Amazon Demotes Token Counts as Anthropic's Customers Spend 9x More
Anthropic's stock-market prospectus, filed October 26, shows its business customers from a year earlier now spending about nine times as much. A day later CNBC reported that Amazon has told its engineers that token counts, the units AI tools are billed by, are a "deeply flawed" way to measure how they use AI. The seller's best growth figure and the metric a large buyer just demoted measure the same thing: volume. Companies that still judge AI by how much of it they consume are judging it by the seller's number.
The sellers are financed on volume that keeps climbing
Anthropic, the maker of the Claude models, filed to list on Nasdaq with first-half 2026 revenue of $16.3 billion, about fourteen times a year earlier. The figure behind that growth is net revenue retention, which compares what last year's business customers spend now with what they spent then. The prospectus puts it near 900% at September 30, a figure the company says has not been independently verified and that leaves out consumer subscriptions. Existing customers multiplied what they spend.
On October 26 three Wall Street banks and the investment firm Blackstone also announced $63 billion to buy chips from the chipmaker Broadcom and lease them to Anthropic from 2027, with Broadcom guaranteeing part of their resale value, per the banks' release. Those lease payments run for years and are paid from customer spending.
Alphabet showed the other side of that bet on October 28: $51.4 billion of quarterly capital spending left its free cash flow at minus $9.1 billion, CNBC reported, and the company plans to spend significantly more in 2027. This week's Newsletter, Alphabet's cash flow and Anthropic's filing put a price on AI, set out that bill; what it depends on is buyers consuming more every year.
The buyers have started to count something else
Amazon, which is also one of Anthropic's largest investors and sells Claude on its own cloud, now names token usage, AI adoption rates and counts of code changes as poor ways to measure how engineers use AI, warning that chasing them raises costs and complexity, per CNBC. The guidance says an engineer who finds that AI does not help with a task is showing good judgment. Amazon shut an internal usage leaderboard in May after staff gamed it, and a spokesperson said it has not capped token use. Consumption is still allowed; it has stopped being the score.
Only 3% of organizations exceed their cost-savings goals through AI productivity gains, according to research the firm Gartner released on October 26, yet 77% of board members expect AI to deliver savings, and 85% of organizations plan to spend more next year. Gartner analyst CV Viverito said organizations will not generate cost savings from AI at this point.
The clearest outcome-first buyer this week was a government. On October 26 the Dutch cabinet set a target of about 25,000 fewer central-government full-time jobs by 2035, with savings of €900 million a year from 2027, Dutch outlets including NL Times reported. AI is one named lever among several, alongside shared IT systems, and the cabinet has not said how much of the saving it expects from AI. What matters is the order of operations: the result came first, with a number and a date, and any tool can now be measured against it.
The strongest case against
The best objection is that the spending is the value. Nobody forces a company to spend nine times more on Claude; it does so because the tasks get done, and Anthropic's prospectus shows the business close to breaking even on operations in the first half. Gartner's own analyst argues that the larger return comes from moving staff time to new work and growth. On that reading, a missed savings target says little about whether the tool pays.
That is fair, and it marks the limit of this week's evidence. It shows the two measures pulling apart; it does not show that the extra consumption is waste. Amazon's memo covers one company's engineers, and Gartner's figures measure cost savings and say nothing about output quality. A customer spending nine times more may well be getting nine times more useful output. What the evidence does show is that one buyer well placed to judge, an engineering organization the size of Amazon's, is no longer willing to assume it.
We argued in Kroger Cuts Jobs for AI as Bank of America Hires to Fix What It Found that usage grows with every employee who finds the tool useful, and that most business cases leave that cost out. The price is moving in the direction that grows volume. This week OpenAI launched GPT-6.1 Terra, a mid-tier model for agents and coding, at about half the per-token price of its main GPT-6.1 Sol model for developers building on it, the tech-news site BleepingComputer reported. A lower unit price makes more volume easy to approve and harder to question.
Who should be uncomfortable
Finance chiefs and operating heads should be uncomfortable if they report AI progress to their board as licenses activated, tokens consumed or share of staff using the tools. Those are the seller's measures, and Gartner's figures suggest most boards hear them as a promise of savings. Also uncomfortable: companies on usage-metered AI contracts with no result written into them, whose bill rises every time adoption succeeds by the only standard they track.
The move, before the 2027 budget closes, is to write the unit of work next to every AI line: contracts reviewed, tickets closed, invoices matched, hours of outside work no longer bought. Then set a consumption ceiling per unit, and approve growth in spend only where the cost per unit falls. The thing to watch is Anthropic's next filing. Retention near 900% will come down as its customer base matures, whatever buyers do. The signal is the split: if growth from existing customers slows while new customers keep arriving, buyers are trimming volume that produced nothing.
Anthropic has told investors, in its own filing, that the business grows because customers spend more every year, and lenders have lined up $63 billion for chips on the strength of that demand. A buyer has no obligation to share the assumption. Consumption is the seller's revenue line; for the buyer it is a cost line that has to be matched against something it produces. The companies that write that match down now, per tool and per unit of work, will negotiate the 2027 renewals from a position the others will not have: they will know which part of the bill is paying for work.