unimportant. It means that searching for the cheapest nominal tariff can be a poor strategy if the location causes substantial delay, low useful utilisation or a loss of suitable customers.

The derivative makes the mechanism visible. Energy cost per productive hour is proportional to PUE × price × `[a/u + (1 − a)]`. The idle term increases as useful utilisation falls; fixed capital per useful hour also varies as `1/u`. At low utilisation both the numerator and denominator penalise the operator. A better cooling ratio alone cannot remove that burden. Nor can an attractive theoretical accelerator efficiency compensate for long periods when the deployed system is unavailable or commercially unsuitable.

The conclusion is conditional. A campus with much cheaper equipment, a different replacement cycle or a power-intensive workload can have a very different cost structure. The model deliberately holds these inputs constant to isolate the mechanism. A real investment committee should rerun it with contract quotes, metered power curves, cohort-specific replacement costs and the expected distribution of saleable service hours.

### 8. Commissioning delay can erase a thin investment margin

A separate timing example assumes an initial USD 800 million investment and USD 120 million of constant annual net cash flow for ten years, discounted at 8%. These assumptions are independent of the campus cost model above; the examples must not be combined as though they describe one project. Without delay, the benefit stream has a present value of USD 805.210 million and NPV of USD 5.210 million.

Delay shifts the entire benefit stream by d months, multiplying its present value by `(1.08)^(−d/12)`. The initial expenditure remains at time zero. No extra construction cost, debt interest or operating penalty is added, avoiding double counting of a financing burden already represented by the discount rate.

| Delay, months | Present value of benefits, USD million | NPV, USD million | Benefit present value lost, USD million |
|---|---:|---:|---:|
| 0 | 805.210 | 5.210 | 0.000 |
| 6 | 774.813 | −25.187 | 30.396 |
| 12 | 745.565 | −54.435 | 59.645 |
| 18 | 717.420 | −82.580 | 87.790 |
| 24 | 690.338 | −109.662 | 114.872 |

![Separate timing-only model: net present value for commissioning delays of zero to 24 months, using USD 800 million initial cost, USD 120 million annual net cash flow for 10 years and 8% discount rate.](https://voltformer.com/content/media/880d0002ba337e7bcf1bafecc392def75eb01444d916e17772d11b8334dcd458.png)

[Open figure at full resolution](https://voltformer.com/content/media/880d0002ba337e7bcf1bafecc392def75eb01444d916e17772d11b8334dcd458.png)

Figure 2 isolates timing, not a forecast of actual project delays. A twelve-month delay turns a narrowly positive illustrative NPV negative even without a cost overrun. If spending is also postponed, if the contract extends the revenue term, or if customers are lost permanently, the calculation changes. Those possibilities belong in distinct cash-flow schedules, not in an arbitrary percentage “risk discount” layered on top of every other cost.

### 9. A power agreement should price abandonment as well as consumption

A long-term electricity contract is partly an insurance arrangement against the possibility that specialised infrastructure outlives its first customer. The economically relevant questions concern minimum payments, ramp schedules, collateral, termination compensation and the ownership of reusable assets. A low energy price can be paired with an expensive minimum-demand obligation; a seemingly higher price may include capacity, reliability or construction risk that another offer leaves with the buyer.

The Black Hills announcement is useful because it describes these categories separately. Its statements about customer protection are issuer descriptions of the structure, not our independent verification that every possible network or fuel cost has been allocated. A serious review requests the executed instruments, applicable tariff and regulatory decisions. It checks collateral durability under stress, not merely the counterparty's name at signing.

A project may also transfer a risk to a party unable to manage it. If a customer guarantees years of fixed payments but depends on a short-lived product advantage, credit exposure becomes part of the electricity investment. If a utility guarantees availability while a third party controls fuel delivery, performance exposure remains in the chain. “Behind the meter” describes an electrical boundary; it does not abolish fuel infrastructure, air permits, water constraints or common-mode failure.

### 10. Water, cooling and community acceptance belong in the operating model

PUE compares facility energy with IT energy. It does not measure whether the computation was useful, whether electricity was carbon-free at the relevant hour, or whether local water withdrawals are sustainable. The US Department of Energy's design guide recommends evaluating IT performance, supporting infrastructure and multiple sustainability metrics together. [DOE design guide](https://www.energy.gov/sites/default/files/2024-07/best-practice-guide-data-center-design.pdf).

For water, state the denominator. Site WUE can be expressed as litres of site water use per kWh of IT energy; multiplying by facility kWh instead would change the result. Site withdrawal, site consumption and upstream power-sector water are different boundaries. A claim of low on-site cooling water does not by itself measure the supply chain. [DOE water-efficiency guidance](https://www.energy.gov/cmei/femp/cooling-water-efficiency-opportunities-federal-data-centers).

The IEA's 14 October report connects data-centre integration with affordability, security and sustainability, and highlights the trade-off between water-saving cooling and electricity use. Those effects are local. Annual global averages cannot establish that a specific site has adequate drought-year water rights or acceptable peak-season heat rejection. [IEA report](https://www.iea.org/reports/powering-artificial-intelligence/executive-summary).

