← Back to Insights

Insight

Headcount Was a Stock

Ariel Agor
•
Headcount Was a Stock

Listen · Read by Leo

0:00 / —

On May 20, 2026, Meta laid off about 8,000 employees and eliminated 6,000 unfilled roles, roughly 10 percent of its 78,865-person workforce. Weeks earlier, the company had raised its 2026 capital expenditure forecast to between $125 billion and $145 billion, nearly double 2025. In an internal town hall, Mark Zuckerberg framed the cuts plainly: this is about capex, not AI productivity. The salary line shrank so the data center line could grow.

On June 3, 2026, Semafor reported that some employees inside JPMorgan's Payments division were spending more on AI tokens than they earned in salary. Zachery Anderson, the unit's chief data and analytics officer, told the reporter the bank was monitoring it, but ran no usage leaderboards and had no companywide rationing program. The number was bigger than the salary, and the company had no instrumentation to decide whether that was good news or bad.

Twelve weeks before that, on March 20 at GTC, Jensen Huang told an audience of engineers and executives that he would be alarmed if a $500,000 Nvidia engineer used less than $250,000 in tokens a year. He proposed token grants worth roughly half of base pay, on top of cash, as a recruiting weapon. He said the question "how many tokens come with the job" was already showing up in Silicon Valley offers.

Three companies, three different framings, and the same shift underneath.

The salary line on a P&L is no longer the easy one to read. It used to be the most visible cost in any operating budget, the line every CFO knew by heart, the line every quarterly forecast walked through first. That role is moving. Compute is moving in.

Headcount Was a Stock

Headcount is what accountants call a stock variable. You count it on a date. You hire, you fire, the number changes step by step. A salary is a multi-year contract with a known burn rate. You plan around it. A 200-person team costs roughly the same on Monday as it did on Friday.

Compute is a flow. It runs continuously. It varies with prompt volume, model size, context length, retry count, agent recursion depth. A single engineer can ten-x a team's token bill on a Tuesday afternoon by flipping a config flag. There is no quarterly stability. There is no two-week notice period before consumption changes. The bill arrives later than the consumption, which means the spike is already in production before anyone sees it.

Uber learned this in April. CTO Praveen Neppalli Naga told The Information that Uber had blown through its entire 2026 Claude Code and Cursor budget in four months. Engineers' monthly API costs ran from $500 to $2,000 each. By June 3, the company had capped employees at $1,500 per month per AI coding tool. President and COO Andrew Macdonald said in late May that he could not yet draw a line between higher token spend and faster useful feature output. His words: that link is not there yet.

Notice the order. Spend first. Cap second. Measure third. That is the inverted shape of a flow cost colliding with a budget process built for stock costs.

AI cost structure versus headcount, side by side

Pick a typical 200-engineer organization in the United States. Fully loaded engineer cost runs about $300,000 a year, so the salary line is around $60 million. Now layer on tokens.

If each engineer averages $1,500 per month per AI coding tool (the new Uber ceiling), that is $18,000 per engineer per year, or $3.6 million across the team. Six percent of payroll. Most CFOs ignore that line until it hits twice that size.

If half the team also runs heavier agentic workloads at the $7,500 per employee per month figure TechCrunch reported on June 10, 2026 for the top one percent of AI-pilled firms, that subset adds another $9 million. Now you are at $12.6 million in token spend, or 21 percent of payroll. That is the difference between a 15 percent and a 20 percent operating margin at most software companies.

If you take Huang's framing seriously, where a $500,000 engineer is expected to consume $250,000 in tokens, the token line is half of payroll. For 200 engineers, $30 million on top of $60 million in salary. The shape of the P&L changes entirely.

That is the meaning of AI cost structure versus headcount in mid-2026. The two lines are converging. In specific functions, payments middle office at JPMorgan being the most cited, they have already crossed. And the second line behaves nothing like the first.

The variable that broke the budget

Salaries are sticky. Layoffs take quarters to plan and execute. Tokens are liquid. A pull request that adds a new agent loop can change the run rate the moment it merges. The half-life of a compute-budget assumption in 2026 is measured in days, not quarters.

That single property breaks three things at once: how companies forecast cost, how they assign accountability for it, and where the operating leverage lives.

Forecasting stopped working

Most finance teams still build the annual plan around headcount. You forecast hires by quarter, multiply by fully loaded cost, add benefits, true up at year end. The model has worked for fifty years because the underlying variable was slow.

