Zeb Evans, CEO of ClickUp, posted on X on May 21 that he’d cut 22% of the company and opened $1 million cash salary bands for engineers who orchestrate AI agents instead of writing code. The tweet has 5.1 million views. His logic: a senior engineer directing agents produces 100x the output of an engineer writing code by hand. So pay the bands that produce the 100x.

The math works under exactly one condition: that the marginal cost of tokens falls faster than human productivity rises. The same week Evans announced his plan, Microsoft canceled internal Claude Code licenses inside its Experiences and Devices division and Uber confirmed it had burned its full 2026 AI budget in four months. Both stories are direct evidence of the opposite condition.

The diagnostic comes before the comp band

Before any $1M band, any 22% cut, or any 100x outcome, your organization needs one thing: a measurement of how many tokens each team consumes today, on which tasks, and what that bill costs compared to the salary of the human the tool is supposedly replacing or amplifying. Without that number, the CEO’s press release isn’t an AI plan. It’s a wager against someone else’s payroll.

That measurement, task by task and not department by department, is the piece most “AI transformation” announcements skip. The band gets decided. The cut gets decided. The vendor contract gets signed. Four months later the Uber bill lands: $500 to $2,000 per engineer per month. The CFO asks whether that’s now a fixed line or a floor, and the team has no answer because nobody measured the baseline.

Yesterday’s post on “lower-value human capital” covered the error of who you cut. Today’s error is the other edge: who you promise a $1M band to before you know what the agent they’re going to be directing actually costs to run.

The week the economics caught up

Four data points published in seven days, all pointing in the same direction.

Microsoft canceled Claude Code internally. In December 2025, the Experiences and Devices division, responsible for Windows, Microsoft 365, Outlook, and Surface, launched a pilot with thousands of engineers. Five months later, executive vice president Rajesh Jha sent an internal memo: licenses get turned off on June 30, the end of the fiscal year. Engineers are moving to GitHub Copilot CLI, not because the product was preferred (Claude Code was the internal favorite) but on unit economics. Microsoft, with effectively unlimited Azure capacity and $13 billion invested in OpenAI, looked at the competitor’s bill and decided it didn’t pencil.

Uber burned its full 2026 AI budget in four months. CTO Praveen Neppalli Naga acknowledged that the company’s 5,000 engineers exhausted the annual AI budget between January and April. Claude Code adoption jumped from 32% to 84% after an internal program that ranked engineers on usage. The result: 95% of the team uses AI tools monthly, 70% of committed code originates from AI, and the monthly cost per engineer landed in the $500 to $2,000 range. The CTO said the company is “back to the drawing board” on its AI budget.

OpenAI raised enterprise pricing 120% year over year. Per consolidated SpendHound data across 851 companies, enterprise OpenAI pricing rose 120.09% year over year while SMB pricing rose 85%. The market price per million tokens fell from $10 to $2.50, but enterprise budgets are rising because usage multiplied faster than the per-unit price dropped.

GitHub is turning off Copilot’s flat-rate plan. Starting June 1, Copilot moves off flat-rate subscriptions and onto usage-based billing tied to AI Credits. One developer documented a projected jump from €67 monthly in April to roughly €966 under the new model. It’s the same transition from insurance plan to electric meter I called out a month ago, now landing across the rest of the stack.

All four are different readings of the same economic fact: when billing shifts from per-seat to per-use, the difference between “productive AI” and “unsustainable AI” depends on a diagnostic most companies didn’t run before buying.

Why the 100x math breaks without the diagnostic

Evans’s premise (pay $1M bands to engineers who produce 100x) assumes the marginal cost of tokens is effectively zero and falling over time. The reality of May 2026 is that the unit cost is falling but consumption is rising faster, and the enterprise bill is going up. Microsoft, an OpenAI shareholder running Azure, couldn’t justify a competitor’s bill. That isn’t a product failure. It’s structural: token economics turn what looks like a fixed commitment into a variable floor that scales with adoption.

I’ve been in computing for 36 years, since 1990 when a Commodore 64 with 64KB was my first machine. I’ve watched this pattern end cleanly twice. The first was Sun Microsystems giving Java away through the nineties to grow the installed base. Oracle bought Sun in 2010 and in 2019 introduced Java SE Subscription, with per-employee pricing that broke enterprise budgets used to treating Java as free for 25 years. The second was VMware, the cheap layer that powered two decades of cloud migration. Broadcom acquired it in November 2023 and customers reported price increases of up to 10x in 2024 after perpetual licenses were canceled and bundles forced. The pattern is identical: a layer enters subsidized to win adoption, the layer becomes critical, the owner charges when reversibility is gone.

The difference in 2026 isn’t the pattern. It’s the clock. Java went 25 years from free to squeeze. VMware took another 25. AI tokens are showing the first squeeze in year three. The compression doesn’t mean the 100x math is dead. It means the room to improvise runs out far earlier than corporate planning calendars assume.

Anyone promising a $1M band on the bet that token costs fall 80% in the next twelve months is repeating what plenty of companies believed in 1998 about Java and in 2014 about VMware. In both cases the subsidy ended. In every single case, the companies that survived were the ones that measured real consumption before the new pricing landed, not the ones that bet on the vendor’s projection.

The one move before the next board meeting

One task for this week, before any salary band or transformation announcement reaches your executive committee. Inventory, team by team, the current spend on tokens and AI tools. Not the vendor’s projection. The actual bill from the last ninety days, segmented by user and by use case. If your organization is still on a flat-rate subscription, the inventory is of raw usage: how many calls, how many tokens, which recurring prompts are consuming the bulk of the allocation.

That inventory is the input to the diagnostic. The output is the question the CFO is going to ask the next time the CEO shares an Evans tweet: which specific task justifies a $1M band when the cost of the agent that person is going to direct grew 120% in twelve months and will keep growing next quarter?

IQ Source’s AI Maestro is built to produce exactly that answer before the decision leaves the committee. Two months of discovery that end with a Process Reality Map, an AI Opportunity Score, and an explicit Go/No-Go gate before any implementation spend. The operating question the discovery answers is the one the CFO just learned to ask: does this task amplify with AI today, at this cost, in this flow? If not, what changes first before promising a 100x outcome or a million-dollar band?

The opportunity this week isn’t to copy ClickUp. It’s to beat the subsidy cycle before the next invoice arrives at the new price and your organization has to choose between the CEO’s public promise and the CFO’s actual math.

Measure the real cost before the next board