AI capacity now faces two underwriting tests: contract quality and power that can actually respond

AI-persona disclosure

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

Key points

• HSBC’s chief executive argues that data-center financing should turn on tenant quality and contract pricing, not on a blanket view of AI infrastructure.

  Sources: S1

• CoreWeave illustrates the tension: it reported substantial customer obligations, but its debt burden and losses leave ongoing access to finance central to the model.

  Sources: S2

• A Morgan Stanley model identifies a U.S. data-center power gap, while a UK grid trial offers evidence that workload flexibility can help manage demand without proving that it replaces new power supply.

  Sources: S3 · S4

The investable asset is not simply a building full of GPUs

The financial question in AI infrastructure is becoming more specific than whether demand for computing will grow. HSBC chief executive Georges Elhedery says lenders and investors need to examine the quality of tenancy agreements and the pricing of contracts, warning that not every arrangement has the strength to support financing. That is a useful distinction because a data center can have high-profile AI exposure while still carrying weak terms, inadequate returns, or a customer base too concentrated to absorb disruption.

Sources: S1

CoreWeave shows why contract analysis and credit analysis cannot be separated. The company had more than $35 billion of debt at its second-quarter earnings and lost more than a billion dollars in the first half of the year, according to The Verge. Yet it also reported $103.7 billion in remaining performance obligations. Its largest customers accounted for almost three-quarters of second-quarter revenue, and the supplied reporting says cancellation rights are limited in specified circumstances. Those obligations can support a capacity build-out, but they do not by themselves erase refinancing risk, operating losses, or customer concentration.

Sources: S2

Sources: S1 · S2

Contracted demand may be an asset, but timing is the liability

Morgan Stanley’s spending model puts a sharper edge on the scale of the dependency. It estimates that hyperscalers and AI cloud providers will spend $1.58 trillion in 2027, including about $195 billion of data-center leases it characterizes as synthetic capital expenditure. The same account says hyperscalers have disclosed more than $1.1 trillion in lease commitments that have not begun. For infrastructure providers, that makes the duration, commencement conditions, pricing, credit support, and termination provisions of a lease as important as the headline commitment.

Sources: S3

The model also warns that financing conditions can change before revenue arrives. It projects the largest cash-flow deficit of its forecast period in 2027 and discusses additional debt, while testing higher borrowing costs. HSBC’s warning aligns with that concern at the underwriting level: broad enthusiasm for AI is not a substitute for a contract that pays sufficiently, begins on time, and remains collectible. The institutional capacity required is therefore financial as well as technical—disciplined project underwriting, credible counterparties, and funding structures that can endure a delay between construction spending and usable capacity.

Sources: S3 · S1

Sources: S3 · S1

Power deliverability is the other side of contract quality

A customer commitment to rent computing capacity implicitly assumes that the underlying hardware can be energized and operated. Morgan Stanley’s power model projects a 32-gigawatt U.S. data-center shortfall through 2028 against 97 gigawatts of need. Its report says operators can buy scarce GPUs ahead of available electricity, supporting chip revenue in the near term but potentially creating hoarded or stranded inventory. This is a different risk from a weak tenant agreement, yet it reaches the same financial result: capital may be committed before the revenue-producing service can be delivered.

Sources: S3

The constraint is not merely the quantity of electricity. Data Center Dynamics describes AI training loads that can move sharply as large GPU fleets start, pause, or complete work. It argues that resilience requires capabilities across different timeframes, including batteries, demand-side flexibility, dispatchable generation, scheduling of non-time-critical computing, and diversified supply. In the UK, contracted offers in the electricity demand connection queue rose from 41GW to 125GW between November 2024 and June 2025; Ofgem indicated that data-center applications were a major driver. A grid connection, in other words, is only one component of operational readiness.

Sources: S4

Sources: S3 · S4

Measured flexibility is promising, not a blanket cure

The strongest operational evidence in the supplied material is narrower than the larger capacity forecasts. National Grid confirmed a UK-first trial in which a London data-center cluster of 96 Nvidia GPUs responded to more than 200 simulated dispatch events over five days in December 2025. The cluster cut power use by up to 40 percent on request while retaining 98.8 percent performance on its highest-priority workloads. That is meaningful evidence that selected AI workloads can participate in real-time grid management without stopping the most important work.

Sources: S4

It does not establish that flexibility closes the projected U.S. shortfall, nor that the trial’s result will transfer unchanged to larger campuses, different applications, or non-simulated grid events. The report itself presents flexibility as a portfolio of measures rather than a single technical fix. Investments in batteries, fast-start generation, scheduling systems, transmission, generation, and power procurement may therefore matter alongside server procurement. Microsoft, Google, and Amazon are cited as pursuing different supply arrangements, underscoring that power strategy is becoming part of capacity strategy rather than a utility detail.

Sources: S4

Sources: S4

What delivered capacity should mean

Inference: the more reliable measure of AI infrastructure progress is not announced capital expenditure, contracted GPUs, or a signed lease in isolation. It is capacity that has a financeable customer contract, installed equipment, a dependable and sufficiently responsive power arrangement, and an operating path to serve workloads at the contracted economics. This follows from the combination of lender emphasis on tenancy terms, the debt sensitivity of compute providers, the prospect of equipment awaiting power, and the evidence that operational flexibility has to be engineered and tested.

Sources: S1 · S2 · S3 · S4

This assessment would strengthen if providers disclose contract start dates, customer concentration, pricing durability, cancellation protections, debt maturities, and the share of facilities with energization secured. It would weaken if power constraints ease materially, broad workload flexibility proves repeatable at much larger scale, or providers demonstrate that equipment ordered ahead of power can be redeployed quickly and profitably. The practical watchpoint is whether commitments convert into energized, revenue-bearing service—not simply whether another financing package or construction announcement is made.

Sources: S1 · S2 · S3 · S4

Sources: S1 · S2 · S3 · S4

Why it matters

AI infrastructure risk is increasingly a systems problem. Credit terms determine whether construction can be funded through the delivery period; grid readiness and flexibility determine whether the resulting hardware can earn revenue. Treating either condition as assumed can overstate the economic value of announced capacity.

Sources: S1 · S2 · S3 · S4

Sources

1. HSBC CEO warns the AI boom won’t make every data center a good bet — CNBC Technology · 2026-10-20
2. CoreWeave’s debt is becoming the whole AI industry’s problem — The Verge · 2026-10-19
3. Morgan Stanley’s favorite AI stock is 92% exposed to the $1.6 trillion build-out it’s bullish on — Fortune · 2026-10-19
4. Power alone won't unlock the next generation of AI data centers — Data Center Dynamics · 2026-10-19