PART 1: the price of the ai buildout

Core view: AI demand looks real enough. The challenge is no longer proving demand. It is financing, powering and ultimately monetizing one of the largest infrastructure buildouts in modern business history. 

AI has become one of the few areas of the economy still expanding at full speed.

Data-center construction continues to surge. Hyperscalers are spending at levels that would have seemed implausible only a few years ago. Equity investors have rewarded them accordingly. Credit investors have become noticeably less enthusiastic.

That divergence caught my attention.

The immediate catalyst was Moody’s decision to frame the AI buildout primarily as a credit issue. Until now, most discussions have focused on adoption, productivity and long-term growth. Moody’s looked at the same companies and focused on leverage, commitments and financing structures instead.

That shift is worth paying attention to.

AI spending is no longer confined to the technology sector. Oxford Economics estimates that AI-related investment and associated wealth effects may explain roughly one-third of recent US economic growth. Data-center construction is supporting activity. AI-linked equity gains have supported household wealth. Credit markets are financing the next phase of expansion.

The buildout has become large enough that financing conditions now matter to more than just technology investors. 

Moody’s saw something different

The headline numbers are already familiar.

Moody’s estimates roughly USD 1.2 trillion of data-center commitments for 2026. It identifies around USD 460 billion of direct debt across Microsoft, Amazon, Alphabet, Meta, Oracle and CoreWeave. Hyperscaler spending could approach USD 1 trillion annually by 2027.

The debt itself is not the interesting part. Most of it is visible and already reflected in market pricing.

What stood out was the agency’s focus on commitments that sit outside the leverage ratios investors typically quote.

If lease obligations, guarantees and other contingent liabilities receive fuller recognition within ratings methodologies, leverage can rise without a single additional dollar being borrowed. Nothing changes operationally. The accounting treatment changes.

Markets do not always react kindly to that distinction.

Even Alphabet, arguably the strongest balance sheet in the group, has commitments that would overwhelm its liquid assets if fully recognized. Investors may want to consider the implications for weaker borrowers.

Bondholders get a vote

Most valuation models, backlog projections and revenue forecasts share one assumption: the buildout continues largely on schedule.

That schedule depends on financing.

Bondholders do not need to shut the project down. They only need to ask for a higher coupon, tighter structures or greater compensation for risk. Any of those slows the pace of deployment.

The investment case and the financing case are becoming harder to separate.

A rating-agency warning therefore carries broader implications than a normal sector-specific credit event. If financing becomes more expensive, the effects extend well beyond bondholders and debt issuers. Construction schedules, infrastructure deployment and earnings expectations all begin to move together.

That relationship matters more today than it would have a few years ago because AI capital expenditure has become one of the economy’s major growth engines. 

Following the financing chain

The financing chain deserves more attention than it receives.

Private-credit firms have increasingly acquired insurers and turned insurance assets into long-duration funding sources. Policyholder premiums become investment capital. Asset managers collect fees. Private-credit vehicles originate loans. Insurers often end up holding those same loans.

For data centers, the structure creates genuine financing capacity.

It also introduces several layers of opacity.

Among the risks worth monitoring:

• Fee extraction through affiliated managers.
• Concentration in illiquid private-credit assets.
• Reliance on private ratings.
• Opaque reinsurance structures.
• The migration of difficult-to-sell assets onto insurance balance sheets.

None of these are necessarily problematic in isolation. Collectively, they make risk harder to observe until conditions deteriorate.

History offers a few reminders of how that tends to end.

Two customers, one trade

The market often discusses AI as though it were a broad and diversified investment theme.

The economics look less diversified.

Moody’s highlights the importance of OpenAI and Anthropic, which account for a meaningful share of growth embedded in hyperscaler backlogs. Across the major providers, remaining performance obligations approach USD 900 billion.

Oracle provides a useful illustration. More than half of its AI-related spending is reportedly tied to OpenAI.

Real-estate investors would recognize the risk immediately. It is tenant concentration under another name.

The largest infrastructure buildout in modern business history increasingly depends on a surprisingly small group of private companies, most of which provide far less transparency than public investors normally demand. 

The bond market has noticed

Amazon, Microsoft, Alphabet and Meta are expected to spend roughly USD 725 billion on capital expenditure in 2026.

Goldman Sachs estimates cumulative hyperscaler capex could exceed USD 5 trillion between 2025 and 2030.

Those numbers would be remarkable in any sector. They become more remarkable when attached to companies investors still largely view as asset-light technology franchises.

The financing burden is now visible in credit markets.

AI-related debt issuance represents a meaningful share of new supply. Governments are simultaneously issuing large volumes of debt. Both are competing for the same pool of capital.

For most of the past decade, large technology companies were ideal bond issuers. They generated abundant free cash flow, maintained substantial liquidity and had relatively limited capital requirements.

The AI buildout has altered that equation.

Data centers, networking equipment, semiconductors and power infrastructure are turning some of the most capital-efficient companies in the market into some of its largest spenders.

Credit investors are pricing that shift.

Equity investors appear less concerned.

The power problem

Financing can solve many problems.

Grid capacity is not one of them.

Data-center electricity consumption continues to rise rapidly, while permitting timelines, transmission infrastructure and generation capacity move at a very different pace.

Microsoft’s Azure backlog illustrates the challenge well. Demand is not the issue. Delivering the electricity required to satisfy that demand is.

The comparison with the late-1990s internet buildout remains useful but incomplete.

Back then, infrastructure arrived faster than demand.

Today, demand may be arriving faster than infrastructure.

Transformers, substations, transmission lines and generating capacity do not scale at software speed. 

Valuation leaves little room for disappointment

The demand story remains persuasive.

Capacity is constrained. Backlogs remain substantial. Compute continues to attract buyers.

The market is no longer debating whether AI matters.

The debate has shifted to valuation.

Current prices imply that several years of extraordinary capital expenditure ultimately convert into extraordinary profits, with limited delays, manageable cost overruns, adequate power availability and sustained pricing power.

Large infrastructure projects rarely follow such tidy paths.

That does not mean the AI thesis is wrong.

It means the margin for error has become very small. 

Plenty of capital, at a price

The good news is that substantial capital remains available.

Money-market funds continue to hold trillions of dollars. Sovereign wealth funds possess both scale and patience.

The financing capacity exists.

Mobilizing that capital may require wider spreads, stronger investor protections and more conservative financing structures.

Those conditions are precisely what rating agencies appear increasingly focused on.

Investors may want to pay as much attention to the cost of financing as they do to the demand for compute. 

Final thoughts

What changed this year is not AI demand.

What changed is that financing, customer concentration, power availability and valuation increasingly depend on each other.

The stronger AI becomes as a contributor to economic growth, the less isolated any disruption to the buildout becomes.

Three things are worth watching closely.

• Ratings-agency treatment of commitments, guarantees and contingent liabilities.
• Credit spreads and terms required to place new AI-related issuance.
• Evidence that contracted power capacity is actually reaching the grid.

If AI revenues begin compounding faster than capital expenditure and infrastructure arrives on schedule, many of these concerns will prove overstated.

Until then, spend more time looking at financing conditions than at AI adoption statistics.

That is probably not where most investors are looking today. 

Footnotes

The AI Boom Is Transforming the American Economy Beyond Recognition, The Wall Street Journal, Justin Lahart, August 3, 2026.

Referenced figures in this note also draw on Moody’s, Goldman Sachs, Barclays, Oxford Economics, Gartner, International Energy Agency and company disclosures as cited in the source research note.