The Great Token Heist: How Capitalism Got Got

That is the comic horror at the center of the Great Token Heist. The boardroom thought artificial intelligence was going to be a labor-saving device, a miracle spreadsheet, a thousand interns in a trench coat. A gentle, glowing rectangle that would write the emails, summarize the meetings, generate the decks, debug the code, classify the tickets, rewrite the rewrites, and generally turn every employee into a tiny sovereign nation of productivity. Instead, the glowing rectangle came with a taxi cab-style meter.

Not a normal meter, either. Not a clean, honest utility meter where you know roughly what the light switch costs. This meter counts tokens. It counts input tokens, output tokens, cached tokens, context tokens, agent tokens, tool-use tokens, web-search tokens, retry tokens, whoops-we-forgot-to-set-a-budget tokens, and the most dangerous category of all: tokens generated by a corporate culture that cannot tell the difference between useful work and visible activity. And now the bill is arriving. Loudly. With teeth.

The Reported Half-Billion-Dollar Oops

The inciting incident is almost too perfect to improve with satire. In May 2026, Axios reported that an unnamed company, according to an AI consultant, recently spent half a billion dollars in a single month after failing to put usage limits on Claude licenses for employees. Tom's Hardware picked up the story with the appropriate level of stunned disbelief: a mystery company, one month, $500 million, and apparently no one standing near the big red spending lever with a clipboard.

Maybe the story is exactly as reported. Maybe the number is rounded, messy, negotiated, enterprise-discounted, or wrapped in some internal accounting fog. Maybe the company will remain unnamed forever, hiding in the tall grass with the other wounded megafauna of the AI boom. The precise identity almost does not matter. The story works because everyone who has watched corporate technology adoption for more than seven minutes can recognize the shape of the disaster.

Someone bought the future in bulk. Someone skipped the usage controls. Someone equated adoption with value. Someone made the tool available before they made the governance real. Someone assumed the cost curve would behave because the slide deck behaved. Then the meter spun like a raccoon had gotten into the walls.

The Genius of the Token

A token is a small unit of text. That is the boring definition. The better definition is that a token is the perfect capitalist particle: tiny, measurable, billable, abstract enough to hide inside infrastructure, and numerous enough to become terrifying at scale.

Capitalism loves units > hours > clicks > impressions > seats > licenses > monthly active users > billable increments > pages viewed > minutes watched > steps taken > calories burned > miles driven. Everything that can be sliced becomes a product, everything that can be counted becomes a metric and everything single damn thing that can be metered becomes someone else's revenue model.Every. Damn. Thing.

So naturally, the newest version of work arrived not as a tool but as a toll road. Every question costs. Every answer costs. Every clarification costs. Every automated step costs. Every agent that thinks, checks, calls another tool, reads the output, decides it is not quite right, tries again, writes a summary, rewrites the summary, and then apologizes for the confusion costs. Everytime you need to correct its sloppy behavior it costs you.

The token is brilliant because it turns cognition into a consumable. It takes the messy human act of asking, thinking, drafting, comparing, searching, revising, and deciding, then chops it into grains small enough that no single grain feels expensive. That is how the heist works. Nobody panics at one token. Nobody panics at a handful of tokens. Nobody panics at a team using tokens. Then one day Finance opens the invoice and discovers that the office has been burning thought-pellets by the boxcar. This is literally Superman 3 style scaping fractions of pennies that compound over time. But, they aren’t scraping pennies, it's a much larger unit.

Usage Is Not Value

The first corporate mistake was believing that AI usage is the same as AI value and it most definitely is not. Usage is a signal that people have access but it is not proof that anything important got better.

This is where the whole thing gets deliciously absurd. For years, executives have begged employees to adopt new systems. Use the CRM. Use the project management tool. Use the knowledge base. Use the portal. Use the dashboard. Use the dashboard that summarizes the dashboard. The assumption was simple: if usage goes up, transformation is working.

AI broke that assumption because AI is very easy to use in ways that feel productive while quietly generating nothing of durable value. It can summarize a meeting that should not have happened. It can rewrite a memo that should have been three bullet points. It can generate five strategic directions for a project that needs one decision. It can create a roadmap for work nobody has agreed to fund. It can produce an executive summary of an executive summary of a deck that exists because nobody wanted to have the hard conversation live.

A company can burn millions of tokens proving that its internal processes are bloated, unclear, duplicative, over-permissioned, emotionally avoidant, and allergic to decision-making. The model did not create that condition. It merely found the buffet.

The Agent Problem: Infinite Interns with Company Cards

The second mistake was believing that agents would make this cheaper. Agents sound efficient because they promise autonomy. Give the agent a goal and it will go do the thing. Wonderful. Magical. Very futuristic. Also potentially like giving a bright, nervous intern a corporate card, a browser, a vague assignment, no spending cap, and the emotional constitution of a Roomba in a hallway full of chairs.

