AI data center debt, the $7 trillion question

The buildout is now 75 percent borrowed money. That means the bond market, not the stock market, decides what happens next.

The AI trade had one referee for three years: the stock market. Earnings beat, stock up, spend more. That era is ending. About 75 percent of the AI buildout is now funded with debt, per SemiAnalysis, and debt answers to a different referee. The bond investor does not care about your story. He cares about your coupon.

This argument dominated the smartest podcasts and investor feeds over the past two weeks. Here is the full picture, both sides, with live data.

How big is the AI debt market getting?

AI data center debt is projected to reach $7 trillion by 2029. On the SemiAnalysis Weekly podcast (July 23, 2026), the firm's data center team laid out the math: roughly $11 trillion of cumulative AI and data center spending between 2024 and 2029, with about 75 percent of it financed by debt, mostly on 5 to 6 year terms.

In the team's words, auto loans and student loans "are all in... 1 or two trillion. The only market that's bigger is the US mortgage market at 13 trillion."

Read that again. A debt market that barely existed three years ago is on track to pass every consumer debt category in America except the roof over your head, within three years.

Why bond investors see AI differently than stock investors

A stock investor gets paid by the future. A bond investor gets paid every quarter. On The Circuit podcast (July 15, 2026), the hosts put it in one line about funding AI with borrowing: "You can convince an equity analyst of that... I don't think you can convince a bond investor... you're going to owe me the coupon every quarter. How you going to pay that?"

For three years the buildout was funded by the most patient money on earth, the operating cash flows of Microsoft, Alphabet, Amazon, and Meta. Patient money tolerates a payoff in 2030. Borrowed money does not.

Two things made this real in July 2026:

• Hyperscale credit spreads widened. The extra interest the market charges big tech borrowers over Treasuries rose, which is the bond market's way of saying risk is going up. You can track corporate spreads on FRED.
• Alphabet posted negative free cash flow of $5.9 billion in Q2 2026 (reported July 22), its first negative quarter in years, then raised 2026 capex guidance to $195 to 205 billion, up from $180 to 190 billion (the full numbers). The same report showed Google Cloud growing 82 percent. That is spending ahead of cash, on purpose.

The same Circuit episode added the cost twist most people miss: memory prices have risen so far that memory alone could become "50 to 60% of hyperscaler capex cost." The bill is growing at the exact moment the lender is getting stricter. We covered why memory supply stays tight into 2028 in the gun you can only fire once.

Mark Cuban's bear case: planning for perfection

Cuban's bear case is not that AI demand disappears. It is that borrowed money leaves no room for surprises. On the All-In podcast (July 21, 2026), he said the hyperscalers are "spending all their... cash flow on capex and then they're borrowing on top of that... That's planning for perfection."

Then he named the surprise that breaks it: efficiency. If models get dramatically cheaper to run, "there's going to be a lot of data centers that are going to be turned into pickleball courts."

His reference point is the 1990s fiber optic buildout. Telecom companies borrowed billions to lay fiber, transmission technology then improved roughly 100 fold, and most of that fiber sat dark for a decade while the debt still had to be paid. The asset did not fail. It became too productive, and it stranded everyone who had borrowed against scarcity.

One more detail matters. Cuban thinks the losses would concentrate in private funds: "it could just destroy a lot of VCs and a lot of funds and a lot of PE... they're going all in." Public stocks are not the epicenter. They are not insulated either.

Gavin Baker's pushback: the assets are underearning

Gavin Baker's rebuttal, posted July 29, 2026 and read 955,000 times in a day: "Market is overreacting to hyperscale credit spreads widening... spot prices for GPU rentals are at least 2x higher than contracted rates."

His logic in plain English: hyperscalers signed long term GPU rental contracts at yesterday's prices. Today's open market rate for the same compute is at least double. That means their AI assets are underearning right now, and as old contracts roll into new ones at market rates, operating cash flow accelerates and funds the capex without more borrowing. Baker runs Atreides Management and is positioned, so this is a defended book, not a neutral model. It is also specific and checkable. We broke down his broader framework in cheap AI eats expensive AI.

