The AI industry tore up flat pricing — now its sellers are eating the risk

An Amazon team running a project on Claude Sonnet finished its work $1.8 million over budget — 860 percent past the plan ^[1]. A year earlier there was no such bill to run up. There was a flat monthly seat, and the cost of a query was the provider's problem, not the buyer's. Then the bill arrived, and it was the whole problem. The industry metered itself because its costs had stopped being optional. A trillion dollars in data-center leases sit in hyperscalers' financial footnotes, uncommenced — up from $725 billion in under a month ^[2]. The financing beneath it is layered: private-credit vehicles called special-purpose entities, which hold a deal's assets and debts off a company's balance sheet, backstopped by Meta's $41.5 billion in guarantees for its Hyperion campus and by Nvidia's guarantees capped at $108.5 billion for partners including OpenAI ^[3]. Oracle has already invoked force majeure — the escape clause a tenant pulls in a disaster — on its Project Jupiter campus, angling to defer rent on a buildout it had committed to ^[3]. Against that fixed floor, flat subscriptions stopped adding up, and the providers said so themselves: flat rates had proven economically unsustainable, and through 2026 OpenAI, Anthropic, Google, and GitHub each tore up the monthly fee for token-metered billing — charging by the unit of computation, like a utility ^[4]. Anthropic showed the meter was a decision rather than a shortage. In March it throttled its flat tiers at peak hours, and some Claude Code users hit their limits within minutes ^[5]. Analysts read the throttling as strategy rather than scarcity — a push to move heavy users off subscriptions and onto consumption billing with revenue locked in.

During weekdays between 5am–11am PT / 1pm–7pm GMT, you'll move through your 5-hour session limits faster than before. — Thariq Shihipar

The meter went in to defend revenue — that, at least, is how analysts read it. What it ended up measuring was something no one had priced before: the distance between what a unit of intelligence costs and what it returns. Greg Rowell, whose firm Revenium sells exactly this measurement in the form of per-agent profit-and-loss statements, put the gap plainly.

You are spending now against a return you cannot measure. — Greg Rowell

Goldman Sachs did the arithmetic in August. Two percent of S&P 500 companies had quantified AI's earnings impact at all, and none showed a statistically significant lift over peers; median monthly AI spend per employee rose from $5 to $12, funded mostly by reallocating existing budgets rather than adding to them ^[6]. Apollo's verdict is harsher — zero payoff outside the tech sector — and MIT's survey found 95 percent of enterprise pilots left no trace on a profit-and-loss statement ^[7][8]. Buyers responded the way people respond to a meter they can finally read: they started conserving. Uber burned its entire annual AI budget in four months and imposed $1,500 monthly per-employee caps on agentic coding tools ^[1]. Work got split across cheap and frontier models, because once a query's cost is visible, you stop sending every question to the expensive model ^[9]. Roughly a third of organizations in a McKinsey survey said they had declined a software purchase because AI let them build the tool themselves ^[10]. Then came the pivot, beginning in late summer and crowding into September and October: sellers stopped charging for effort and started charging for results. Consulting moved first, accepting skin in the game — fees contingent on hitting agreed goals — because fixed fees failed to demonstrate clear returns ^[11]. Accenture expanded into fixed-price, outcome-based contracts; despite $84.5 billion in bookings and $11.6 billion in free cash flow, its stock has sagged back toward 2019 levels ^[12]. Law followed: Cooley is walking away from billable hours and toward results-based fees ^[13]. Rachel Proffitt, who leads the firm, put it plainly.

Clients have never been excited about the fact that it took us 82 hours to do something. They don't care. — Rachel Proffitt

B2B software is doing the same, selling consumption-and-outcome hybrids as agents decouple a product's value from the number of human seats; demand for forward-deployed engineers — the people who embed with a customer to force the result — rose roughly 729 percent in job postings ^[14]. The books can barely keep up: RightRev warns that token- and outcome-based pricing makes revenue hard to predict and recognize, raising audit risk ^[15]. Read the direction of travel and it is a risk transfer. The fixed debts sit at the top of the chain; the meter sits on the customers; the customers read it and capped it; and now the services layer — the consultants, integrators, law firms, and software vendors — is signing contracts that guarantee the results the meter just showed nobody can yet measure. They are writing warranties against profit-and-loss statements Goldman found mostly empty. Is the wall real, or a lag? The stress is real but not yet on any balance sheet. Meta's credit-default swaps — insurance against its default — hit records, CoreWeave fell more than 30 percent in October, and below-investment-grade data-center bond issuance is up roughly 460 percent year over year ^[3][16]. Yet no one has written down a data center, and the big money is not treating this as a crash. Morgan Stanley reads September's lull in AI debt issuance as a scheduling artifact rather than weakening fundamentals — summer front-loading and a rates selloff that forced renegotiations ^[17]. BlackRock and Bank of America are blunter still, rejecting the bubble frame and projecting $5 to $8 trillion in global AI spend by 2030 ^[18].

Is this 2000? Are we in a bubble? No — Savita Subramanian

The economists who study adoption lags say the J-curve is normal — the PC boom's payoff didn't show up in productivity statistics until the mid-2000s ^[19]. And there are genuine outliers: TDK has deployed roughly 2,000 AI agents to about 40 percent of its staff, with 10,000 more planned ^[20]. The wall may turn out to be a lag measured in years. But the contracts are being written now. The risk transfer is already marked to market. Accenture's stock sits at 2019 levels on $84.5 billion of bookings — the market pricing the guarantee before any outcome is proven ^[12]. The wall is not on any balance sheet yet. It is in the contracts, and the contracts are already signed.