Token billing is going crazy: Silicon Valley giants hit the brakes together

Yesterday, as soon as that The Information report came out, HR groups and FinOps groups across Silicon Valley basically exploded. AT&T, Meta, Uber, Microsoft, Walmart, Amazon—virtually every big-name company you can list—have all been doing almost the same thing over the past quarter: putting tight restrictions on employees’ AI tools.

The turning point actually became apparent back in April. Uber CTO Praveen Neppalli Naga admitted during an internal briefing that the company had allocated an annual budget of $3.4 billion for AI coding tools for 2026, and by April it had already been spent. After deploying Claude Code to 5,000 engineers, monthly active usage went straight to 85%-95%, with per-person API call bills ranging from $500 to $2,000 a month, and invoices rolling in like an avalanche. Uber now has a hard rule: each person, for each tool, has a monthly limit of $1,500—go over that, and access stops.

A year ago, things looked very different. In 2025, everyone was shouting tokenmaxxing—Amazon, Disney, JPMorgan, KPMG all had internal rankings: whoever burned the most tokens was considered an innovation pioneer. Visa even had a points system—burn tokens heavily and you could trade them for a coffee machine. The logic was simple then: AI is a new productive force; if employees dare to use it, it’s a good thing. Not using it was what deserved criticism.

Now, it’s the complete opposite.

From “Leaderboard Incentives” to “Quota-Based Controls”

Meta’s story is the most dramatic. In April, an employee spontaneously created an unofficial leaderboard called Claudeonomics. The top-ranked person burned between 281 billion and 328.5 billion tokens in 30 days—according to Anthropic’s published pricing, that’s close to $2 million—a single person in a single month. The day the report came out, the leaderboard was taken down. Zuckerberg reportedly admitted internally that “the incentive mechanism was wrong,” and then quietly began cutting spending on the Anthropic platform.

Amazon’s case is more typical. They had a leaderboard called Kirorank, based on internal Kiro platform scores for employees’ AI usage. People discovered a simple arbitrage: using Agents to run a lot of unnecessary tasks just to rack up points. Amazon Senior VP Dave Treadwell was blunt in an internal memo last month: “Please don’t use AI just for the sake of using AI.” They took the leaderboard offline in May.

We can list the actions of several companies:

• Uber: $1,500 per person per tool monthly limit
• Walmart: Usage limits on internal AI assistants
• Amazon: Canceled the Kirorank leaderboard
• Meta: Cut Anthropic platform spending, shut down Claudeonomics
• AT&T: Restricted employees’ GitHub Copilot access
• Coinbase: Weekly budgets set by job level, ranging from $500 to $5,000
• Microsoft: Ended Claude Code group license for the “Experiences & Devices” department on June 30

The Microsoft example is especially extreme—internal audits found some engineers’ monthly coding token usage reached $2,000. The Claude Code pilot program, rolled out with fanfare in December 2025, was forcibly withdrawn in less than half a year in favor of GitHub Copilot CLI. They had cheaper in-house tools but chose to burn Anthropic’s tokens; finance definitely couldn’t let that slide.

Price Hikes Are the Real Trigger

On the surface, this “hard brake” looks like corporate internal controls waking up. But the underlying trigger is simple: model vendors have collectively raised prices.

From February to June this year, OpenAI, Anthropic, and GitHub all practically synchronized their pricing model changes—from fixed rates to per-token billing. GPT-5.5 doubled its price: $5 per million input tokens, $30 per million output tokens; Gemini Flash 3.5 costs 3–6 times more than the previous version. Over the last six months, high-quality reasoning tokens from cutting-edge models have risen about 40% in price.

This is driven by several forces at once: continued scarcity of high-performance GPUs, data center energy cost increases of 15%-20%, and exploding demand from Agent-based scenarios. Even though model efficiency has improved about 2× over the past year, the 40%-50% token premium still causes API-dependent companies’ net costs to surge by 20%-30%.

GitHub Chief Product Officer Mario Rodriguez explained: “In the old model, a casual question and a multi-hour autonomous coding task cost the same; that subsidy is no longer sustainable.” Translated: model vendors used to subsidize heavy users—now they don’t. A senior Deloitte engineer estimated that under usage-based pricing, a single detailed prompt causing the model to work for hours can cost more than $100.

Priceline IT finance senior director Chris Reed used a harsher metaphor, calling it a “crack-cocaine-style addiction epidemic” — “They let you try it, get you hooked, and then you’re locked in.”

ROI: No One Has Figured It Out

If burning money really brought a corresponding productivity boost, it would be justifiable. But here’s the problem—the money is burned, and there’s no visible output.