Community acceptance is consequently a cash-flow issue as well as a question of distributional fairness. A credible arrangement specifies who bears reinforcement costs, how environmental performance is measured and what happens when the promised benefits do not appear. Jobs, tax revenues, ratepayer credits and environmental burdens occur on different schedules and accrue to different people. Summing their headline values without a common horizon, counterfactual and incidence analysis is not a social cost-benefit assessment.

### 11. The strongest counterarguments—and tests that could defeat our thesis

The first counterargument is strategic scarcity: an operator may rationally reserve more capacity than it immediately uses because losing access would be even more costly. That can be true. The appropriate treatment is to value the option, its expiry and the cost of carrying it. Calling every idle asset an option without specifying an exercise condition makes the claim impossible to falsify. Useful evidence includes conversion of reservations into saleable service, customer retention and the cost of equivalent capacity elsewhere.

The second counterargument is efficiency. Better models, batching, networking and systems software can expand useful output from the same capital stock. That strengthens rather than defeats the utilisation argument if quality-adjusted output is measured honestly. The thesis would weaken if cheap efficiency gains consistently offset delay and underuse, while customers paid enough for the resulting service to cover replacement and supporting infrastructure. Benchmark throughput without reliability, latency and quality controls cannot establish that result.

The third counterargument is speed through dedicated generation. Dedicated supply may reduce exposure to one grid constraint, but the test is an earlier credible commissioning date with acceptable full-system cost and performance. The alternative remains vulnerable to construction, fuel and permitting dependencies. Comparing one option's advertised completion target with another option's conservative contracted date biases the decision before any arithmetic begins.

### 12. An evidence standard for the next investment cycle

The monitoring panel should include energised and accepted capacity, installation-to-production time, useful utilisation by hardware cohort, realised service revenue, cash capital outlays and lease commitments on reconciled definitions, replacement spending, contracted minimum electricity payments, and site-specific water and reliability performance. These indicators connect engineering delivery with commercial conversion. A rising announced gigawatt total does not substitute for any of them.

Decision makers should require a dated register of assumptions and evidence, a one-page cash-flow boundary, a physical commissioning critical path and a contract termination map. Stress cases should change specified inputs, including correlated failures where justified. They should not present arbitrary optimistic, central and pessimistic labels as statistical confidence intervals. A project should have explicit pause, resize and exit conditions before the first irreversible long-lead order is placed.

Our conclusion is conditional but clear: the economic unit of AI infrastructure is useful service delivered through a complete, financed and socially workable system on time. October's power contracts and financial disclosures are valuable because they expose parts of that system. They do not remove the need to reconcile its boundaries. The next phase of investment quality will be determined by that reconciliation, not by the largest announced number.

### 13. Primary bibliography and reproducibility notes

- [Meta, third-quarter 2026 results, 21 October](https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-Third-Quarter-2026-Results/default.aspx): consolidated cash-flow definitions and forward capex guidance; no campus attribution.
- [AEP, planned 702 MW fuel-cell project, 14 October](https://www.aep.com/news/stories/view/12314/): announcement and expected completion, not operating output.
- [Black Hills, Google service agreements, 6 October](https://ir.blackhillscorp.com/news-releases/news-release-details/black-hills-corp-executes-definitive-agreements-serve-googles): separate generation, service and managed-resource quantities.
- [Oracle, proposed New Mexico energy agreement, 13 October](https://www.oracle.com/news/announcement/oracle-proposes-plan-to-offset-customers-utility-costs-2026-10-13/): proposed spending and credits, not realised savings.
- [Microsoft, fiscal fourth-quarter earnings call, 29 July](https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q4): cash, lease and useful-life distinctions; fiscal period differs from calendar quarters.
- [IEA, Powering Artificial Intelligence, 14 October](https://www.iea.org/reports/powering-artificial-intelligence): policy and electricity-system context.
- [DOE, energy-efficient data-centre design, July 2024](https://www.energy.gov/sites/default/files/2024-07/best-practice-guide-data-center-design.pdf): established metric definitions, not a 2026 operating survey.
- [DOE, cooling water efficiency](https://www.energy.gov/cmei/femp/cooling-water-efficiency-opportunities-federal-data-centers): WUE measurement boundary.

Evidence cutoff: 22 October 2026. All model prices, asset costs, useful lives, utilisation, idle-power behaviour and required returns are assumptions. Numerical outputs retain their units; display values are rounded only at the final step. The two models are independent and exclude the items identified in their respective sections. Their purpose is disciplined comparison and falsifiable reasoning, not a security-price forecast or a claim that any named project's returns have been measured.

[← Back to Scope](https://voltformer.com/articles.md) · [View Related Equipment](https://voltformer.com/catalog) · [JSON](https://voltformer.com/articles/data-center-economics-power-utilisation-delay-october-2026.json)