The Numbers That Matter

Alphabet, Amazon, Meta, and Microsoft have collectively guided for $650 billion in 2026 capital expenditure — a 60% year-over-year increase. Amazon alone: $200B. Alphabet: $185B. These numbers dwarf the combined capex of the 21 largest US industrial companies ($180B combined).

The market’s reaction was instructive. Over $950 billion in combined market cap evaporated post-earnings. The Street is beginning to price what infrastructure lenders have known for a year: the gap between capital committed and the physical capacity to deploy it is enormous and widening.

Capital Flows Downhill

Hyperscaler capex doesn’t stay at the top. It flows to developers, colocation providers, equipment manufacturers, and power suppliers — entities that are structurally unprepared to absorb a 60% demand increase in 12 months.

The transmission mechanism looks like this:

  1. Hyperscaler signs lease with data center developer for 100MW+ campus
  2. Developer must now deliver — land, power, cooling, network, on schedule
  3. Developer needs equipment — transformers (18-month lead), generators (12-month lead), switchgear, GPUs
  4. Developer needs capital — construction debt, equipment financing, working capital
  5. Developer approaches banks — who increasingly say no

Step 5 is where the credit opportunity lives.

Why Banks Are Stepping Back

The risk profile of data center development has shifted beyond traditional bank lending models:

  • SPV structures with single-asset risk
  • Pre-revenue borrowers building against forward contracts
  • Construction timelines that keep extending due to supply chain constraints
  • Collateral that banks struggle to value (GPU clusters, custom power infrastructure)
  • Concentration risk — a single hyperscaler tenant, often itself levering up to fund expansion

Mitsubishi HC Capital’s mission-critical finance team put it plainly this week: the risk profile no longer matches bank lending models. Banks haven’t exited the sector, but they’ve moved up the quality spectrum — investment-grade tenants, stabilized assets, proven operators. The marginal borrower, the one building the next 200MW campus with a growing-but-not-IG tenant, is being pushed toward private credit.

The Three Stress Points

1. Power Delivery

This is the binding constraint and it’s getting worse. Utilities in Northern Virginia, the Bay Area, and Atlanta are projecting power delivery timelines 2+ years longer than operators expected. This gap has widened in the last six months, not narrowed.

Bloom Energy reports that one-third of hyperscalers and colocation providers are now planning fully self-powered campuses by 2030. This isn’t preference — it’s necessity. If the utility can’t deliver 100MW in 24 months, you build your own power plant.

But self-generation creates its own bottleneck: gas turbines and transformers have the same multi-year lead times as utility interconnections. Going behind-the-meter doesn’t magically bypass the supply chain. It just shifts who’s waiting.

2. Stranded Assets

Here’s the scenario playing out across the country: A company orders $200M in GPUs. Secures a colocation contract or builds a shell facility. Power delivery slips 18 months. The GPUs sit in crates — or partially racked but unenergized. Revenue is zero. Debt service is ongoing.

This is a classic distressed lending setup. The borrower has hard collateral (the GPUs, the facility, the power equipment) but no cash flow. They need liquidity to survive until energization. A lender willing to underwrite against the equipment value — rather than against projected cash flows — can command significant returns.

The wrinkle: there is no functional secondary market for GPUs. Multiple prior generations remain in active use, which disrupts the depreciation curves lenders typically rely on. Anyone underwriting GPU residuals needs independent conviction on hardware value, not just spreadsheet models.

3. The Inference Buildout

The AI infrastructure narrative has been dominated by massive training clusters — 500MW campuses for frontier model development. But the next phase is different.

Nvidia, EPRI, Prologis, and InfraPartners this week announced a partnership to study 5–20MW micro data centers near utility substations for distributed AI inference. Five pilot sites are planned across the US by year-end 2026.

The framing matters: Nvidia’s Marc Spieler explicitly said the goal is to “unlock stranded power to scale AI inference.” They’re targeting substations with existing grid capacity that’s underutilized — sites that don’t need new interconnections.

The inference market is projected to hit $254B by 2030, with power demand growing at 45% CAGR (versus 30% for training). This buildout will be more distributed, smaller in scale per site, and faster to deploy than training infrastructure.

From a credit perspective, this creates a new class of borrower: smaller colocation providers in Tier 2 and Tier 3 cities who have the real estate and utility relationships but are running out of capital to deploy GPU racks for inference workloads. These are $25M–$100M equipment finance opportunities against hard assets in known locations.

What To Watch

The credit stress in digital infrastructure won’t announce itself with a press release. It will show up in:

  • Bond market pricing — watch 144A paper from digital infrastructure issuers. A 10+ point drop is a signal, not noise.
  • Delivery delays — every month of slippage increases financial pressure on developers carrying construction debt.
  • Tenant credit — the hyperscalers and GPU cloud providers signing leases are themselves levering up. Their credit trajectory directly impacts the developers building for them.
  • Equipment orders — cancellations or deferrals of transformer and generator orders signal that projects are stalling.

The $650B in committed capex will create enormous opportunity. But it will also create enormous stress for the physical infrastructure layer that must absorb it. The lenders who understand equipment value, power infrastructure, and construction risk — rather than just projected EBITDA multiples — will be the ones positioned to capture it.

Celeborn Research publishes independent credit analysis on digital infrastructure. Views are our own. Not investment advice.