Token spend does not behave that way. Anthropic, OpenAI, and Google ship new models on rolling weeks. A reasoning model that costs ten cents per million input tokens on Monday can be replaced by one that costs forty cents but answers in one third the prompts by Friday. Both the denominator and the numerator move. You cannot extend a Q1 trendline.

Uber's April overrun is the canonical case. Engineers were told to adopt the new agentic coding tools. The tools delivered enough value that engineers refused to ration their use. The 2026 plan was written before Claude Code or Cursor's agent mode existed in their current form. By the time the CTO had numbers, the year's allocation was already gone.

CFOs at hyperscaler customers are now running monthly token re-forecasts the way airlines re-forecast fuel. The annual plan is a rough envelope. The real planning happens every four weeks against actuals, and the variance band is wider than anything the finance team has ever lived with on the human side.

Accountability went sideways

Under the old model, a manager owned a team's salary line. The org chart told you who was responsible for which dollar of headcount cost.

Token spend does not respect that map. A single prompt template inside an internal application can be invoked by every team in the company. An agent built in one division can recursively call services owned by another. The bill lands in one cost center. The value, if any, lands in another.

JPMorgan's Anderson described middle office employees in Payments spending more on tokens than they earn in salary. That is not yet a scandal and it is not yet a success. It might be the highest leverage spend in the building, or it might be the most wasteful. The company has no instrumentation that tells the difference. There are no leaderboards because there is no agreed model of what good looks like.

Two governance failures collide here. The first is attribution. Companies that already struggled to allocate SaaS costs to business units now have to do the same for tokens, except the consumption is fractal: prompts call prompts call prompts. The second is intent. A token spent on a $50 million decision is cheap. A token spent on a meeting summary is expensive at any price. The line item on the bill looks identical.

The new metric is per-decision cost

The companies that have done this well have stopped reporting tokens per engineer or dollars per team. They report cost per business event. Cost per loan decision. Cost per claim adjudication. Cost per code review that ships. Cost per resolved support ticket. The unit changes by business, the principle does not. You attribute compute to outcomes, not to humans.

Until that metric exists, the token line is just a tax. The CFO sees a number that grows. The board asks what the company gets for it. Nobody can answer because the data is organized around people, not around the work the AI actually does.

Leverage moved

The third break is the one operators care about most. Headcount was a leverage instrument. You added bodies to add capacity. You cut bodies to protect margin. The lever was well understood, slow, and reversible at a known cost.

The compute lever is faster and harder to reverse. Capacity goes up the moment you raise a rate limit. It goes down the moment you cut it. But the human dependence on that capacity, the workflow assumptions, the customer expectations, the downstream products that ship on top of it, those reset more slowly than the rate limit itself. Cutting tokens after engineers have rebuilt their workflows around them is like cutting power to a hospital. The official line drops. The real cost compounds elsewhere.

This is why Microsoft, which moved to cut up to four percent of its workforce in 2026 and announced its first voluntary buyout program in 51 years, explicitly exempted Azure OpenAI, GitHub Copilot, and the Turing research group from the cuts and the hiring freeze. The compute side is the new headcount. Cutting there is treated as cutting into bone.

The architecture problem most companies face

Walk into a typical mid-market company in mid-2026 and you find three things in tension.

The first is a finance team running last year's headcount model on a process that takes four weeks to close the books. The model has not been redesigned for compute volatility. It produces a single annual number that no longer maps to reality.

The second is an engineering team using whatever AI tools they can put on a corporate card, with informal Slack channels recommending Claude Code one week and Cursor the next. There is no central inventory of which prompts, agents, or model versions are in production. There is no kill switch. There is barely a usage report.

The third is a leadership team that has read the same headlines about Meta and Nvidia and JPMorgan and Uber and concluded that AI is an infrastructure decision. They sign a contract with a hyperscaler and assume the cost structure problem will solve itself. It does not.

The gap between these three is where the new operating loss lives. Salary cuts pay for capex the company cannot allocate, on top of token spend it cannot forecast, against outcomes it cannot measure. The numbers are getting bigger and the operating discipline is getting weaker at the same time.

What a usable architecture looks like

The companies that have started to solve this share three structural moves.

