The $2.5 T IOU – Big Tech’s Off-Balance-Sheet AI Bet

One of the most durable comforts in the AI trade has been the belief that hyperscaler capex is discretionary. If demand for chatbots and coding tools disappoints, the thinking goes, Big Tech simply turns off the taps. Bloomberg Opinion’s Chris Bryant has taken a hard look at the accounting footnotes and concluded that this optionality is largely an illusion. Credit markets, meanwhile, have stopped waiting for the equity crowd to catch up.

Bryant tallies the future spending commitments of the major hyperscalers at more than $2.5 trillion, counting leases not yet commenced plus contracted purchases of equipment, energy, and services. None of it appears on a balance sheet. The Q2 disclosures show how fast the lock in is building. Alphabet now carries $811 billion in purchase commitments, including $707 billion with terms beyond one year, a stunning $475 billion sequential increase. Meta’s noncancellable commitments have reached $349 billion, more than double the figure at the close of 2025, alongside $279 billion in future data center leases. Microsoft sits on $557 billion combined. Even Oracle, the smallest of the group, has $260 billion in pending leases and $19 billion of unconditional purchase commitments running five years.

What changed is the balance of power in the supply chain. On GPUs, the hyperscalers have preserved flexibility. Microsoft’s Amy Hood and Amazon’s Andy Jassy both stressed this week that chip purchases can be slowed if demand softens. But memory is a different animal. Micron has locked customers into agreements of roughly five years covering at least $100 billion of future revenue. Sandisk extracted $42 billion from just three clients. Samsung may dedicate up to 70% of capacity to customers on long contracts secured by deposits and guarantees that make exit punitive. The suppliers, scarred by past cycles, are forcing customers to carry the downturn risk. The tell is buried in Meta’s filings: nearly $11 billion of its money market cash is now restricted from general corporate use, collateralizing infrastructure purchase agreements through 2030.

Credit Risk Rising

The credit market has moved from noticing to trading. A record $127 billion of AI linked corporate bonds changed hands in July across the eight largest names, a more than fourfold rise from a year earlier, lifting the complex to 14.5% of total US investment grade volume from an average of 4.5% in 2025. Much of that paper is only months old, the residue of the bond flood that Meta, Nvidia, SpaceX, and Oracle unleashed to fund their buildouts. After initially absorbing the deluge, investors have pushed funding costs sharply higher and reached for protection.

Credit default swaps (CDS) tells the story in miniature. Oracle single name notional stands at a record $25.1 billion, up from $6.9 billion a year ago, with spreads on five year protection out to 218 basis points, which is junk pricing for an investment grade name. Meta CDS, which only began trading regularly in November, has already amassed $5.8 billion notional. SpaceX hit $1 billion within weeks of its debut bond. Even the picks and shovels are getting hedged. Nvidia five year CDS touched 82 basis points this week, double its level in the middle of June, and tech now accounts for nine of the ten widest names in the CDX investment grade (IG) ndex. As S&P’s Gavan Nolan puts it, “Hyperscalers are the big driver of the investment-grade CDS market right now.”

Importantly, this is repricing, not panic. Traders report healthy flow in both directions, with protection buyers offset by investors selling CDS on the view that the selloff is overdone, and bid/offer spreads have held stable even as credit spreads widened. Hedge funds are working relative value between bonds and CDS and across the capital structure, while insurers and asset managers are returning to single name CDS for the first time in years. A functioning hedging market is a feature, not a bug. But its very emergence confirms the thesis: credit risk is now central to the hyperscaler investment case.

Step back and the macro picture is uncomfortable. Silicon Valley has abandoned its capital light DNA and converted variable spending into quasi debt, precisely as its funding shifted from internal cash flow to public bond markets. If cheaper open source models undercut OpenAI and Anthropic, the contracted revenue backing these pledges erodes, but the pledges remain. That is negative convexity at the heart of the market’s largest concentration of equity value. The capex put is smaller than advertised, the exit doors are narrowing, and the CDS market is now quoting the cover charge.

Markets are getting antsy for a return on all this invested capital, and the pressure is slowly building.

Stay frosty, folks.

*Sources: Chris Bryant, “AI’s Wildest Spenders Are Hitting the Accelerator,” Bloomberg Opinion, July 31, 2026; IFR/market reporting on AI linked credit trading, July 2026.*