Previous programme · preserved history

This is the previous Summary, not the pool’s current assessment. Its original text follows, including its 11 and 13 September 2026 evidence dates. References below to “new,” “current” and the next check describe that earlier programme.

The previous programme examined useful parts of the question but did not connect them into a whole assessment. The replacement connects customer value, financing, investment recovery and ownership prices. This is a change in scope and weighting—not a reversal from imaginary activity to universal financial soundness.

The Amazon payment correction was already present here and in report 02. The March CoreWeave schedule below is historical; the current assessment uses a June snapshot and separately identifies later financing, without claiming a complete 18 September balance.

Read the current answer · What changed and why

Is AI a Bubble?

The spending is real. Whether it earns enough back is still open.

Financing and debt evidence checked 13 September 2026 · Cash and capacity comparison checked 11 September. Event dates are identified below.

The answer so far

AI can be useful while some AI investments lose money. The evidence checked here shows real spending and working infrastructure, but not enough to establish returns across the investment chain or justify every price. That leaves the broad bubble question open.

A “bubble” claim is about expectations outrunning what an investment can eventually earn—not a claim that the technology is imaginary. What a bubble means, and what a burst could change →

A long loan can still require earlier repayments

The new CoreWeave case separates a loan’s final day from its principal schedule, borrower collateral from parent guarantees, and a covenant ratio from a cash-flow headline. It examines March 2026 documents—not a current September debt balance.

See the repayment timetable and creditor protections →

Evidence checked 13 September 2026. March financing disclosure; CoreWeave Q1 filing.

Real payments, conditional promises, different risks

A closed funding round is not one bank transfer. SoftBank reports two $10bn OpenAI installments executed, with another scheduled for October. Amazon reports two separate $5bn Anthropic investments, alongside a larger facility whose funding depends on conditions. SoftBank payment schedule; Amazon investment note.

That is stronger evidence than an empty promise. It still does not tell us whether independent customers will ultimately pay enough to cover the whole investment chain. The new report follows the money without pretending a purchasing commitment or guarantee is cash already spent.

The round closed. Did the money arrive? →

What the 11 September cash and capacity checks changed

A lower spending headline did not necessarily mean less building. Microsoft reduced its 2026 capital-spending outlook from $190bn to $175bn while saying its underlying investment expectation was unchanged, excluding an accounting-life effect. The lease definition matters. Microsoft’s explanation

One funding uncertainty is now narrower. Amazon’s filing reports the full $50bn OpenAI investment funded by the July filing, with $28.7bn through June and $21.3bn afterward. Our first baseline had not checked this disclosure. Payment record

Power use is real, but payback is another test. A July PJM event involved about 3,800 MW of actual data-center load switching to backup generation. It does not reveal AI-only use or customer profitability. Independent grid observation

Read the cash and capacity evidence →

The test is not just “does AI work?”

Useful technology: does it help with a real task after cost and quality are considered? Business returns: does the company keep enough value to cover its spending? Investment price: how much future success is already assumed? A yes to the first does not settle the other two.

The first research record examines reported profits and task studies. The second follows cash, leases and capacity. The third separates financing payments from conditional commitments. Together they narrow the question without pretending to answer it from a single headline. See the four dated records

See the money without mixing the labels

Compare Microsoft, Alphabet, Amazon and Meta over the same January–June 2026 period. Cash paid now is kept separate from future leases and management forecasts. The accounts cover whole companies, not only AI.

Open Spending & promises →

What would change the answer?

Stronger evidence of paying customers, high useful utilization and returns after replacement costs would support durable expansion. Persistent underuse, falling realized prices, losses on completed assets or dependence on refinancing would support a less favorable account. Those outcomes must be measured, not inferred from a promise to spend.

How we check the evidence

A guarantee can change from one loan to the next

The CoreWeave case now follows March’s limited parent guarantee into May’s broader payment promise and June’s unsecured note issuance. That makes “who owes whom?” more useful than one blanket label for the company. The March timetable stays dated; later issues do not create a verified current debt balance.

See what May and June changed →

This is research on an open question, not a prediction of a crash or a personal investment recommendation. The answer so far · Research record