In March 2026, NVIDIA CEO Jensen Huang declared at the GTC conference and on the All-In Podcast: "If an engineer earning $500,000 a year isn't consuming $250,000 worth of tokens annually, I'm deeply alarmed." A token is the smallest unit an AI model uses to process input and output—roughly 100 tokens equals 75 English words. Huang framed tokens as "capital that amplifies engineer productivity tenfold," going so far as to say that token budgets had become a recruiting tool in Silicon Valley.
Around the same time, Meta CTO Andrew Bosworth made similar remarks at a February tech conference, noting that the company's top engineers were consuming tokens equivalent to their salaries while boosting productivity up to tenfold. He encouraged the trend: "This is easy money. Keep doing it. No limit."
This atmosphere gave rise to an internal phenomenon known as "tokenmaxxing"—treating AI token consumption itself as proof of productivity, and simply burning through as many tokens as possible. At Meta, one employee built an internal leaderboard called "Claudeonomics," ranking token consumption across more than 85,000 employees. Titles and badges like "Token Legend," "Cache Wizard," and "Session Immortal" turned consumption into a competition. Over a 30-day period, total consumption reached 60 trillion tokens, with the top individual consumer hitting 281 billion tokens. At Anthropic's public pricing for Claude Opus models (roughly $15 per million tokens), that one person's usage alone would translate into costs in the millions of dollars.
When the Bill Caught Up, the Metric Collapsed
The turning point came in April. Two days after The Information reported that Claudeonomics data had leaked externally, its creator voluntarily shut down the leaderboard. Around the same time, Amazon also shut down an internal dashboard after employees began "using AI to use AI." SVP Dave Treadwell told employees, "Don't make using AI itself the goal."
In June, Bloomberg reported that Uber had set a $1,500 monthly cap per AI coding tool—a measure taken after the company had already exhausted its 2026 AI coding budget by April. Some companies reportedly saw AI spending reach $7,500 per employee per month.
Then in July, Adam Mosseri, who leads Meta's Instagram division, said on a podcast that "within a year or two, the cost of token consumption per engineer might match their salary. In that world, we'll need caps"—signaling a shift toward treating token budgets as a managed resource, on par with payroll and capital expenditure.
In just six months, the industry's mood reversed entirely.
| Timing | Key Development | Direction |
|---|---|---|
| Feb 2026 | Meta CTO Bosworth: "No limit" | Encouraging consumption |
| Mar 2026 | NVIDIA's Huang: "Use $250K worth of tokens" | Encouraging consumption |
| Apr 2026 | Meta shuts down Claudeonomics, Amazon shuts down dashboard | Restraint begins |
| May 2026 | Microsoft terminates Claude Code license | Cost reduction |
| Jun 2026 | Uber's $1,500 monthly cap, GitHub Copilot moves to usage-based billing | Budget management |
| Jul 2026 | Meta's Mosseri: "We'll need caps" | Affirming restraint |
| Aug 2026 | Microsoft's Parikh: "Tokenmaxxing is not what we're optimizing for" | Official shift |
The Three Concrete Measures in Parikh's Memo
In early August 2026, Jay Parikh, EVP overseeing Microsoft's CoreAI division, sent a company-wide email. The core of the memo, obtained by 404 Media, is captured in a single line: "Tokenmaxxing is not what we're optimizing for. I want everyone focused on delivering outcomes for customers and the business."
The memo came with three concrete measures.
First, setting AI token budget targets by department. Internal guidelines state: "We're not sharing specific target figures at this time, but data shows that many engineers are currently consuming anywhere from several hundred to several thousand dollars' worth of tokens per month."
Second, switching the default model for internal use from Anthropic's models to OpenAI's GPT-5.6 Sol. Parikh explained that "by shifting more workloads to OpenAI models, we can extract greater value from our token investment"—a decision apparently driven by cost efficiency.
Third, providing an internal dashboard that lets employees track their own token consumption. "We will manage token spend with the same discipline we apply to every other critical resource," Parikh wrote.
A Microsoft spokesperson, responding to an inquiry from The Register, said only: "We have nothing to add."
The Shift to Usage-Based Billing Made the 'Invisible Cost' Visible
Behind this directive lies a change in GitHub Copilot's billing structure. On June 1, 2026, GitHub moved all Copilot plans to usage-based billing based on AI Credits. This replaced the previous Premium Request Unit (PRU) model with billing tied directly to token consumption.