A simple chatbot interaction is already a meter. An agent is a meter with legs. Agents do not merely answer. They plan. They search. They inspect files. They call tools. They read outputs. They revise plans. They ask themselves if they have enough context. They add more context. They make a mistake. They correct the mistake. They generate logs. They produce a final answer. Then a human says, 'Can you make it more concise?' and the little furnace starts again. This is useful when the task is worth it. It is rather ridiculous when the task is corporate confetti.

A great AI agent attached to a valuable workflow can save real time. A great AI agent attached to a bad workflow can automate the badness at industrial speed. It will not fix the fact that seven departments all own one field in a spreadsheet. It will not solve a governance model where everyone can comment and no one can decide. It will not repair a culture that confuses escalation with leadership and documentation with alignment. It will simply convert that confusion into billable computation. The machine is not eating jobs first. It is eating ambiguity. And ambiguity, in corporate America, is an all-you-can-eat buffet.

Capitalism Thought It Was the House

The funniest part is that capitalism thought it was the house. Corporations looked at AI and saw leverage. They saw reduced headcount, faster output, cheaper operations, fewer handoffs, fewer meetings, fewer contractors, fewer writers, fewer analysts, fewer developers, fewer people asking annoying questions about whether the strategy made sense. The dream was not subtle. It was the oldest dream in management: keep the output, reduce the humans, call the difference efficiency.

But the AI stack has landlords all the way down. The model provider wants revenue. The cloud provider wants revenue. The chipmaker wants revenue. The data center wants power. The enterprise vendor wants margin. The integration partner wants services. The security layer wants a seat. The monitoring layer wants a seat. The governance platform wants a seat. The cost-optimization startup wants a seat because, of course, once the invoice gets insane, a second market appears to help you understand why the first market ate your lunch.

The company trying to save money by replacing labor does not escape cost. It changes landlords.

This is precisely how capitalism got got. It built a system where every inefficiency becomes a product, every product becomes a subscription, every subscription becomes a platform, every platform becomes infrastructure, and every infrastructure layer gets to charge rent. Then AI arrived and did the funniest possible thing: it made the invisible thought-work of the organization meterable.

The Token Heist Is Also a Mirror

The word heist implies a criminal mastermind. Perhaps, a crew in black turtlenecks and van idling in the alley. Lasers. Blueprints. Someone saying, 'We have ninety seconds.' Then running down the street looking for the van that wasn’t there. But the Great Token Heist is not that cinematic. Nobody had to break in because the doors were opened from the inside.

The model companies did not have to sneak into the enterprise. The enterprise invited them in with procurement paperwork, executive sponsorship, innovation theater, and a sincere belief that the next quarter could be improved by adding a shimmering layer of autocomplete to everything. That does not mean the providers are innocent woodland creatures. They know the beauty of usage-based pricing. They know the magic of making the first experiments feel cheap. They know that once a tool becomes part of the workflow, the switching cost becomes a moat. They know that employee enthusiasm, executive fear, and competitive panic are a powerful sales team.

But the deeper indictment belongs to the buyer. The buyer wanted the machine to make judgment unnecessary. The buyer wanted transformation without redesign. The buyer wanted cost savings without the political pain of saying what work should stop. The buyer wanted innovation without governance, speed without discipline, and adoption without a theory of value.

The Weather Query That Explains Everything

One of the most revealing details in the broader reporting around enterprise AI costs is almost comically small: employees using AI models to check the weather.

On one level, who cares? People waste tiny amounts of company money every day. They print documents nobody reads. They attend meetings where the agenda is vibes. They spend twenty minutes choosing the right reaction meme. A weather query is not the apocalypse.

But symbolically, it is perfect. Checking the weather with an enterprise AI model is the whole AI bubble in miniature. It takes a solved problem, routes it through an expensive abstraction layer, consumes resources, produces a worse version of a free answer, and then gets counted as adoption.

That is the danger. Not that employees are foolish. Employees are rational within the incentives they are given. If leadership says use AI, people will use AI. If leadership praises usage, people will create usage. If leadership launches leaderboards, people will climb them. If leadership asks for transformation but cannot define the transformed state, people will produce artifacts that look like transformation. The weather query is not the scandal. The scandal is the measurement system that could confuse it for progress.

The Productivity Trap

AI can absolutely make people more productive and that is the uncomfortable part for both the true believers and the reflexive skeptics. The tool is not fake. It can draft, summarize, classify, translate, brainstorm, debug, compare, generate, and explain. In the hands of a skilled person with a clear goal, it can be astonishing.

But productivity is not the same as output volume. A worker who creates ten more documents is not automatically ten times more useful. A team that ships twice as many drafts is not necessarily closer to a decision. A department that generates beautiful slides at machine speed may simply be producing a more elegant fog. It’s like the whole million monkeys at a typewriter thing.

The productivity trap is the belief that because AI lowers the friction of producing something, the something must be worth producing. That is how organizations drown themselves in plausible material. More options. More summaries. More variants. More analysis. More automated follow-ups. More synthetic confidence. More words around the same unresolved center.

