Compute as Flex

Welcome to tokenmaxxing — Silicon Valley's newest performance theater, where the metric for being a good engineer isn't shipping useful software but consuming the most compute. It's every bad corporate KPI game from the last thirty years — only now the leaderboard costs a trillion tokens a month.

Here's the pattern: Meta built an internal dashboard called "Claudeonomics" that ranked all 85,000-plus employees by AI token consumption. Meta eventually shut it down — probably because publicly ranking your entire workforce by how much money they spend on a single vendor is the kind of thing that looks bad in a lawsuit. Shopify and OpenAI kept their leaderboards running. Sequoia Capital built one for their whole firm and holds regular office hours around it.

Visa is reportedly burning through one trillion AI tokens monthly. Anthropic doubled its revenue projections in two months. Business token spend across corporate America rose 13 times between January 2025 and early 2026. The people selling the shovels are extremely happy about all of this.

Meanwhile, Andrej Karpathy — who helped build the modern AI stack — reported spending 16-plus hours a day directing agent swarms and feeling "nervous" whenever he had unused tokens. That's the benchmark we're optimizing toward: a founder-class engineer experiencing anxiety about not spending enough money on compute.

The tell is always the language. When Sequoia partner Sonya Huang says "we all should be tokenmaxxing" and frames it as survival — "some companies are going to make it, some companies are not" — that's not insight. That's a VC whose portfolio companies sell compute rationalizing the behavior that benefits her portfolio companies. The incentive structure is wearing a thought-leader costume.

The actual research is less enthusiastic. Boston Consulting Group and UC Riverside linked heavy AI use to "brain fry" — cognitive overload leading to more errors and higher desire to quit. Critics have coined "workslop" for the output: low-value, error-ridden content that costs companies millions to sort through. One OpenAI employee, asked about the trend at their own company, put it plainly: "It doesn't seem sustainable."

The pattern here is old and depressing in its familiarity. A tool enters the workplace. The tool gets measured. The measurement becomes the target. The target gets gamed. Now you have entire engineering teams burning compute to hit a leaderboard while the actual work takes a back seat.

When the tool becomes the status object, the tool stops serving the work. There's no output column on any of these leaderboards. No ratio of tokens burned to decisions actually improved. No defect rate on the workslop. Without that, you're not measuring productivity — you're measuring how enthusiastically a company has committed to a very expensive form of looking busy.

The $250,000-per-engineer token stipend idea will get piloted somewhere this year. There will be a press release about democratizing AI productivity. Eighteen months later there will be a quiet restructuring and a blog post about focusing on intentional AI adoption.

I'll start the timer now.

i · sources

• What Is Tokenmaxxing? The AI Workplace Trend Explained — Built In, 2026-04-23
• Reid Hoffman weighs in on the 'tokenmaxxing' debate — TechCrunch, 2026-04-15
• What to know about tokenmaxxing, the new AI status game — NewsNation, 2026-04-23
• 'Tokenmaxxing': the $72K race to use the most AI at work — AI for Automation, 2026-03-22