One AI Credit costs $0.01. Copilot Business, at $19 per month, includes 1,900 AI Credits; Enterprise, at $39 per month, includes 3,900 AI Credits. Code completion and Next Edit suggestions remain unlimited, but agentic coding, chat, and code review consume credits. Once included credits are exhausted, additional charges apply within budget limits set by administrators.
What this change brought about was a situation where every time an engineer consumed tokens, the cost became visible in dollar terms. Under the old flat-rate system, there was no need to be conscious of "how much you're using." The shift to usage-based billing set the stage for Parikh's memo.
In May, Microsoft had already terminated Claude Code licenses for its Experiences + Devices division, directing engineers to migrate to GitHub Copilot CLI. An internal memo from EVP Rajesh Jha stated: "When we started offering both Copilot CLI and Claude Code, the goal was to learn quickly and benchmark them against real engineering workflows"—declaring the end of the learning phase and a consolidation around Microsoft's own tools.
Set Against $145.3 Billion in Annual Capital Expenditure
Microsoft's capital expenditure for fiscal year 2026 (July 2025–June 2026) reached $145.3 billion. In the fourth quarter alone, spending hit $41 billion, up 69% year-over-year. Even while achieving $331.8 billion in annual revenue and expanding operating margins, investment in AI infrastructure kept accelerating. The company projects roughly $175 billion in capital expenditure and finance leases for fiscal year 2027.
A company making investments at this scale scrutinizing a few hundred dollars a month in internal engineers' token consumption might seem strange. One anonymous Microsoft employee told 404 Media: "This is the ultimate confession. We, the host of AI infrastructure, can't even afford our own AI products. If that's even partly true, how are the enterprises we sell to supposed to manage?"
However, this view oversimplifies the situation. What Microsoft is taking issue with is not the total amount of spending, but the relationship between spend and outcomes. Parikh's memo also states: "We are not optimizing for fewer tokens. We are optimizing for more impact per token."
| Metric | Old Model (Flat Rate, Unlimited) | New Regime (Usage-Based, Budget Managed) |
|---|---|---|
| Billing model | Flat per-seat pricing (fixed monthly fee) | AI Credits usage-based (1 credit = $0.01) |
| Consumption visibility | Invisible at the individual level | Real-time tracking via personal dashboard |
| Budget management | None (unlimited use) | Department-level targets, limits on overage |
| Default model | Anthropic-based (auto-routing) | OpenAI GPT-5.6 Sol (cost-efficiency prioritized) |
| Success metric | Token consumption volume | Customer and business outcomes |
Why 'Token Consumption = Productivity' Never Worked as a Metric
The reason token consumption failed as a productivity metric is structurally clear. Tokens measure the volume of input and output to an AI model, and they can grow independently of output quality or actual business outcomes. When agentic tools run autonomously for hours, tokens get consumed regardless of whether they align with human intent. At Meta, reports indicate that some Claudeonomics participants had AI agents run lengthy research sessions purely to climb the leaderboard rankings.
Logan Wolfe of Kyndryl, speaking to CIO.com, noted: "When you use token usage as a proxy for productivity, you incentivize volume of output, which moves you further away from outcomes like efficiency, quality, and risk reduction." In software development terms, it's like rewarding the total volume of code written rather than the lines of code actually deployed to production.
Pendo CEO Todd Olson also observed: "I think there's a correlation between maximizing token consumption and productivity, but making it the sole metric creates a vanity metric."
Microsoft's directive represents the loudest acknowledgment yet, from within the industry, of this structural flaw. A company that sells AI infrastructure and champions an "AI-first" identity has now publicly admitted, as an internal problem, the gap between consumption and outcomes.
Questions That Remain
Parikh's memo does not offer a concrete methodology for measuring token efficiency. Since the definition of "outcomes" varies by department and role, designing a unified metric will not be easy.
The fact that department-level budget targets remain undisclosed also obscures how effective this directive will actually be. What restrictions kick in when targets are exceeded, and what happens in cases of noncompliance—both remain unclear.
A more fundamental question also looms. Microsoft positions Copilot as a core tool for "AI-first companies" while simultaneously signaling restraint on its own consumption. How will customers interpret this contradiction? The anonymous employee's question—"If we can't afford our own AI products, how are customers supposed to manage?"—cuts straight to the heart of the cost-benefit debate across the entire enterprise AI market.
With no sign of token prices falling anytime soon, how many companies will be able to demonstrate, in quantitative terms, the relationship between consumption and outcomes? Microsoft's directive has thrown that question at the entire industry.