The real scarce resource in modern work is not text. It is judgment. It is taste. It is prioritization. It is the ability to say: this matters, this does not, this can wait, this is a decision, this is noise, and this meeting should have been a paragraph. AI can assist judgment. It cannot replace the need for it. When companies forget that, tokens become a tax on indecision.

The Coming Cost Council

One predictable outcome of the Token Heist is the rise of a new C-Suite position: AI FinOps. Dashboards will bloom, governance councils will convene, consultants will arrive with frameworks. There will be token budgets, prompt hygiene guidelines, model routing policies, cache strategies, chargeback models, approved-use matrices, and at least one executive steering committee with a name so bland it could be used as packing material. Some of that will be necessary. Cost control matters. Model choice matters. Caching matters. Guardrails matter. Usage limits matter, as the mystery half-billion-dollar bonfire rather dramatically suggests.

But there is a risk that companies will respond to the wrong problem. They will optimize the meter while preserving the madness that feeds it. They will reduce token spend on a bloated process instead of asking why the bloated process exists. They will route low-value work to cheaper models instead of eliminating low-value work. They will create policies for summarizing unnecessary meetings instead of canceling unnecessary meetings.

That would be the most corporate outcome imaginable: building an entire governance apparatus to make waste more cost-efficient.

The right question is not merely, 'How do we spend fewer tokens?' The right question is, 'What work deserves machine assistance, what work deserves human judgment, and what work deserves the trash can of accountability?'

Will the Token Heist Swallow Capitalism?

Will the Token Heist swallow capitalism? Not in one dramatic gulp. Capitalism is too flexible, too shameless, and too good at turning its own injuries into new markets. It will not look at runaway AI costs and walk quietly into the sea. It will create budget tools, token auditors, cheaper models, private deployments, efficiency benchmarks, procurement rituals, and a fresh vocabulary for pretending the lesson has been learned.

But the heist exposes a structural problem that will not go away. AI turns more of work into metered infrastructure. The more companies rely on AI to draft, analyze, monitor, code, support, sell, summarize, and decide, the more their operating costs become tied to computational consumption. That consumption is not always predictable. It is affected by user behavior, model architecture, workflow design, governance quality, context size, retry loops, tool calls, data access, and the endless corporate appetite for generating one more version.

In other words, capitalism may have created a new cost layer that grows in direct proportion to its own inability to simplify itself. That is the swallowing motion. Not collapse. Enclosure. Every process wrapped in AI. Every AI call wrapped in tokens. Every token wrapped in billing. Every bill wrapped in optimization. Every optimization wrapped in another platform. The system consumes itself by metering the work it once hoped to eliminate.

The Smarter Way Forward

This is not an argument against AI. It is an argument against treating AI like a magic solvent for organizational stupidity.

Use it. Absolutely use it. Use it where the workflow is clear, the output is valuable, the human remains accountable, and the cost is visible. Use it to reduce drudgery, accelerate drafts, improve accessibility, test ideas, find patterns, explain complexity, and help people move from blank page to useful first pass. Use it like a power tool, not a religion.

But count outcomes, not vibes. Measure cycle time, defect reduction, customer value, revenue impact, employee capacity, decision quality, and work avoided. Do not celebrate tokens consumed. Do not reward people for feeding the furnace. Do not confuse prompt volume with transformation. Do not buy a thousand licenses and call that a strategy. And for the love of every budget owner who has ever aged three years in one invoice cycle, set the usage limits.

The Great Token Heist is funny because it is absurd. It is alarming because it is plausible. And it is instructive because it reveals a truth companies keep trying not to learn: tools do not save you from management. They amplify management. Good management gets leverage. Bad management gets a larger blast radius.

Capitalism thought AI would help it squeeze more value out of labor. Instead, AI revealed how much of corporate labor was already a ritual of confusion, delay, documentation, and performance. The meter did not create the monster. It gave the monster a price per million tokens.

And somewhere, deep in the enterprise stack, the taxi meter is still running.

Closing: The House Always Wins Until It Does Not

The house always wins because the house controls the game. That was the old wisdom. But the Great Token Heist suggests a stranger possibility: the house can win so hard that it builds a game it no longer understands.

The corporation wanted a machine that could think cheaply. It got a machine that made thinking billable. It wanted to reduce dependence on people. It became dependent on vendors, models, clouds, chips, data centers, consultants, governance tools, and the fragile hope that the next billing cycle would be less humiliating than the last.

That is how capitalism got got. Not by communists. Not by poets. Not by some raccoon in the dumpster behind the data center, although I would not rule out his involvement. It got got by its own favorite trick: take something human, abstract it, meter it, scale it, financialize it, and then act surprised when the meter becomes the product.

The Token Heist is not coming. It is already here. It is in the dashboards, the usage reports, the agent logs, the CFO panic, the canceled licenses, the suddenly serious meetings about governance. It is in every organization that believed intelligence could be purchased by the scoop and poured over a broken process until the process became smart.

But intelligence was never the scarce ingredient. Discipline was. Judgment was. Clarity was. Courage was. The tokens were just the receipt.