A second voice supports him. On Forward Guidance (July 22, 2026), Steve Hou, an economist who builds GPU compute price indexes, said the one year GPU forward curve has "actually been monotonically going up... Every single time we saw multiple providers actually raising prices." If a glut were coming, forward prices would fall. They are rising. (Hou sells compute price hedging products. His data is still data.)

What live GPU pricing shows right now

The live numbers side with tightness, for now. Bargo's GPU price tracker shows Nvidia B200 rentals as of July 27, 2026, across 8 tracked providers:

Two signals in that table. Renting compute today costs 15 percent more than committing long term, which is the opposite of a glut, where providers discount today to fill idle machines. And 92 listings are sold out entirely, more than the on demand and reserved listings combined. This squares with what we found in GPU compute is tightening at the high end.

Why the price level matters: on The Circuit (July 20, 2026), Creative Strategies CEO Ben Bajarin ran the payback math. At "$6 an hour and... 75% utilization, they can pay back their NVIDIA capex... in a little over two years," and a full gigawatt facility including power and buildings in about 7 years. Market rates sit above his $6 threshold today, and he notes hyperscalers earn "double digits" per hour on contracted Grace Blackwell capacity. Compute renting above the payback line while most listings are sold out means the debt is being serviced by revenue, not by hope.

The day that table flips, on demand below reserved and availability everywhere, is the day Cuban's scenario starts.

Who carries the risk if it goes bad

Per the SemiAnalysis Weekly discussion, essentially one structure gets financed today: a data center backed by a 5 year purchase commitment from an investment grade hyperscaler. That sorts the market into three tiers:

• Hyperscalers (Microsoft, Alphabet, Amazon, Meta): the kingmakers. Their signature is what makes a project fundable, and they earn the rental spread Baker describes. Microsoft's $190 billion question is the next test of that framework.
• Neoclouds: CoreWeave, Nebius, and their peers live on those purchase commitments. The ones without an investment grade anchor are where a credit squeeze bites first.
• The circular question: Nvidia is reportedly in talks to backstop $250 billion in financing for an OpenAI data center in Ohio that could cost $500 billion in total, per the Wall Street Journal (July 27, 2026). The chip seller guaranteeing the debt of its biggest customer is either smart vendor financing or the cycle funding itself.

And per Cuban, the deepest losses in a bust land on private credit, venture, and PE funds holding unanchored projects, not primarily on public shareholders.

What to watch

• The AI debate has moved from "is demand real" to "who carries the paper." Credit signals now matter as much as earnings.
• The best early warning indicator is GPU rental pricing. On demand above reserved with sold out listings means the debt is being serviced. That table updates daily, alongside the Compute Tightness Index.
• A hyperscaler capex cut would be the signal everyone is waiting for. The opposite is happening: Alphabet raised 2026 guidance on July 22.
• Efficiency breakthroughs are the wild card. Cheaper compute is bullish for AI adoption and bearish for anyone who borrowed against compute scarcity. Both can be true at once. The demand side of that equation is tracked on the Token Demand Index.

This is not investment advice.

Sources

• SemiAnalysis Weekly, Episode 021 (July 23, 2026) — AI debt market sizing, financing structures — semianalysis.com
• The Circuit with Ben Bajarin and Jay Goldberg, episodes of July 15 and July 20, 2026 — bond investor framing, memory share of capex, gigawatt payback math
• All-In podcast, Mark Cuban interview (July 21, 2026) — the structural bear case — allin.com
• Forward Guidance, Steve Hou interview (July 22, 2026) — GPU forward curve data — blockworks.co
• Gavin Baker (@GavinSBaker), X post (July 29, 2026) — the rebuttal
• Alphabet Q2 2026 results (July 22, 2026), cash flow statement via @pequityresearch
• Wall Street Journal report on the Nvidia backstop (July 27, 2026), via @jukan05
• FRED, ICE BofA US Corporate Index option-adjusted spread — fred.stlouisfed.org
• Bargo GPU price tracker, B200 snapshot (July 27, 2026) — bargo.ai/gpu-prices

More research at bargo.ai/research.