ThinkPatternGet the app
Perspective
TECHNOLOGY · NOV 17, 2026

AI's Sellers Cheered the Token Burn, Then Rationed It First

AI's suppliers spent the spring telling engineers to burn more tokens, then became the first to ration them — and the accounting that would say what any of it earned didn't exist until October.

Amazon set an internal target that 80% of its developers use AI every week, and it ran an internal token leaderboard to make the number move; at Meta, engineers competed on a leaderboard called "Claudeonomics" [1]. The standard-setters were not coy about what the contest was for. Nvidia's chief executive set the going rate for a $500,000 engineer.

if your $500K engineer isn’t burning $250K in tokens, something is wrong. — Jensen Huang

OpenAI's chief executive watched the same race with approval.

I'm excited to see what will happen with tokenmaxxing startups, both for how they work internally and the products they can build. — Sam Altman

That was the spring of 2026: maximum consumption, institutionalized. But the earliest rationing on the record predates the contest's peak. In March, Anthropic quietly shortened Claude session allowances during weekday peak hours without formal notice, touching roughly 7% of Pro-tier users [2]. Analysts read part of the throttling as a nudge toward consumption-priced API plans. The supplier was metering its own demand months before any customer complained. The mechanism that turned a burn contest into a billing problem was simple enough. Through 2026 the major AI providers dismantled flat-fee pricing for consumption-based token billing, so every query became a visible, variable cost, the way a cloud instance is [3]. Those costs often exceeded development budgets once agents ran in production [4], and firms like Uber blew through a year's AI coding budget in months [5]. Then came the human tax: roughly 70% of executives say staff now spend more time reviewing and correcting AI output than the tool saves them [6]. By late July the buyers were saying it out loud. Token expenses were nearly doubling every other month without proportional value, and firms began routing routine work to cheap open-weight models [7]. Palantir's chief executive put the mood plainly.

The basic view among enterprises in this country is, 'I’m going to chillax and waste my time with tokens. I'm going to get no value and they're going to get my IP, — Alex Karpovsky

Moody's head of AI analytics named the moment the bills arrived.

As bills started to pile in, people realized that those new tools are quite expensive and you need to use them wisely. — Vincent Gusdorf

This was not a price revolt. Usage-weighted token prices actually fell roughly 40% between late June and August, and enterprise spending accelerated anyway [8]. Citadel Securities' Frank Flight chalked the pattern up to the Jevons Paradox — the old observation that cheaper inputs get consumed in greater total volume, not less. The July backlash was an accounting revolt: the complaint was never that tokens cost too much apiece, but that nobody could say what the tokens earned. September turned the ledger onto the sellers themselves. Salesforce's cost of goods sold rose 18.2% year over year against revenue growth of 10.8%, dragging gross margins down — its own AI token costs were outrunning its revenue [9]. Benioff's year compresses the reversal to five months, and his own words mark each step. June was the promise.

I am going to probably use $300 million of Anthropic this year at Salesforce. Coding. Everything’s going to be cheaper to make. — Marc Benioff

Across August, September, and October he held the line — dismissing Wall Street's fear, promising the Anthropic spend would come back multiplied, calling the SaaS-apocalypse story nonsense.

Wall Street's Fears That AI Will Kill Salesforce Are "Dead Wrong" — Marc Benioff
I think we'll make not just billions but maybe tens of billions of dollars on it. — Marc Benioff
This ‘SaaSpocalypse’ narrative has been such nonsense. — Marc Benioff

ServiceNow's maturity index, published in early August, had quantified the same gap across the whole market: corporate AI spending up 110% while operational maturity rose just 16 points, to 51 out of 100 [10]. Synthesia's chief executive pushed an AI impact scorecard measured at the workflow level rather than the model level, and called out the tokenmaxxing habit by name — firms grading their AI success by the sheer volume of tokens they burn [11]. Five days before Salesforce capped its own employees' tokens, the public ledger still read as adoption — the company was reporting rapid Agentforce 360 uptake with customer returns and expanding margins [12]. One company, two ledgers. October was when the missing instrument began to be invented. Aon's global head of data and AI proposed an AI FinOps framework — the cloud-cost-management discipline of tying every token spent to a business outcome — arguing that traditional FinOps can't handle variable token pricing or fast-moving models [13]. Flexera's chief architect reported that 36% of leaders overspend on AI applications, and that organizations with full visibility into AI costs are five times more likely to show established returns [14]. Seramount's president urged firms to stop grading AI by how much they adopt and start grading what it changes.