They meter every agent call at the source, with attribution to the business event that triggered it. Not the engineer who wrote the prompt. The decision the prompt was made to support. A claims adjudication call is tagged to a claim. A code review call is tagged to a pull request. A pricing call is tagged to a quote. The unit economics fall out of the tagging.

They treat the prompt and agent layer as code under the same review discipline as production software. Changes to prompts go through pull requests. Token-intensive loops are flagged in CI. New agents declare an expected cost per event before they go live. Shadow AI becomes a code review problem instead of a finance problem.

They move the cost owner. The token bill stops landing in IT and starts landing in the P&L of the business owner of the underlying process. The claims VP owns the claims AI bill. The collections leader owns the collections AI bill. When the spend sits with the operator who controls the workflow, the operator builds the instinct to cut what does not pay.

None of this is exotic. It is what good engineering organizations already do with cloud spend, ported to a different layer. The reason it is rare is that it requires architecting the AI stack as a managed cost center on day one, instead of bolting tools onto a workflow that was designed around salaried bodies. By the time a company is paying $12 million in tokens a year, the wiring is buried under three years of point integrations and an army of vendors. The retrofit is expensive.

Don't buy the tool, architect the layer

The pattern that creates real leverage is not a procurement choice. It is an architectural one. Buying Cursor seats or Copilot enterprise or a vendor agent platform without redesigning the cost attribution layer underneath produces exactly the Uber outcome. Adoption climbs, value is real but unmeasured, the bill spikes, and the CTO caps usage at a flat per-seat number that punishes the heaviest leverage users and protects the lowest.

Capping per seat is the worst available response. It treats a flow variable as a stock variable. It assumes every engineer should consume the same token volume, which is roughly equivalent to assuming every salesperson should make the same number of calls. It also discards the actual signal in the spend, which is that some workflows pay back at twenty times the token cost and others do not pay back at all.

The right response is to know which is which. That requires instrumentation, attribution, and a small amount of new accounting discipline. None of it comes from a vendor.

The argument for going first

The hyperscaler capex tells the broader story. Amazon guided to roughly $200 billion in 2026 capital expenditure, mostly AI infrastructure. The combined hyperscaler bill across Amazon, Microsoft, Alphabet, and Meta is between $630 billion and $700 billion this year, close to double 2025. That money is being deployed whether or not your company has a coherent token strategy. The marginal price of compute will fall, then rise, then fall again as supply catches up to demand. The variance will get worse before it gets better.

Companies that survive this transition will be the ones that built the cost attribution layer before the line items got large. Companies that lose will be the ones that ran the old headcount playbook against a flow cost and watched the bill outgrow the salary line they were trying to replace.

The 142,000 tech layoffs tracked by TechTimes through May 29, 2026, with 48 percent explicitly attributed to AI, are a leading indicator. The companies doing the cutting are betting that compute substitutes for salary. The bet only works if the compute is governed at least as well as the salary it replaced. Most of the bet is being made without that governance in place.

Why architecting this beats buying a tool

Every major vendor will tell you their platform solves cost attribution. They are selling the part of the stack they own. Microsoft will sell you Copilot reporting. AWS will sell you Bedrock dashboards. Anthropic and OpenAI ship usage analytics. Cursor and Claude Code emit telemetry. None of these tools see each other. None of them attribute to your specific business events.

The integration layer that makes the cost numbers usable for your operation does not come in a SKU. It is built. It runs against your specific prompts, your specific agents, your specific business processes. It is owned by you because nobody else has the context to build it well. And it pays back the moment the first executive asks "what is the cost per loan decision after AI?" and gets a real number back.

That layer is what we build at Agor AI Advisory. We treat the AI stack as a managed cost center from the first prompt forward. We instrument agent calls at the source. We attribute tokens to business events, not to people. We give finance a forecast that survives a new model release on a Tuesday. We give the operator the kill switch and the leaderboard at the same time. We do this once, properly, so the bill never outruns the value the way it did at Uber, and so the cuts never outrun the capacity the way they will at the next company that swings the salary axe blindly.

The companies winning in 2026 are the ones that know, every week, what each token paid for. The companies losing are the ones that cut the salary line because the token line was easier to ignore.

Schedule a strategic consultation with us today.

Sources

Want this kind of automation working for your business?

Agor AI designs and ships the systems these posts describe, scoped in weeks, not quarters.

Book a Free Strategy Call