Uber built its 2026 AI budget the way every company builds a software budget: headcount times license price, plus a cushion. It was gone in four months.
Internal Claude Code adoption jumped from 32% of engineers to 84%, and the bill ran $500 to $2,000 per engineer per month. Nobody misused anything. The tool worked, people used it more, and the budget model broke on contact. The budget was not wrong. The unit was.
That unit, the seat, has quietly run enterprise software economics for twenty years. It is dying now, and it is worth understanding why, because the answer explains what your next three renewal negotiations are going to look like.
Why the seat existed at all
Per-seat pricing was never a law of nature. It was an artifact of a specific economic fact: software had zero marginal cost.
Once the code was written, one more user cost the vendor approximately nothing. There was no meter worth reading, so price attached itself to the only number that scaled with value: people. Seats were easy to count, easy to budget, easy to true-up at renewal. Finance teams built twenty years of planning muscle on that assumption. Headcount goes up, software spend goes up, in a straight and predictable line.
The entire apparatus of corporate software, procurement templates, budget cycles, vendor sales compensation, was built downstream of one premise: usage is free.
AI broke the premise
Every prompt burns compute. Every agent run consumes tokens, and tokens are a real cost of goods sold, paid by the vendor to a cloud provider or a chip fab in actual dollars.
In budget terms it shows up simply: the team that ran 500 prompts a week now runs 10,000, and the line item tracks that volume, not the headcount.
For the first time since software came in shrink-wrap, a heavy user costs the vendor real money. Under a flat seat price, that inverts the whole business: the vendor loses the most on its best customers. The engineer running an agent all day is a margin problem. The team that finally adopted the tool everyone hoped they would adopt is a margin catastrophe.
No pricing model survives losing money on your most enthusiastic users. So the model is changing, in public, on a schedule you can watch.
The migration is already receipted
On June 1, GitHub moved Copilot to usage-based AI Credits. That is developer tools moving first: the vendor category sitting closest to the compute bill blinked before anyone else.
On July 8, the model vendors followed. Anthropic moved its top model to usage credits: $10 per million input tokens, $50 per million output. The detail that matters is the phrase on top of existing subscriptions. The seat did not go away. It became a cover charge. The meal is metered.
Enterprise applications are next in line. Salesforce has spent months repricing core clouds around agent bundles that push effective costs up double digits. We called the shift in Issue #39 when Copilot's usage billing first surfaced; by Issue #41 the confirmations were arriving weekly.
The future is not usage instead of seats. It is seats plus a meter.
None of this is vendors getting greedy. It is vendors getting invoiced. The economics moved first, and the price sheets are catching up.
Get 10 AI workflows + the weekly briefing. Free.
Subscribe FreeWho gets squeezed
The squeeze does not land on the vendor. It lands on the person whose planning model assumed seats.
Seats were predictable. A meter is not. Finance teams that could forecast software spend to within a rounding error are now looking at line items that can double in a quarter because a team found a workflow that works. That is the perverse part: the variance is a symptom of success. Adoption is the thing every AI budget was built to encourage, and adoption is now the thing that detonates the budget.
The chatbot-era budget, a few subscriptions, a pilot, a cushion, is meeting agent-era consumption. Agents do not take lunch. The gap between those two curves is going to surface in upcoming planning cycles for the next fiscal year, and it will land as a surprise in exactly the organizations that treated AI spend as a per-seat line item.
The Budget Math playbook
Three moves, in order of urgency.
1. Put caps and alerts in before adoption spikes, not after. Every major AI platform now ships spend controls. If your tools are metered and your alerts are not configured, your budget protection is hope. A simple starting pattern: alert at 50 percent of the monthly cap, review at 80. The time to set the tripwire is while the line item is still boring.
2. Route work by cost tier. The most expensive model should not be the default for everything. In practice: drafts, summaries, and internal tickets run on the inexpensive tier; contract review and customer-facing work get the frontier model. Teams that route by task are reporting materially lower bills for the same output. Teams that default to the flagship are funding the vendor's margin recovery.
3. Renegotiate at renewal assuming the meter is coming. If your vendor has not moved to usage pricing yet, the renewal is where they will. Walk in with your own consumption data, ask what the effective per-user cost looks like at your actual usage curve, and get a written overage schedule before you sign. The vendors have done this math. Most buyers have not.
Watch the meter
Per-seat SaaS is not dying because vendors changed their minds. It is dying because usage finally costs something, and prices always find their way back to costs.
The seat count on your renewal quote is becoming the least interesting number on the page. The meter is the strategy. Watch the meter.
We read 47 AI and tech newsletters so you don't have to. getaiminute.com