For the last 1-2 years, corporate AI adoption has operated with the financial discipline of a teenager holding a stolen Amex at a 7-Eleven.
Use more AI. Use it everywhere. Put it in every workflow. Ask every employee how they are using it. Add it to performance reviews. Shame the skeptics. Celebrate the “AI-native” power users. Let agents run overnight. Feed entire repositories into coding tools.
Welcome to the age of tokenmaxxing: the corporate habit of treating AI usage itself as a sign of productivity.
The first phase was predictable. Companies saw generative AI as a new operating system for work. Early adopters used it to write code faster, analyze documents, build dashboards, automate support, draft emails, and generally compress the time between “I have an idea” and “something exists.” The upside was real. The enthusiasm was earned. Adoration even.
But the second phase is arriving now: the bill.
And unlike a SaaS seat license, token costs do not politely sit still at $30 per user per month. Agentic AI can consume tokens like a coal furnace. Every step, retry, tool call, context window, chain-of-thought-like intermediate operation, and “just one more refinement” is metered. The more autonomous the system becomes, the more compute it burns. The more context it has, the more expensive each turn becomes. The more employees are encouraged to “use AI,” the more they learn to spend invisible money.
This is not a morality play. It is a unit economics problem.
A token is not magic. It is a billing unit. It is the new cloud compute meter wearing a productivity costume. And many companies have rolled it out before they understand the labor/output economics.
The corporate mandate often sounds like this: “Everyone should be using AI.” Reasonable enough. But the next question is usually missing: “For what business outcome, at what cost, compared to what alternative?”
That missing sentence is where the reckoning lives.
Consider the most extreme public example so far. Peter Steinberger, creator of the open-source AI project OpenClaw and an OpenAI employee, reportedly burned through about $1.3 million in OpenAI API tokens in 30 days, covering roughly 603 billion tokens, 7.6 million requests, and about 100 autonomous coding agents. The bill was reportedly covered by OpenAI, not Steinberger personally. Still, as a real-world data point, it is astonishing. One project. One month. Seven figures of inference. (source)
Now, OpenClaw may be a legitimate frontier experiment. That is the point. The question is not whether the spend was “bad.” The question is whether the output justified the cost. Did it generate more than $1.3 million of enterprise value? Did it create reusable infrastructure? Did it produce a step-change in engineering velocity? Maybe yes.
But if the same spend happened inside a normal enterprise IT department to build a dashboard that six people look at twice a quarter, the CFO would be legally required to throw a chair.
The weirder part is that some executives have started to frame token consumption almost like a professional obligation. Nvidia CEO Jensen Huang has reportedly argued that highly paid engineers should be consuming AI tokens at the scale of a meaningful share of their annual compensation — even half of salary in some commentary. (source)
Intellectually, the argument is not crazy. If a $500,000 engineer can spend $100,000 of compute to produce $2 million of output, please proceed. Buy the tokens. Buy more tokens.
But notice the assumption: output.
The danger is treating token burn as the proxy for output. This incentivizes employees start optimizing for visible AI usage instead of business results. Amazon recently shut down an internal employee-made AI token usage leaderboard called KiroRank after it reportedly encouraged some staff to perform tasks that did not necessarily solve problems, merely to climb the rankings. A senior Amazon executive told employees, essentially: do not use AI just to use AI; use it to solve customer and business problems. (source)
That sentence should be engraved above every enterprise AI budget request.
Because the cost curve is about to get very real. A normal chatbot interaction may be cheap. A coding agent operating for hours with repository context, tools, retries, tests, browser access, memory, and long-context reasoning is different. Recent research on agentic coding tasks found that agentic workflows can consume orders of magnitude more tokens than simple code chat, and that higher token usage does not reliably translate into higher accuracy. In some cases, accuracy peaks at intermediate cost and then saturates. Translation: the agent may be thinking harder, but not necessarily better. (source)
This is where companies may become schizophrenic.
On Monday, they will mandate AI adoption.
On Tuesday, they will receive the inference bill.
On Wednesday, they will ask why headcount has not gone down.
On Thursday, someone will produce a 47-slide deck titled “AI ROI Governance Framework.”
On Friday, procurement will block Claude.
The reckoning will not be philosophical. It will be financial. Companies will ask a brutally simple question: if we spent $10 million on AI tools, inference, and internal enablement, where is the $10 million of labor savings, revenue growth, risk reduction, or cycle-time compression?
This is especially awkward because AI is being sold, explicitly or implicitly, against labor. The economic pitch is not merely “better software.” It is “same output with fewer people” or “more output with the same people.” That means every token bill will eventually be compared to FTE cost.
