Who pays for the gigawatts

2026-09-15

Six companies now set the pace of AI investment: Amazon, Alphabet, Meta, Microsoft, Oracle and SpaceX. Add up their capital spending and the street has it near nine hundred billion dollars this year and around one and a half trillion in 2027, a rise of roughly 70%. Their disclosed future commitments for leases and purchase contracts already exceed two and a half trillion dollars.

The number that matters for investors is not the spend. It is the cash flow after the spend. Across the six, free cash flow turns negative in 2026 and the cumulative burn over 2026 to 2028 is in the order of nine hundred billion dollars, with the trough in 2027. Companies that were cash machines for twenty years are, for a few years, cash consumers.

Where the money comes from

Net cash balances are being run down and the gap is being filled with debt. Hyperscalers have raised several hundred billion dollars of debt and equity since the start of 2025, and estimates for new debt issuance from 2027 to 2029 are above a trillion dollars. Spreads on their bonds widened over the summer as the market absorbed supply it did not expect. Issuers are starting to look outside the dollar market, which is the first time in a long time that Asian and European fixed-income desks have been courted by this group.

Three things follow.

Credit, not equity, is the instrument. The equity story is already priced. The incremental dollar of AI infrastructure is being funded by bonds, bank facilities, leases and project-style structures around data centres and power.
Power is the constraint, not chips. Every operator says the same thing: turbines, cooling and grid connections are longer-dated than GPUs. Capital that can build or finance power sits upstream of the whole industry.
Returns are visible but lagged. Backlogs across the cloud providers total around two and a half trillion dollars. Payback on servers is quoted at one to three years depending on who built the silicon. The spend is not speculative. It is front-loaded.

The second-order trade

Below the six, the money flows into a long tail of counterparties: contractors, equipment suppliers, landlords, power developers and the neoclouds that rent capacity by the watt. Compute is now priced per watt per year, and the quoted range for the newest chips sits several times above what commodity providers charge. That gap is where new entrants and private capital are competing, and where some of them will be displaced when the largest builders bring their own capacity online.

For the receivables and structured-credit work we do, this is the interesting layer. Suppliers to a capex programme of this size carry large invoices against strong names. Financing those invoices is a cleaner way to own the build than buying the equity of the builder.

The consumer side is already moving

The other half of the picture is where the returns show up. AI-referred traffic to retailers is still a low single-digit share of visits, but it converts markedly better and spends more per visit than ordinary traffic. On one large commerce platform, orders from AI channels tripled year on year and new buyers reached their first purchase twice as fast. The infrastructure bill is being paid, in part, by people who do not know they are using it.

Our read

The build is real, the money is borrowed, and the lenders are now a global group. For allocators, the question has moved from "which model wins" to "which part of the capital structure, and which part of the supply chain, is paid for taking the risk". That is a credit and trade-finance question, and it is the one we spend our time on.