Who Picks Up the Tab Now That AI Borrows?

Disclaimer: This article is for informational and educational purposes only. It does not constitute financial, legal, or investment advice. Views are based on public information as of the publication date and are subject to change.

In February 2026, Alphabet raised roughly $31.5 billion in the global bond market. One tranche was a £1 billion note maturing in one hundred years.

The stock closed down 1.78% that day.

Nobody knows what will be standing a century from now. What we do know is that the GPUs bought with that money get outclassed in about two years. That mismatch is the whole story. The question is no longer whether AI works. It is who is funding it, at what cost, and how soon the money has to come back.

For three years the playbook was simple. Hyperscalers raised capex, chipmakers booked orders, analysts revised up estimates, and the market paid a higher multiple. As long as the tab could be settled out of operating cash flow, you only had to track orders, revenue, and valuation. The funding source was somebody else’s problem.

It is now the only problem. And the answer runs through people who have never owned a single AI stock, possibly including you.

Capex Now Comes With a Coupon Attached

What AI adds to GDP and what it takes out of capital markets are two different numbers.

Build a data center and GDP counts value added, net of imported equipment, intermediate inputs, and whatever older investment got displaced. Capital markets net out nothing. Land, power, GPUs, networking, and cooling are all paid gross, and most of the cash goes out the door before a dollar of AI revenue shows up.

GDP measures value added. The bond market gets the full invoice. AI can add modestly to current growth while creating enormous demand for capital.

Alphabet is the example: In August the company printed another $25 billion and lifted 2026 capex guidance to $195 billion to $205 billion. The book peaked near $115 billion, more than four times covered.

Early in the cycle, a capex raise meant management was seeing more demand than the Street was modeling. The same announcement now also means more depreciation, weaker near-term free cash flow, and a bigger call on external funding.

Bondholders underwrite whether a company can service the debt. Shareholders own whatever is left after it does.

Alphabet can borrow easily, and that tells you about its credit quality. Whether each borrowed dollar earns an excess return for shareholders is a separate question, and it depends on project returns net of the cost of capital.

Capex is no longer growth by default. It has to clear the marginal cost of money first.

Once the Tab Moves to Credit, the Price Changes

Protection on Microsoft, Meta, and Oracle costs more than it did a year ago. None of those balance sheets deteriorated.

Credit is pricing three other things:

1. a sudden increase in bond supply
2. genuine uncertainty about capex returns
3. creditor demand to hedge exposure to the entire AI funding chain.

Through 2026, new issue spreads for Amazon, Alphabet, Meta, and Oracle have generally priced wider than the prior year, with more compensation demanded further out the curve.

I want to be precise about what this does and does not mean. My argument is not that AI is heading into a debt crisis. Alphabet’s four-times-covered August deal shows that quality issuers with real repayment capacity still find plenty of demand. What has changed is the price of that demand, and therefore the return hurdle every new project has to clear.

Equity investors care whether AI revenue eventually arrives. Creditors also care when. Land, chips, and power are cash out today, and coupons are due on schedule. Revenue from AI applications, cloud services, and productivity gains may take years.

That timing gap used to be absorbed by tech balance sheets and their cash piles. Part of it now sits with bondholders.

If AI cash flow compounds faster than funding costs, leverage amplifies shareholder returns. If funding costs compound faster, the pressure travels back to the equity through interest expense, depreciation, and a higher discount rate.

Credit is not asking whether AI succeeds. It is asking whether AI succeeds in time.

Washington and the Hyperscalers Are Bidding for the Same Money

Insurers, pensions, mutual funds, banks, and sovereign funds are the natural owners of long duration. They can buy Treasuries, buy Alphabet’s long corporates, or lend directly into data center projects.

They manage a great deal of money. It does not expand simply because AI requires more of it.

So the Treasury and the hyperscalers now compete for the same pool. Rising government supply lifts the benchmark rate beneath every corporate bond. Concentrated tech issuance widens credit spreads on top of that. When both increase together, borrowers face a higher risk-free rate and a higher credit premium simultaneously.

That is why the AI cycle is now linked directly to the 30-year.

The more difficult development is that stock-bond correlation has become unstable.

