Big Tech Has Committed $1.2 Trillion to Unbuilt Data Centres

Moody's published a research note this week with a warning it does not often make about companies of this quality: unprecedented artificial intelligence spending is threatening the credit standing of six of the largest technology businesses in the world.

The companies tracked are Microsoft, Amazon, Alphabet, Meta, Oracle and CoreWeave. The ratings firm projects combined capital expenditure reaching 785 billion dollars in 2026, and approaching one trillion dollars the year after.

The capital expenditure number has been discussed for eighteen months. It is not the interesting one. The interesting one is 1.2 trillion dollars, which is what Moody's calculates these companies have committed in data centre lease obligations. More than 820 billion of that total relates to leases that have not yet commenced, meaning the facilities are still under construction.

Those obligations do not appear as debt. Moody's treats them as debt-equivalent liabilities, because they bind the companies to substantial rent payments for years regardless of what happens to demand.

The transition Moody's is describing

The ratings firm's framing is worth stating precisely, because it identifies a change in business model rather than a change in spending level.

These companies previously operated asset-light structures built on software, intellectual property and scalable cloud services requiring modest capital investment. The move to asset-heavy models demands, in Moody's assessment, unprecedented levels of investment and capital raising.

That is a description of technology companies becoming infrastructure companies. Infrastructure companies carry different risk, are valued on different multiples, and require different financing. The equity market has been slow to reprice this because the revenue still arrives from software, but the balance sheets no longer look like software balance sheets.

The financing evidence supports the point. The largest hyperscalers are expected to raise substantially more debt in 2026 than in 2025, and Alphabet has issued a 100-year bond, the first century bond from a technology company in decades. Its initial 20 billion dollar sale reportedly drew orders many times over. Demand for the paper is not in question. The need to issue it is the signal.

Where the money went, in cash terms

Alphabet reported roughly 44 billion dollars of capital expenditure in a single recent quarter, taking trailing twelve-month spending to around 132 billion, and raised its 2026 guidance to between 195 and 205 billion dollars. Free cash flow, the metric equity investors have historically prized above all others in this sector, turned negative for the first time in the company's history.

Meta's trajectory is steeper in percentage terms. Amazon and Microsoft are on comparable paths. Roughly 90 percent of the group's operating cash flow is being recycled into infrastructure, which leaves very little for anything else.

The competitive logic is not mysterious and the participants have not hidden it. They believe AI compute is a winner-take-most market and none of them is prepared to lose it. That is a defensible strategic position. It is also the precise condition under which capital discipline fails, because the cost of underinvesting is framed as existential and the cost of overinvesting is framed as temporary.

The three exposures that matter

Depreciation is arriving on a delay. Capital expenditure hits the cash flow statement immediately. The depreciation of those assets hits the income statement over their assumed useful life. While spending accelerates, the gap between what is being spent and what is being expensed widens, and reported earnings look stronger than the underlying cash position. That gap closes eventually, and it closes on the income statement. Useful-life assumptions for AI hardware are an estimate about how long a generation of accelerators remains economically productive, made during a period when each generation is being superseded faster than the last. If those assumptions prove optimistic, the correction appears as a step change in depreciation expense across the sector simultaneously.

The lease structure removes the exit. Owned assets can be written down, repurposed or sold. A commenced lease is a contractual obligation to pay rent whether or not the capacity is used. With more than 820 billion dollars of leases yet to begin, these companies have pre-committed to capacity for facilities still in construction, based on demand forecasts made before those facilities were designed. If demand grows as expected, this is efficient financing. If it does not, it is a fixed cost that cannot be reduced by reducing activity.

The demand side is smaller than the supply side. Every operator reports being supply-constrained rather than demand-constrained, which is the strongest available argument for the build-out and is presented as such. It is also, at present, a statement about the current period. Roughly 785 billion dollars of annual spending is being deployed to serve a customer base whose revenue, across the entire pure-play AI vendor category, remains a fraction of the infrastructure committed on its behalf. That gap can close. Closing it requires enterprise AI revenue to grow at a rate that has been forecast repeatedly and not yet demonstrated at the necessary scale.

What this means for companies that are buying, not building

Executives outside this group of six have three practical exposures, and none of them requires taking a view on whether the build-out is wise.

Pricing has a direction. Compute has been sold below the cost of the capital deployed to provide it, because market share matters more than margin during a land grab. Land grabs end. When they do, the operators carrying 1.2 trillion dollars of lease obligations will need those assets to generate returns. Multi-year commitments signed at current pricing are worth more than they appear. Architectures that assume compute costs fall indefinitely are carrying an assumption that the financing structure argues against.

Concentration risk is now credit risk. A ratings agency has formally flagged the credit quality of the companies most enterprise technology now depends on. This does not imply distress at any of them, and none is close to it. It does mean the standard assumption that hyperscaler counterparty risk is effectively zero deserves a fresh look, particularly for the smaller and more leveraged names in the category rather than the four largest.

Your own AI capital plan faces the same arithmetic in miniature. The mechanism generating strain at trillion-dollar scale operates identically at a smaller one: capital committed against a demand forecast, with depreciation lagging and obligations that outlive the assumption. Any organisation with a multi-year AI infrastructure commitment should be able to answer what happens to it if the revenue case slips by two years. The largest companies in the world are currently discovering the value of having answered that in advance.

The read

Nothing here suggests a collapse. These are among the most profitable businesses ever built, they have access to capital on terms almost no other borrower can obtain, and the technology they are building capacity for is real and in demand.

What Moody's has documented is narrower and more useful. The financing structure of the AI build-out has moved from cash flow to debt, and from debt to off-balance-sheet obligations that behave like debt without being counted as it. Each step in that sequence extends the runway and reduces the flexibility.

The industry has committed 1.2 trillion dollars in lease obligations, more than two-thirds of it against buildings that do not yet exist. That is a bet on a demand curve, placed with contractual money rather than discretionary money.

It may be the right bet. It is worth being clear that it is one.