A few data points for comparison:

• Engineering management platform Jellyfish found that driven by Agents, per-developer token consumption rose 18.6× over nine months. However, developers who burned the most tokens had productivity only about twice that of low-usage developers, but consumed 10× more tokens.
• Startup EntelligenceAI aggregated data from over 2,000 companies using advanced AI coding tools. Conclusion: only 18% of token spending ultimately resulted in delivered code that reached actual users.
• Bain’s June report: among companies able to quantify AI cost reduction effects, 40% saw actual cost drops of 10% or less; of the 37% who aimed for 11%-20% cost reduction, only 31% hit their target.

Modal co-founder Akshat Bubna put it bluntly: “I’m quite sure 50% of internal token spending is completely useless, but it’s hard to know which 50%.”

Uber COO Andrew Macdonald’s statement last week was representative: “I haven’t seen an increase in token consumption directly improve productivity.”

Faros AI CEO Vitaly Gordon shared a real case: a CTO discovered one engineer burned $40,000 worth of tokens in a month—debating whether to stop them or encourage them. This “afraid of waste yet afraid of missing out” tension is basically the mindset of every CIO in the first half of 2026.

“AI Gateways” and “Model Alternatives” Are the New Buzzwords

Companies aren’t hitting full reverse. Two new trends are rapidly becoming consensus.

First: the AI gateway layer. Microsoft, Databricks, and Factory are all pushing this—basically an intelligent router that assigns tasks based on complexity: administrative tasks like writing emails, reformatting, and document lookup go to cheap open-source models, while complex reasoning is sent to high-end models like Claude Opus or GPT-5.5. Salesforce CTO Parker Harris admitted last week that fiscal year 2026 token spending is “far beyond” plan, and the company is pushing an internal metric called “Effective Output Score” to predict returns and constrain spending.

Second: lightweight Chinese models have become the preferred substitutes. Coinbase has already shifted basic work to lightweight domestic models. Coding Agent startup Command Code saw demand for cheap models surge, gaining 10,000 new customers in 30 days. Harness SVP Trevor Stuart’s analogy nails it: “Using a top-tier AI model for a basic text summary is like driving a Ferrari to buy groceries.”

This is actually the true backdrop for OpenAI Hub-type aggregation platforms being pushed to the forefront—a single key controls all mainstream models (GPT, Claude, Gemini, DeepSeek, etc.), letting companies do model routing and cost control at the gateway layer: simple tasks to DeepSeek or Qwen, complex reasoning to Claude Opus or GPT-5.5, immediately cutting bills. Direct domestic connections and OpenAI-compatible formats eliminate engineering costs for maintaining multiple SDKs. This isn’t an ad—it’s genuinely what enterprise FinOps teams are doing now.

Microsoft CEO Satya Nadella’s recent industry piece advocated “interchangeable AI architecture”—plainly, he doesn’t want a few big models to absorb all enterprise economic value. That statement comes with Microsoft’s own agenda (since OpenAI is both its biggest rival and ally), but the direction is correct.

It’s Not a Bubble Burst, It’s a Subsidy Ending

Should this be interpreted as the “AI bubble bursting”? I think that’s overstating it.

Wells Fargo chief equity strategist Ohsung Kwon shifted from “bullish” in April to “firmly neutral,” emphasizing that token demand growth might be peaking, not demand itself. Nvidia VP for applied deep learning Bryan Catanzaro admitted “in my team, compute costs have far exceeded personnel costs.” But read that in reverse—compute can already replace labor—that’s strong evidence of commercialization.

A more reasonable description: the cheap ‘all-you-can-eat AI’ era is over; the refined operations era has begun.

The Linux Foundation will officially launch a “Tokenomics” foundation in July, with IBM, Oracle, and JPMorgan all signing on. FinOps Foundation executive director J.R. Storment made a vivid comparison—tracking cloud costs involves hundreds of millions of rows of data per month; tracking token costs is a trillions-of-rows-per-month problem. The new foundation’s two core metrics are “cost per unit of intelligence” and “tokens per watt”—AI spending is being folded into the same financial discipline framework as cloud computing.

Hardware won’t save things in the short term. Nvidia acquiring Groq, AMD and Intel redesigning AI accelerators—but Gartner analyst Will Sommer rightly warns: “Chief product officers shouldn’t confuse commodity token deflation with the popularization of cutting-edge reasoning.” Basic models will get cheaper, but token consumption growth from advanced reasoning and Agent tasks will likely outpace unit cost declines.

Goldman Sachs predicts that by 2030, global monthly token usage will surge 24× to 120 trillion tokens. This means the pitfalls everyone is encountering today with AI gateways, quotas, and token routing will be repeated many times over the next five years. The only difference: companies who can afford to trip will survive; those who can’t will be buried by the bills.

The term tokenmaxxing will likely remain as a legacy buzzword for 2025. And 2026’s buzzword has already switched to tokenminimizing.

---

References

• AI Spending Out of Control: AT&T, Meta, and Other Giants Hit the Brakes—From “Crazy Consumption” to “Strict Limits” - IT Home: IT Home’s complete Chinese summary of The Information’s original report—this article’s core source.