If an employee spends $10,000 in Claude Sonnet tokens to build a dashboard, is that good or bad?
The correct answer is annoying: it depends.
If the dashboard replaces three weeks of analyst work every quarter, helps executives detect churn risk earlier, and saves a $1 million customer, it is the cheapest $10,000 the company ever spent. If the dashboard is a prettier version of an existing Zendesk report and becomes digital wallpaper, it is an expensive arts-and-crafts project.
AI spend should be evaluated like any other capital allocation decision: cost, output, durability, reuse, and business impact.
Instead, too many companies are still in the “look, the robot made a deck” phase.
Meanwhile, some companies are pre-empting the economics with headcount reductions. Meta has large layoffs linked to an AI-driven efficiency push while investing heavily in AI infrastructure. (source) Block cut roughly 4,000 employees — nearly half the company — while arguing that a smaller team using AI tools could do more and do it better. (source) Salesforce has also cut roles while simultaneously hiring around AI products, a pattern that reflects the new corporate posture: reduce legacy labor, fund AI distribution, and hope the productivity bridge holds. (source)
This is the uncomfortable pre-reckoning. Companies are paying for both the old operating model and the new one. They still have the people. They now also have the AI bill. The promised efficiency has not always translated into lower operating expense because workflows, incentives, controls, and management systems have not caught up. But they must.
And yet, dismissing AI as overhyped would be lazy. There are real productivity gains, and some are enormous.
In customer support, an NBER study found that access to a generative AI assistant increased productivity by about 14% overall, with the largest gains — around 35% — for less experienced and lower-skill workers. (source) That is not vapor. That is labor leverage.
In consulting-style knowledge work, a Harvard/BCG study found that consultants using GPT-4 completed 12.2% more tasks, worked 25.1% faster, and produced higher-quality results on many tasks inside the model’s capability frontier. But the same study also found performance could worsen when users applied AI outside that frontier. (source) Again, not magic. Leverage with boundaries.
In software development, GitHub reported that developers using Copilot completed a controlled coding task 55% faster than those without it. (source) Even if later studies complicate the picture in real-world repositories, the core lesson stands: for certain tasks, especially boilerplate and well-bounded coding work, AI can produce dramatic speed gains.
Klarna, in Pied Piper fashion, fluted its intoxicating case study: its AI-powered assistant handled 2.3 million conversations in its first month, managed about two-thirds of customer service chats, performed work equivalent to 700 full-time agents, reduced repeat inquiries by 25%, cut resolution time from 11 minutes to under 2 minutes, and was estimated to drive $40 million in profit improvement. (source)
That is what good looks like. Not “we used a lot of AI.” Not “our token usage is up 400%.” Not “employees are really leaning in.” Good looks like throughput, cycle time, customer experience, margin, risk reduction, and revenue.
The next phase of AI adoption will belong to companies that stop asking, “How do we get employees to use AI?” and start asking, “Which workflows deserve expensive intelligence?”
That distinction matters. The right tool for the right job.
A $20 chatbot answer does not need a frontier model with a million-token context window. A legal-risk review of a major contract might. A customer support macro can probably use a cheaper model. A complex engineering migration may justify a premium agent. A board presentation can tolerate a few dollars of inference. A production finance workflow needs accuracy, auditability, and cost controls. Take a look at our portfolio company TrueFoundry if you’re keen to optimize your AI gateway and reduce costs.
The future is not “AI everywhere.” That is toddler strategy. The future is model routing, token budgets, ROI instrumentation, and workflow-level accountability.
Every enterprise—yes, even startups—should know five numbers:
AI spend by team.
AI spend by workflow.
Cost per completed output.
Human time displaced or revenue/risk impact created.
Reuse rate of AI-generated assets.
Without those numbers, companies are not running an AI strategy. They are running an expensive vibes program.
The great tokenmaxxing reckoning will not kill AI adoption. It will professionalize it. The early period rewarded enthusiasm. The next period will reward disciplined outcomes.
There will still be room for frontier experiments. There should be. Some people should be allowed to burn tokens like rocket fuel. But most corporate AI work is not a rocketry. It is creativity, delivery, and workflow. And workflows need unit economics.
The winners will not be the companies with the highest token usage. They will be the companies with the clearest conversion of tokens into business outcomes.
The losers will be the ones who confuse motion for leverage, usage for productivity, and inference spend for innovation. Meta’s rocking horse metaphor applies.
AI is powerful. Tokens are useful. Agents may be transformational.
But the bill is coming, and this time the CFO is going to read it.