Enterprise AI buyers learned to pay less per task as CoreWeave's debt implies 63% five-year default odds

The cost of a unit of AI work is falling inside the companies that buy it ¹, and the credit behind the data-center buildout is repricing at the same time: yields on bonds of CoreWeave▼−0.59% — CoreWeave, down 0.59 percent today, a cloud provider that rents out Nvidia▲+1.42% — Nvidia, up 1.42 percent today chips and borrowed heavily to buy them, hit 15% last week ⁴⁵. The buyers' side of that repricing now has a documented ledger.

FUNDA, a research service built on expert interviews, published its sixth enterprise AI survey today↗, drawn from six companies interviewed between September 19 and October 11, and finds AI spending growth slowing at most of them ¹. The mechanics are the story. A subprime auto lender built a workflow on Claude, Anthropic'sPVT$350B — Anthropic, private, latest valuation $350B model family, for feature engineering, the work of building the variables a credit model uses, one its compliance team can interrogate. It is canceling the roughly $80,000–100,000-a-year specialist tool that workflow replaces ¹.

Buyers have learned to route, self-host and rebuild

The first lever is model choice. After training on which model suits which task, teams at a fintech roughly halved their spending while using more tokens, the chunks of text that models bill by ¹. The second is routing: a pharmacy and health-insurance group is evaluating an automatic router to shift routine work off its priciest models ¹. The third is self-hosting. A pharmaceutical group downloads, verifies and tunes open-weight models, downloadable model files it runs on its own servers, and those now carry about half its calls ¹.

Volumes rise as prices fall

Consumption still grows underneath. The insurer's monthly bill for API access, the pay-per-use way of calling models, roughly doubled in six months, and its DatabricksPVT$201.64B — Databricks, private, latest valuation $201.64B and Snowflake·+2.88% — Snowflake, up 2.88 percent as of its last quote data-platform costs are still climbing ¹. The sellers made the same move: Anthropic priced its small Haiku 5.5 model about 75% below the previous version↗ ². Token volumes keep rising while the price of each unit of work falls.

Whose money this moves

Anthropic is the entrenched supplier: developers at the sports streamer default to Claude Code, its coding tool, production runs on Claude, and one expert's explanation is simply "it got here first" ¹. The survey also shows rising demand for the Snowflake and Databricks data platforms beneath these workflows ¹. Amazon▲+5.06% — Amazon, up 5.06 percent today, named Morgan Stanley's·−1.06% — Morgan Stanley, down 1.06 percent as of its last quote top pick today, is the landlord of the model choice itself, hosting other companies' models through Bedrock ³. The bank expects AWS, Amazon's cloud arm, to grow 47% next year ³.

The payers are seat-priced software vendors. The lender's canceled subscription is one line. The sports streamer has separately proposed cutting its bill for Atlassian's▲+3.41% — Atlassian, up 3.41 percent today Jira ticketing and Confluence wiki products as their use declines ¹.

The exposure is credit priced on yesterday's token margins

Default insurance on CoreWeave's debt implies nearly a 63% chance of default within five years↗ ⁴. A year ago that reading was about 30% ⁴. Bank of America's▼−1.64% — Bank of America, down 1.64 percent today strategists frame the wider question: the largest cloud providers' cost of capital has risen to about 10%, from 4% in 2023, and "monetization vs. cost of capital is the only question that matters" ⁶.

Anthropic sits on both sides of the deflation. Revenue grew about 46% last quarter↗ ⁷, yet some investors now see its November IPO at $1.5 trillion rather than the roughly $2 trillion once expected ⁸.

The bottleneck has moved to monetization

The build is real. Chip equipment is effectively sold out, ASML·+0.52% — ASML, up 0.52 percent as of its last quote, the Dutch maker of the machines that print chips, is taking orders for 2029 and 2030, and rental rates on older Nvidia GPUs are still rising this year, a signature of scarcity ⁶.

AllianceBernstein portfolio manager James Russo expects a significant deceleration in capex growth in 2027 and again in 2028↗; his stated worry is the cost of capital ⁹. On the survey's record, the buyers supply a second, demand-side reason: they have learned to pay less per task ¹.

The payoff goes to the workflow builders

The value is being created one hop above the model, by the builders who turn cheap tokens into workflows that explain themselves to a compliance team and retire a subscription. That lender's in-house workflow deserves the reward, and the debt that assumed per-token margins would hold is paying for the lesson.

The test is close. Third-quarter earnings for the major hyperscalers, the largest cloud providers, are just around the corner ⁹, and the capex lines in those prints are where a buyer-led slowdown would show up first.