Debt, not equity, is now paying for the AI buildout
Hyperscaler spending is projected at about $3.8 trillion from 2026 through 2028, with credit funding more than half. UBS puts AI capital spending at $1 trillion this year alone. The binding limits are physical and political.
- Total projected hyperscaler spending from 2026 through 2028 is approximately $3.8 trillion, and credit markets are expected to finance more than 50% of it. m:…0709
- US investment-grade corporate bond issuance reached $2.2 trillion through September 2026. m:…0709
- Returns on invested capital for AI infrastructure range from 25% to 50%. m:…0709
- UBS forecasts aggregate AI capital spending of $1.0 trillion in 2026, $1.4 trillion in 2027 and $1.5 trillion in 2028. m:…0099
- Power shortages, equipment backlogs and local opposition were already slowing or deterring data center construction before November 2026, and midterm election results add a political layer. m:…5329
- A 12-month delay on a median $2.7 billion data center project can cost between $162 million and $270 million, according to ICE. m:…5639
The money has moved to debt
Technology companies are increasingly tapping credit markets to fund AI infrastructure, and the scale is now the story. Total projected hyperscaler spending from 2026 through 2028 is approximately $3.8 trillion, and credit markets are expected to finance more than 50% of it. Financing is shifting from corporate cash flows and equity markets toward debt markets. US investment-grade corporate bond issuance reached $2.2 trillion through September 2026. m:…0709
UBS puts the same buildout in annual terms: aggregate AI capital spending of $1.0 trillion in 2026, $1.4 trillion in 2027 and $1.5 trillion in 2028. New chip restrictions could affect 5 to 10% of next year's AI demand. The spending keeps rising even as the market argues about whether it will pay. m:…0099
Why lenders are comfortable
Lenders are not being asked to take the risk on faith. Returns on invested capital for AI infrastructure range from 25% to 50%, which is why capital availability is a more pressing constraint for AI investment than borrowing costs. Credit spreads remain near multidecade lows despite rising bond yields. And the structure of the lending is changing: AI infrastructure projects are moving toward corporate credit underwrites because the contracts are single-tenant and non-diversified. m:…0709
The logic is straightforward: when a single tenant signs the contract, the credit decision looks like lending to a company, not like pooling lots of small loans. That concentrates the risk on one borrower, and it also makes the underwriting legible to a bond market that already knows how to price a big corporate borrower. The implication for founders and finance chiefs is that the bond market, not the venture market, is now the marginal funder of the buildout. m:…0709
The real constraint is physical
The binding constraint is not money. Physical constraints like power availability and permitting are major constraints for AI infrastructure investment. Data center expansion was already being slowed or deterred by power shortages, equipment backlogs and local opposition before November 2026. Risks to the continued build-out existed before November 3, 2026, and the November 2026 midterm election results add a new political layer to those existing constraints, according to Meketa. m:…0709 m:…5329
The market is starting to price the physical risk directly. Intercontinental Exchange launched the ICE Data Center Site Tool, which helps developers, investors, lenders and municipalities assess data center project feasibility and delay risk. Its disruption model estimates the probability of construction disruption within 24 months, and achieved 77% accuracy when tested against historical permit outcomes. A 12-month delay on a median $2.7 billion data center project can cost between $162 million and $270 million, and the tool covers more than 3,100 US counties. m:…5639
That reframes the question investors should ask. The debate about whether AI spending is too high assumes the money is the scarce input. The evidence this week says the scarce inputs are megawatts, permits and local consent, and that the political environment around them has just become harder to predict. Credit can fund a buildout that cannot get power or permission, but it cannot fix either. The next bottleneck to watch is not a spread or a rate, but a substation queue and a planning meeting. m:…0709
What is still unresolved
- Whether credit spreads stay near multidecade lows if AI returns disappoint
- How much of the planned spending survives permitting and power delays
- Whether single-tenant corporate underwrites concentrate risk that securitization used to spread
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.