1. Strong AI demand can lift tech while capex and issuance rise, so long yields need not fall.
2. Weak AI demand hits tech, but reduced corporate supply and softer growth expectations can rally Treasuries.
3. If inflation reaccelerates, borrowers face both a higher funding cost and a higher discount rate, and stocks and long bonds decline together.

None of these combinations mattered much over the past two years. Once debt is carrying the capex, equities, rates, and credit can no longer be analyzed in separate windows.

Treasury Is Moving Duration, Not Printing

On August 19, the Treasury announced an expansion of its long-end buybacks, raising the per-operation cap in the 10-to-20-year and 20-to-30-year buckets from $2 billion to at least $4 billion, running September 9 through November 4.

The 30-year fell nearly 10 basis points on the announcement and then retraced most of the move. The market quickly connected the operation to QE, liquidity injection, and yield curve control.

That comparison does not hold.

QE is a central bank operation. When the Fed buys assets, it creates bank reserves and expands its own balance sheet. The Treasury has no such capacity. The cash it uses to retire long bonds ultimately comes from taxes, its cash balance, or new issuance, and under the current financing mix the marginal dollar most likely comes from bills.

The net result is fewer long bonds outstanding and more bills outstanding. Total debt is unchanged. Bank reserves are unchanged.

What the Treasury is doing is maturity transformation. Retiring illiquid off-the-run long bonds takes inventory off dealer balance sheets, reduces dealer sensitivity to rate volatility, and clears capacity to absorb upcoming auctions. It improves trading conditions in the long end. It adds no cash to the financial system.

Dealer shelves get cleaner. The cost is a larger stock of bills that has to be rolled. Long-end rate risk has been converted into front-end rollover risk. A problem locked in for decades becomes a problem repriced every few months.

Bills are bought by money funds, bank liquidity portfolios, and stablecoin reserves. They do not object to short duration, but their balance sheets are finite. Keep expanding supply and they will eventually require a higher yield.

Inflation tightens the constraint further. If the Treasury genuinely suppresses long yields, financial conditions ease, growth and asset prices stay resilient, and inflation becomes harder to bring down. The Fed then cannot cut quickly, and each new bill rolls at a high rate.

The more effective this tool proves, the longer front-end funding may stay expensive.

The Treasury can clear the shelf. It cannot generate revenue for AI companies, and it cannot absorb return risk on behalf of creditors.

Wall Street Is Turning Compute Into Paper

Once the corporate bond market began demanding more compensation, AI buildouts had to look elsewhere for capital. Wall Street’s approach is to take compute assets that were previously unfinanceable and convert them into something that can be valued, hedged, and packaged.

Nvidia is building an AI infrastructure financing platform with large asset managers, which addresses the supply of capital. CME plans to launch H100 and B200 rental index futures on October 5, which addresses the pricing of risk.

The second one is necessary because a GPU makes poor collateral.

Product cycles are fast, so rents and residual values on installed equipment can reset as soon as a new part ships. The secondary market for high-end GPUs lacks depth, and decommissioning, testing, freight, and redeployment all carry real cost. The most serious issue is the correlation: when compute demand falls, operator revenue declines and default probability rises, and the lender takes possession of precisely the asset whose rents and residual value are falling at the same moment.

Borrower health and collateral value are driven by the same variable. Collateral of that kind provides almost no protection in a crisis.

The CME contracts are quoted in dollars per GPU hour, at 730 GPU hours per contract, equivalent to one GPU for one month. Settlement is financial, with no delivery of GPUs or cloud capacity.

The value is visibility. A public spot index and forward curve indicate whether compute is expected to be tight or oversupplied, allow creditors to test the rental assumptions embedded in project finance models, and allow operators to sell forward rents and hedge part of their price exposure.

What futures cannot cover: customer churn, utilization declines, power costs, facility leases, hardware failure, and operating expense. The clearinghouse assumes counterparty risk on a futures contract, not credit risk on a GPU loan.

Financializing compute proceeds in three stages:

• Commoditization: public indices, forward prices, standard contracts
• Securitization: operators can hedge, and creditors accept hedged compute cash flows
• Collateralization: banks, insurers, rating agencies, and regulators broadly accept the valuation and risk methodology

Only the first stage is confirmed. The second depends on volume, open interest structure, and basis risk after launch. The third remains a projection.

