The Token Reckoning: Five Predictions for How AI Spending Breaks the Corporate World
Someone predicted eight days ago that companies would crack down on runaway AI token costs. Then Sam Altman went on stage and confirmed it. That kind of turnaround earns the right to make five more predictions. Here they are, ranked from heartwarming to genuinely unhinged.
The Confirmation That Started All of This
Altman said at OpenAI's Intelligence at Work event that token costs went from an issue that never came up at the start of 2026 to "all of a sudden, a huge issue." The meme spreading through enterprise hallways: "My company spent the entire 2026 budget in Q1. Can you make this more efficient?"
The numbers behind that panic are staggering. Six and a half years ago, OpenAI's heaviest user consumed 100,000 tokens per month. That figure is now the global per capita average. The current top user inside OpenAI burns 100 billion tokens a month. Altman admitted, to his own embarrassment, someone outside the company uses even more.
Uber's COO Andrew Macdonald put it plainly: despite 95 percent of Uber engineers using AI tools monthly and 70 percent of committed code being AI-generated, he cannot draw a clear line between that spend and improvements in consumer-facing products. "That link is not there yet." This followed the Uber CTO's earlier disclosure that the company burned through its entire 2026 Claude Code and Cursor budget by April.
Prediction 1: Open Source Token Donations
The nicest one. Companies sitting on unused token budgets will begin donating them to open source projects to cover the cost of running automated tests, CI pipelines, and other tasks that have quietly become expensive as AI tooling embeds itself into every workflow.
Think of it like the old model of donating spare compute cycles for protein folding research, except the currency is inference instead of CPU time. This one feels close. It requires no regulatory change, no new product, just a policy decision from labs and a matching mechanism. Expect it within months.
Prediction 2: Token Stipends as a Standard Benefit
This one is already partially real. Jensen Huang proposed at GTC 2026 that engineers should receive token budgets worth roughly half their base salary on top of their pay. His framing: if a $500,000 engineer is not spending at least $250,000 on tokens annually, something is wrong. He described token allocation as an emerging recruiting tool in Silicon Valley: "How many tokens come along with my job?"
The broader prediction is that yearly token stipends become a fourth standard employment benefit alongside salary, equity, and health coverage. Engineers who come in under budget at year end keep a share of the savings. Good behaviour gets incentivised. Token waste stops being the default.
Prediction 3: Token Poker
Remember planning poker from agile? The ritual where teams held up cards estimating story points for a task, debated the outliers, and landed on a consensus? That process is coming back, rebranded and retooled for a world where the cost of a feature is measured in millions of tokens rather than developer-days.
Before a long-running agent task is kicked off, teams will sit down to estimate the token cost. Someone will say ten million tokens. Someone else will say fifty million because it is urgent and complexity is being underestimated. The team will negotiate. A prompt will be workshopped collaboratively before anyone hits run.
Consulting firms will package this as a methodology. Workshops will be sold. Certification programmes will follow. The cycle continues.
Prediction 4: Org-Wide Token Budgets and the Management Class They Create
This is where things start to rot. Rather than a single company-wide AI budget, organisations will cascade token allocations down through divisions, teams, and individuals. And then the consequences arrive.
One person on a team burns through the quarterly allocation in a week running agents on tasks nobody asked for. The rest of the team goes back to writing code by hand. Resentment follows. Meanwhile, a new layer of middle management emerges whose entire job is negotiating token limits upward with whoever controls the master budget.
The knock-on effect will be pair prompting: two engineers huddled over a single prompt before a multi-day agent run, arguing over context window strategy and whether threatening the AI with a bad performance review actually helps. Long-running agent tasks will get treated like production deploys: reviewed in pull requests, debated in comments, scrutinised for efficiency before anyone commits to the spend.
Prediction 5: Lines of Code as a Proxy for Token Budget Allocation
The most destructive prediction, and the one most likely to already be happening somewhere.
Management will pull the git logs, identify who produced the most code last month, and allocate more token budget to those people. The engineer who shipped 100,000 lines gets more. The one who shipped 10,000 gets less. Output volume becomes the metric that unlocks resources.
ClickUp already pointed the way. The company cut 22 percent of its workforce in May 2026, deployed 3,000 internal AI agents, and introduced salary bands reaching $1 million per year for employees who produce what CEO Zeb Evans calls "100x impact." The model: find the people generating the most output and give them effectively unlimited resources. Everyone else becomes redundant.
The problem, as George Hotz argued in his blog post "The Eternal Sloptember", is that volume and quality are not the same thing. High performers can spot bad agent output. Weaker developers often cannot, and they are the ones now producing ten times the volume. Rewarding the biggest output numbers with the biggest token budgets accelerates the production of what Hotz calls "buckets and buckets of slop." At some point, a company will bury itself under the weight of it.
Where This Is Heading
The token cost crisis is a corporate governance problem that the industry was not built to handle. Companies gave every employee access to tools with no spending limits, no accountability mechanisms, and no way to measure whether the output justified the bill. Altman himself acknowledged that OpenAI is now actively trying to help customers get more value for less spend.
What comes next is not the end of AI in enterprise software. It is the arrival of the same budget scrutiny that applies to every other line item. Someone will have to justify the spend. Someone will have to own the number. And somewhere, a company that skipped that step and just kept the agents running is going to fail in a way that makes a very instructive case study.