Adoption matters, but it is ultimately a measure of activity. — Subha V. Barry

The cost-to-outcome meter that would tell a firm what its AI usage actually earns did not exist before this October, and even now it exists mostly as proposals. Then Apollo's chief economist Torsten Slok put the flaw in the buildout case in one line.

The core economic flaw in the AI bull case is that the marginal cost of producing compute exceeds the marginal benefit it delivers to users, raising questions about claims of unlimited demand for data centers. — Torsten Slok

The demand case behind the data-center buildout rested on usage figures that included the very numbers the spring's leaderboards were inflating — a point analysts had been making since May [1]. November moved fast. Salesforce capped its own employees' tokens at $1,000 a month for technical sales roles and $100 for everyone else, after individuals ran up thousands of dollars in a single month [15]. Its chief operating officer framed it as a move away from counting adoption and toward measuring impact.

Starting this month, every role has a monthly token budget, so we can focus our AI spend on the work that matters most for our customers and business. — Miguel Milano

Pipefy's chief executive told the market the era of simple AI adoption was over, and that AI now had to answer to the P&L.

The demo gets the meeting, but the P&L gets the budget. — Alessio Alionço

And Anthropic — the lab that had rationed its own sessions back in March — began selling corporations a tool that scans Claude transcripts to flag "low-value" sessions as cost inefficiencies [16]. The waste-detector is the bookend to the leaderboard. Salesforce's caps and Pipefy's mandate landed on November 17, the same day OpenAI shipped its GPT-6.2 models and Cognition launched its Devin Bots manager system [15][17][18][19]. Anthropic's waste-detector had shipped two weeks earlier, on November 2 [16]. The pile-up is coincidence, not coordination. Set the spring's leaderboard beside November's waste-detector and the year reads as a single inversion: the industry built a scoreboard for burning tokens a full year before it began building a meter for measuring what any of them earned.


Sources
  1. 1. Amazon Employees Game AI Metrics in Tokenmaxxing Trend
  2. 2. Anthropic Reduces Claude Session Limits During Peak Hours
  3. 3. Major AI Providers Shift to Consumption-Based Token Pricing
  4. 4. Enterprise AI Agent Costs Exceed Initial Development Budgets
  5. 5. AI Firms Struggle With Unpredictable LLM Token Costs
  6. 6. Executives Consider AI Budget Cuts Amid ROI Struggles
  7. 7. Corporate America Rejects AI Tokenmaxxing Over Rising Costs
  8. 8. AI Token Prices Drop 40% as Enterprise Spending Surges
  9. 9. Salesforce Gross Margins Decline as AI Costs Outpace Revenue
  10. 10. ServiceNow Index Finds Corporate AI Spending Outpaces Operational Readiness
  11. 11. Synthesia CEO Urges Shift From AI Token Metrics to Outcomes
  12. 12. Salesforce Sees Rapid Adoption of Agentforce 360 AI Platform
  13. 13. Aon Executive Proposes AI FinOps Framework for Scaling GenAI
  14. 14. Flexera Proposes Framework to Link AI Spend to ROI
  15. 15. Salesforce Implements Monthly AI Token Budgets for Employees
  16. 16. Anthropic Launches AI Usage Tracking for Enterprise Users
  17. 17. Pipefy CEO Outlines Four AI Operational Trends for 2027
  18. 18. OpenAI Launches GPT-6.2 Astra and Sol AI Models
  19. 19. Cognition Launches Devin Bots for Autonomous Software Engineering

Keep reading in the app

The full perspective, free in the app.