The free-for-all is over: AI infrastructure now has to pencil out

Meta reported negative free cash flow this week. Operations produced $32.79 billion in the third quarter and capex took $34.91 billion, so free cash flow was minus $2.12 billion. A year earlier it was plus $10.6 billion. Meta also approved a new class of non-voting stock, to raise capital among other uses. The market noticed. The Nasdaq fell 4.5% last week and the semiconductor index had its longest losing streak since 2008.

Source: Meta Form 8-K filings for Q1, Q2 and Q3 2026.

Show data

Meta’s free cash flow turned negative in the third quarter of 2026

	Meta free cash flow
Quarter 1	$12.39B
Quarter 2	$0.78B
Quarter 3	-$2.12B

Tom Lee told CNBC this week that AI is still the main source of S&P earnings growth, and he called the pullback a buying chance. I think he is right on the top-down view. AI is producing real earnings.

The bottom-up view is different. I see neoclouds shopping projects with project IRRs as low as 7%. That is below the cost of the debt. A sector can earn a good return while individual projects do not. This post is about that gap and what closes it.

Who pays for the gap

Morgan Stanley counts $2.9 trillion of data center spending through 2028 and finds that hyperscaler cash covers about half. The rest has to come from outside. The table shows the split and what changed this month.

Exhibit 2. Who funds data center spending through 2028

Source	Morgan Stanley estimate through 2028	What changed
Hyperscaler cash flow	~$1.4T (implied)	Meta free cash flow turned negative in Q3
Investment-grade bonds	~$200B	Already passed: Alphabet, Amazon, Meta and Oracle issued more than $228B by October 13
Private credit	~$800B	Lends against contracts and assets
ABS and CMBS	~$150B	Needs a stable lessee and stable cash flow
Equity, sovereign, other	~$350B	Meta adds a non-voting class in January 2027

The hyperscaler line is implied: our arithmetic on Morgan Stanley’s figures. Scopes differ across forecasters, so do not add them.

Source: Morgan Stanley; company filings; LSEG via investinglive, October 13, 2026.

The pressure is rising. On October 21, Meta narrowed its 2026 capex guidance to $135 to $145 billion, from $130 to $145 billion in July and $125 to $145 billion in April, raising the low end each time. Goldman expects hyperscaler bond issuance of about $250 billion this year and $400 billion in 2027, roughly 35% of capex. A Morgan Stanley note this month puts hyperscaler and neocloud cash burn near $600 billion in 2027, as reported by Crypto Briefing.

Almost every cheap line in that table lends against a company. A GPU fleet has no company behind it, only a contract and an asset. The credit that reaches it is priced on the contract and on how long the contract runs against the loan. I covered cost of capital by tier separately. At the weak end, CoreWeave’s 2032 notes yielded about 15% this month and S&P rates the company B+. Those financings differ greatly in size, structure and credit quality from ours.

That is the public price signal. Private lenders ask the same two questions on smaller deals: who signed the contract, and how long it runs against the loan. A loan that outlasts the contract leaves a balance at the end that nobody has agreed to pay.

Top-down is fine. Bottom-up needs discipline

Gartner’s October 14 press release does the arithmetic. It expects AI providers to spend nearly $1 trillion in 2026 and $6 trillion more from 2027 to 2030. To earn a 12% return on that, data center owners need $8.5 trillion of cumulative revenue. Gartner forecasts no more than $3.6 trillion. It says more than $3.5 trillion of spending is exposed to impairment by 2030, and it expects more than 80% of neoclouds to lose their independence by then. Hyperscalers get a cushion from diversified revenue. Neoclouds do not.

Lee’s view and Gartner’s fit together. Revenue is real and growing. It does not grow fast enough to pay for every project at every price. Gartner’s John Lovelock put the return neoclouds need at 18%. Top-down numbers average across everything that gets built. A lender funds one project at a time. The chart shows where a 7% project sits.

Hyperscaler and small neocloud costs of capital are asset-level WACC estimates from our post on the cost of capital by tier (September 29, 2026).

Source: Gartner, October 14, 2026 and October 20, 2026; TIKR, October 2026; GPU Capital estimates.

Show data

A 7% project IRR sits below every cost of capital and hurdle rate

Rate	Percent a year
Project IRR shopped by some neoclouds	7%
Hyperscaler cost of capital	About 11%
Gartner general hurdle	12%
CoreWeave 2032 bond yield	About 15%
Gartner neocloud hurdle	18%
Small neocloud cost of capital	About 23%

Rents are not collapsing, but they are not keeping up. The SemiAnalysis H100 one-year contract index rose from $1.70 to $2.35 per GPU-hour between October 2025 and March 2026, up 38%. Equipment rose faster. NVIDIA’s list price for the RTX PRO 6000 went from about $8,565 at launch to $16,000 in August, up 87%, and Meta raised its capex range this year partly on higher component pricing. Those are different products, so treat the comparison as rough.

Neoclouds also compete for the same customers, and many of those customers do not necessarily have free cash flow. That gives the buyer the leverage on price. The neocloud faces a cost that rose and a rate that cannot follow, and some respond by underwriting projects that do not work and hoping a later rate or contract fixes them. That is how a 7% IRR gets shopped.

Here is what 7% means in plain terms. A project that earns 7% before financing cannot pay a lender 10% and leave anything for the equity. Each dollar of debt makes the equity return worse. That is why these projects get shopped around and not financed.

There is a fair counterpoint. Gartner counts only revenue attributable to data centers, and CoreWeave’s CEO says its backlog covers all of its debt. Both can be true while weaker projects still fail.

The new rule: no capex without a contract and a healthy IRR

For two years the rule was to be first to market. Capacity sold on demand, and the economics came second. Lenders are ending that. I held off underwriting B300s until three- and five-year contracts started to show up. A contract turns a hope for demand into cash flow a lender can test.

This is the test I run before capex, in order.

Exhibit 4. The test before capex, in order

Test	What I look for
1. Signed offtake	Three to five years, from a customer whose own future cash flow from this workload covers the payments. A recent VC raise shows cash today. It does not show who pays in year three
2. Term against loan	The contract runs at least as long as the loan, or a prepayment covers the gap
3. Project IRR	A healthy project IRR at the contracted rate, with real room above the cost of the debt
4. Whole-node payback	Capex per whole node recovered inside the contract. On a three-year term, we estimate H200 and RTX PRO 6000 pay back in about 22 months and B300 in 27

Source: GPU Capital. Payback figures are GPU Capital estimates from the September 29, 2026 post.

For an operator, the rule means signing the customer before ordering the nodes. For a lender, it means asking for the contract first and the hardware quote second. Structure moves risk around. It does not fix a project whose IRR is below the cost of the money.

A project returning below cost of debt fails the third test before the others matter.