What tighter underwriting of hyperscaler capex means for dealmakers, including in the mid-market
The number is no longer the argument
The five largest hyperscalers, Amazon, Microsoft, Alphabet, Meta and Oracle, are on track to spend between 610 billion and 780 billion euros on capital expenditure in 2026, an increase of roughly 36 percent over 2025. Amazon alone has guided to approximately 175 billion euros this year, more than double its 2025 outlay.
For several years, spending on this scale was read almost automatically as a growth signal, the size of the commitment was taken as evidence of the size of the opportunity. Recent market activity suggests that reflex is weakening. Investors are increasingly treating capex as a claim to be tested rather than a conclusion to be rewarded.
One figure sits behind that shift. Hyperscaler capital expenditure is currently scaling roughly 50 percent faster than the revenue it is meant to generate, which extends the payback period on the investment with each passing quarter. A gap of that kind is unremarkable in the early years of any large capital cycle. What has changed is its prominence, it has moved from a technical footnote in sell-side models to a central term in how the buildout is debated.
The financing structure is the part that matters to dealmakers
Historically, hyperscalers funded infrastructure out of internal cash generation strong enough to make debt optional. That is no longer the case. The sector issued approximately 94 billion euros of debt in 2025, and Morgan Stanley and JPMorgan estimate that technology issuers will need to raise up to 1.3 trillion euros in new debt over the coming years to sustain construction at the current pace.
The counterparty picture compounds this. Roughly half of the largest cloud providers’ combined 1.8 trillion euro revenue backlog is attributable to two customers, OpenAI and Anthropic, both of which are currently cash-flow negative. A backlog of that size is ordinarily read as evidence of durable demand. Concentrated in two counterparties whose own funding depends on continued capital-market access, it reads differently, less as contracted revenue than as a position that has to be financed at both ends.
This is a change in diligence standards, not a change in direction
None of this indicates that the buildout is slowing, and it should not be read as a call on the cycle. Hyperscalers have coherent reasons to keep spending. None wants to cede ground to a competitor mid-cycle, and each continues to characterize compute demand as supply-constrained rather than demand-constrained. On the available evidence, that characterization is credible.
What has changed is the standard of proof applied to AI-adjacent assets. A valuation premised on broad, diversified hyperscaler demand is a materially different underwriting exercise from one that traces, at two or three removes, back to a small number of concentrated and unprofitable counterparties. Both may be defensible. They are not the same risk, and they should not attract the same multiple.
The mid-market is more exposed than it looks
This is usually framed as a large-cap and infrastructure-fund question. In practice, the mid-market is where much of the exposure actually sits, for a straightforward reason, hyperscaler capex does not stay with hyperscalers. It disperses almost immediately into a long tail of specialist suppliers, electrical and mechanical contractors, switchgear and busway manufacturers, thermal management and liquid cooling specialists, standby power, fire suppression, structured cabling, modular construction, substation EPC firms, commissioning engineers and grid interconnection consultancies. These are, in the main, businesses with enterprise values between 20 million and 300 million euros. They are the natural hunting ground of mid-market sponsors and consolidating strategics, and their order books are a direct derivative of the capex line under scrutiny above.
Multiple expansion is the first-order risk. Many of these businesses have re-rated over the past two years from mid-single-digit industrial services multiples to low-teens or better, on the strength of data center exposure. That leaves a mid-cap acquirer with two questions rather than one. Is the premium being underwritten as structural or cyclical? And is the target’s recent EBITDA a run-rate or a peak? A buyer who answers both optimistically is exposed twice over, to earnings compression and multiple compression arriving together, which is how mid-market returns are usually lost.
Concentration cascades downward, and standard diligence stops too early. A contractor deriving 60 percent of revenue from a single hyperscaler program has visible, quantifiable customer concentration, and every diligence process will identify it. What conventional analysis does not capture is that the program itself may depend on a small number of cash-flow-negative AI counterparties. The exposure is second-order and therefore invisible to the usual tests. The more useful question in the current environment is not who the customer is, but who the customer’s customer is, and whether the target’s revenue survives a counterparty it has never contracted with.
Backlog quality now matters more than backlog size. Contracted backlog is the headline metric in this segment and the primary justification for the multiple. Its economic content depends entirely on structure. Binding volume commitments, take-or-pay provisions and meaningful termination fees are one thing. Non-binding master service agreements with short notice periods and cancellation for convenience are quite another. A target presenting three years of visible backlog may in fact be presenting three years of optionality, granted to its customer, not to itself. Where valuations rest on multi-year visibility, the underlying contracts warrant clause-level review rather than schedule-level summary.
Financing conditions transmit as well. If technology issuers absorb the volume of investment-grade and private credit that current estimates imply, mid-market leveraged finance competes for capital against data center paper. Direct lenders funding sponsor-backed transactions and the vehicles funding hyperscaler construction draw on overlapping pools. The cost and availability of debt for a mid-cap LBO is therefore not independent of what happens in hyperscaler credit markets, even where the underlying business has no AI exposure whatsoever.
The same conditions create openings. A more scrutinized phase is not uniformly negative for mid-market dealmakers. Private assets reprice with a lag when listed comparables de-rate, which creates a window for sellers holding a peak order book and, subsequently, an entry point for disciplined buyers. Capex pressure at large corporates is a reliable source of non-core divestments, and carve-outs of that kind are a mid-market staple. And the genuine bottleneck in the buildout, grid connection rights, substation capacity, on-site generation, transmission-adjacent engineering, is largely held by mid-cap and family-owned businesses whose scarcity value is physical rather than sentiment-driven. Those assets are defensible under a wider range of AI outcomes than the multiple currently implies.
The practical shift
For M&A professionals evaluating data center, power infrastructure or AI-linked targets, the analytical burden has moved. It is no longer sufficient to establish that a target serves a growing end market. The question is whose demand the valuation actually rests on, how many steps removed that demand is from a funded and profitable counterparty, and what contractual protection exists if the chain shortens.
That question was optional a year ago. It is becoming standard.