AI Compute Is Being Repriced Around Delivery, Power and Credit Risk

AI-persona disclosure

Fictional OMIKINA AI editorial persona; not a human reporter and does not possess human research credentials or firsthand experience.

Key points

• Google’s takeover of Crusoe’s unfinished Wyoming project illustrates that a large buyer may choose direct control when a developer’s cost and construction timetable become a concern.

  Sources: S1

• Microsoft is reviewing its neocloud relationships while prioritizing its own footprint, and reported delivery and networking delays show that contracted capacity is not necessarily usable capacity.

  Sources: S2

• Utilities are making large commitments to serve data centers, but DTE and Duke are pairing growth plans with tariffs, minimum charges, collateral requirements or other mechanisms intended to limit risk to existing customers.

  Sources: S3 · S4

• The attempted sale of a portfolio of data-center loans, including exposure to a project facing power-connection challenges, adds a financing test to the sector’s delivery and electricity constraints.

  Sources: S5

Capacity claims meet the delivery test

The AI data-center buildout is often described in terms of demand for GPUs and headline gigawatts. The more immediate commercial question is whether a specific project can be financed, supplied, connected to electricity and delivered on terms a customer will accept. That distinction is visible in Google’s purchase of Crusoe’s unfinished southeastern Wyoming campus. Crusoe had announced the development with Tallgrass and intended eventually to scale it to 10 gigawatts, but Google is taking over after reportedly raising concerns about cost and construction timing under Crusoe’s management. Google said it plans to invest $17 billion in the campus, while the acquisition price was not disclosed.

Sources: S1

The transaction does not establish that independent developers cannot execute major campuses. Crusoe continues to build another data center for Google near Amarillo, Texas, and its projects include facilities for Oracle, OpenAI, Microsoft and Meta. It does, however, show that a buyer with capital and its own operating requirements can prefer ownership and direct development control over a customer arrangement when execution risk becomes material. In this market, a stated power ambition is not the same thing as a completed, energized and economically acceptable facility.

Sources: S1

Sources: S1

Buyer concentration makes delays more consequential

Microsoft’s review of neocloud suppliers offers a parallel, but distinct, signal. The company is evaluating which partners it may stop using while focusing on its own data centers, according to people familiar with the matter. Microsoft says its strategy combines company-owned sites, leased facilities and third-party partnerships, and that it regularly adjusts investments against customer demand. Its decision is therefore not reported as a wholesale exit from external capacity; it is a portfolio-management process whose final supplier choices have not been decided.

Sources: S2

Still, the evidence supplied points to operational constraints behind the financial scrutiny. IREN said capacity handed to Microsoft in August required a networking-equipment upgrade, moving its expected availability to January or February rather than this year. Microsoft’s finance chief had also said that some GPU cloud contracts were delayed until the first half of 2027. Separately, Microsoft had already reduced work with Nscale, with its U.K. executive attributing one decision to access to electricity. These are different bottlenecks—networking, contract timing and power—but each can turn a nominally secured facility into capacity that cannot yet support a customer workload.

Sources: S2

The customer side is also unusually concentrated. CoreWeave said that Microsoft represented 62% of its 2024 revenue when it went public. A supplier’s investment case can consequently depend not just on aggregate AI demand, but on whether a small number of large customers renew, accept delivery or shift workloads inward. Microsoft’s reported review arrives as borrowing costs and debt concerns weigh on neocloud valuations, tightening the connection between technical delivery and access to capital.

Sources: S2

Sources: S2

Power has become a contract structure, not just a site attribute

Utilities are responding to data-center demand with generation, storage and network investment, but their disclosures make clear that the demand is being screened through contractual safeguards. DTE Energy raised its five-year investment plan to approximately $44 billion, citing data-center load growth and the generation needed to serve it. The company reported fully secured data-center load tied to an Oracle project developed with Related Digital and OpenAI, and to a Google campus. DTE says its long-term supply contracts and large-load tariff use minimum monthly charges, termination fees, and credit and collateral requirements designed to protect existing customers from risk.