Compute futures are a thermometer first and a financing tool second.

Thermometers expose things. Private credit and data center projects are typically marked on internal models. If the futures curve sits persistently below the rents assumed in those models, the gap currently hidden inside private marks becomes publicly visible for the first time. The same instrument can expand financing capacity or accelerate a repricing.

Where the Paper Ends Up

Hundreds of billions of dollars of AI debt cannot remain with underwriters. It has to find balance sheets capable of holding it for decades.

Insurers and pensions collect premiums and contributions today and pay out over the following decades. They require long duration, investment grade quality, and a yield above Treasuries. Data center debt, once guaranteed, rated, and structured, satisfies all three.

Meta’s Hyperion project with Blue Owl is a clear template. It is financed through an SPV, roughly $27 billion in total size, with Blue Owl vehicles holding 80% of the joint venture equity and Meta holding 20%. PIMCO took down the bulk of the investment grade debt, with BlackRock and other long-term money participating.

The load-bearing term is a 16-year residual value guarantee from Meta. If leases are not renewed or are terminated early, Meta pays the project under the agreement. Rating agencies treated that guarantee as the primary support for the investment grade rating.

Consider what actually happened. Technology obsolescence risk on the GPUs did not disappear into the SPV. It returns to Meta through the guarantee, and the rating and structure then convert it into debt suitable for long-term institutional portfolios.

The path runs from data center, to guarantee, to rating, to SPV, to private credit, to insurance general account, to pension fund.

Insurers are not passive buyers sitting at the end of that chain. They run their own credit teams and are actively reaching for spread. But long-dated liabilities create a standing need for this kind of product, and AI infrastructure debt is now supplying it in size.

Once risk reaches those accounts, its behavior changes. Money market and repo funding can withdraw within a day, and losses turn into runs. Insurance general accounts and pensions carry far longer liabilities, and private assets have no continuous mark. A problem does not have to announce itself as a bond price collapse or a closed funding market.

It can surface years later as returns below actuarial assumptions, lower policyholder payouts, wider pension deficits, private asset writedowns, or a regulator requiring additional capital.

Short-term volatility has been suppressed. The risk has moved from portfolios that mark daily into portfolios that rarely mark at all.

The debt did not disappear. It changed owners. The risk did not disappear. It changed how and when it will appear.

So Where Does That Leave Retail?

The chain above terminates in insurance and pension balance sheets. The money in those accounts is yours.

I mean that literally. Your capital touches AI capex in at least four places, and three of them are easy to miss.

One: what your tech position actually owns has changed. You bought a high-growth, asset-light, high-free-cash-flow business model. As capex shifts from cash flow to debt, what you hold increasingly resembles a leveraged infrastructure company whose core asset depreciates every two to three years. Same ticker, different risk profile. Which is why free cash flow, net leverage, and new issue spreads now matter more on the print than the revenue growth rate.

Two: your policy and your pension are already allocated to AI. You may own no AI stocks at all. But if you hold a savings-type policy, an annuity, a corporate pension, or any sleeve with private credit exposure, Hyperion-type debt is probably already in your money. It carries no daily quote, so you observe no volatility. The absence of a price is not the absence of exposure. If it goes wrong, it appears as a lower realized payout rate or an actuarial gap years from now, not as a bad session on the screen.

Three: this chain is repricing every asset you own. Long yields set government interest expense, corporate borrowing costs, mortgage rates, and equity multiples. AI debt and Treasuries are competing for the same long-duration money, which means the scale of AI capex now reaches your mortgage, your bond fund, and your equity multiples through the long end. It has stopped being a sector story.

Four, and the most practical: you now have new instruments to watch. Retail tracking of AI had exactly one entry point, the order number on Nvidia’s print, which is lagging and has already been telegraphed by the sell side. The funding chain quotes publicly every day: hyperscaler new issue spreads, CDS, how long the buyback keeps the 30-year down, and the compute futures forward curve. Free, updated daily, and earlier than orders.

The retail disadvantage in this cycle was never slow news. It was watching a single dashboard. There are several now, and all of them are public.

The Market Already Changed the Test