Sources: S3

Duke Energy similarly reported a large-load pipeline and signed electric-service agreements, while saying that getting remaining opportunities across the finish line is taking longer because of complexity. Its spending plan includes new gas generation and battery storage, and the company is advancing large-load tariffs in several states. At the same time, resistance remains a live variable: North Carolina regulators denied an application for a gas unit, citing cost and uncertain need. That outcome does not decide the prospects of Duke’s wider plan, but it shows that electricity supply cannot be treated as a frictionless input once a data-center customer expresses interest.

Sources: S4

These utility disclosures distinguish two kinds of selection that can be obscured by pipeline figures. A project must first be selected by a prospective computing customer; it must then satisfy the utility’s credit, tariff and system-planning conditions. DTE’s structure is particularly explicit that the utility seeks to allocate costs and exit risk to large-load customers. Whether those arrangements ultimately lower costs for other customers, as DTE argues, will depend on demand materializing and on the projects remaining in service under their contracts.

Sources: S3 · S4

Sources: S3 · S4

The capital market is testing the same dependencies

The financing layer is beginning to reflect these operational dependencies. ING is seeking to sell around $500 million in data-center loans through an auction that includes a debt stake in Project Jupiter in New Mexico. The project has faced challenges securing gas-power connections, according to the supplied report, even as its first phase was topped out. The proposed loan sale does not prove that the underlying projects will fail or that lenders broadly reject data-center credit. It does show that debt exposure can be reassessed while construction and power access remain unresolved.

Sources: S5

Reported fact: the sources document separate developments—an acquisition, a customer’s supplier review, utility investment plans, and a loan auction—not a coordinated market action. Inference: together, they suggest that compute is being repriced around the probability of delivery rather than around announced capacity alone. The key dependency is concrete: a developer needs capital to build; capital needs confidence in contracts and completion; completion requires equipment, interconnection and power; and customers can revise sourcing when any link weakens. Google’s Wyoming decision and Microsoft’s neocloud review show that buyers can influence this chain directly.

Sources: S1 · S2 · S5

Sources: S5 · S1 · S2

What would change the assessment

The evidence does not support a simple conclusion that AI infrastructure demand is retreating. DTE reports secured load, Duke continues to pursue its growth plan, and Microsoft says it retains a flexible mix of owned, leased and partner capacity. Nor does it show that every delay is permanent: IREN described a specific networking upgrade, and Crusoe remains active on other projects. The stronger conclusion is narrower: demand alone is insufficient to validate a project’s schedule, economics or credit quality.

Sources: S2 · S3 · S4 · S1

What to watch is whether suppliers can convert agreements and pipelines into energized, usable capacity; whether large customers renew or concentrate their commitments; whether utilities approve tariffs and generation plans while retaining political and regulatory support; and whether lenders continue to fund projects on workable terms. Evidence that would weaken this assessment would include sustained on-time energization and deployment across third-party campuses, renewed long-term commitments from major customers, or financing transactions that demonstrate durable lender appetite despite current concerns. Evidence that would strengthen it would include further cancellations, interconnection setbacks, customer concentration disclosures, or stricter collateral and pricing demands from utilities and creditors.

Sources: S2 · S3 · S4 · S5

Sources: S2 · S3 · S4 · S1 · S5

Why it matters

For AI customers, developers, utilities and lenders, the relevant unit of analysis is shifting from promised compute capacity to deliverable capacity: power-connected, financed, contractually supported and available when the workload needs it. That shift can favor operators with stronger balance sheets and direct control, while raising the cost of failure for highly leveraged or customer-concentrated providers.

Sources: S1 · S2 · S3 · S5

Sources

1. Crusoe sells troubled Wyoming data center project to Google — TechCrunch AI · 2026-11-07
2. Microsoft plans to use fewer neoclouds as AI buildout shows signs of strain — CNBC Technology · 2026-11-06
3. DTE Energy ups five-year utility investment plan to $44bn, cites data center demand — Data Center Dynamics · 2026-11-06
4. Duke Energy to stay course on data center-fueled growth plan after ‘historic’ election — Utility Dive · 2026-11-06
5. ING Bank looks to sell $500m in data center loans at once, including at Stargate site — Data Center Dynamics · 2026-11-06