No Backsies. You Told Us to Burn More Tokens or Lose Our Jobs.

They ordered the survivors to torch tokens to prove “adoption,” then capped them when the bill came. The executives giving the orders can barely use the tool.

Six months ago, the fastest way to look like a serious engineer was to set a number on fire. Not ship better software — burn more tokens. The metric that mattered wasn’t whether the code held up in production; it was consumption, how many billions of tokens your agents could devour while you looked busy. Finance teams cheered the line going up. Managers built leaderboards. Jensen Huang reportedly decided a $500,000-a-year Nvidia engineer should be torching at least $250,000 in tokens a year, as if combustion were the same thing as output. And the message to everyone below the executive floor was unambiguous: feed the machine, or the machine feeds without you.

Now read this week’s news. Uber blew through its entire annual AI budget in four months and capped employees at $1,500 a month. Amazon and Meta quietly deleted the leaderboards they’d spent the spring celebrating. Lindy, an AI startup run by Flo Crivello, ripped Claude out and moved all of its traffic to DeepSeek, watching the cost curve collapse to save millions. CNBC’s Friday piece has a tidy label for the new mood: a shift from tokenmaxxing to efficiency.

Cute. Except a lot of people don’t get to shift. They got fired on the way up so the buildout could be funded, the survivors got whipped to feed the machine that replaced their coworkers, and now they’re being told the machine costs too much and could they please use it less. This piece is longer than the last one because the story is bigger than one reversal. It’s about who gave these orders and whether they had the faintest idea what they were talking about, what this technology was supposed to be for, what it’s actually being used for, and the financial sinkhole underneath the whole performance. There are no backsies on any of it. You told us to burn more tokens or lose our jobs. We remember. Let’s go through it properly.

The mandate was never optional

Let’s not pretend tokenmaxxing was a quirky meme engineers cooked up for fun. It was policy, enforced from above, with your performance review stapled to it.

JPMorgan Chase, per Economic Times reporting, told its tens of thousands of software engineers in writing that AI usage would factor into their performance ratings. Use the tools, or it counts against you. Amazon tracked whether developers hit its internal benchmark of 80% weekly AI usage and ranked them on leaderboards by consumption; the company said the stats weren’t used in reviews, but the Financial Times reported employees spinning up agents to run pointless busywork purely to keep their numbers high — which tells you precisely how much they believed the disclaimer. Meta factored AI usage into reviews outright. The figures this produced weren’t signs of productivity; they were signs of fear. One Disney employee pinged Claude 460,000 times in nine days. One OpenAI engineer burned 210 billion tokens in a single week, enough output to fill Wikipedia thirty-three times over. One anonymous worker ran up more than $150,000 a month in token charges. Token budgets even started showing up as a line of compensation, next to equity and bonus — the company had decided your worth was partly a function of how much of its own compute you could set on fire.

A New York developer named Dina gave The Guardian the version from the floor. Her job used to be writing code; now it’s mostly cleaning up after a hallucinating tool that hands her flawed output she has to comb through line by line. It wasn’t making her faster. But the message from management never changed: go faster, this will make you go faster, speed is the number one priority. That’s the whole arrangement in one breath. The tool didn’t have to work. You just had to be seen feeding it.

The men giving the orders don’t know what a token is

Here is the part nobody wants to say in a town hall. A staggering share of the executives who mandated all of this cannot actually use the technology they bet the company on.

Start with the numbers. The National Bureau of Economic Research surveyed nearly 6,000 CEOs, CFOs, and senior executives across the US, UK, Germany, and Australia, and roughly 90% of firms reported that AI had made no measurable difference to productivity or employment. The executives who said they personally used AI averaged about an hour and a half a week with it — less time than most people spend doomscrolling before lunch — and a quarter of them didn’t use it at all. PwC, surveying 4,454 CEOs across 95 countries for its Davos report, found that 56% of companies were getting essentially nothing out of AI, and its own global chairman blamed executives for forgetting the basics. And MIT’s NANDA initiative, after interviewing leaders and analyzing 300 deployments, landed on the number that should have ended the leaderboard era by itself: 95% of enterprise generative-AI pilots delivered no measurable impact on the bottom line, against $30 to $40 billion in enterprise spend. A manufacturing executive told the researchers the quiet truth — the hype says everything has changed, but in operations, “nothing fundamental has shifted.”

Now layer the competence gap on top of the usage gap. Grant Thornton found that 78% of executives lacked confidence they could pass an independent AI-governance audit within ninety days — meaning they had deployed systems they couldn’t explain, with accountability they couldn’t locate. These are the people who decided your job could be done by an agent, and who then decided you should be graded on how hard you fed it.

Box’s CEO, Aaron Levie — an AI bull, not a hater, a man who angel-invests in this stuff — put a clinical name to the condition on X. CEOs, he wrote, are “uniquely prone to AI psychosis” precisely because they sit so far from the last mile of work where value actually gets created. When an executive plays with a chatbot, he sees the happy path: the clean demo, the impressive first draft, the magic. What he does not see is the next ten or twenty things — the hallucinated library call, the contract clause the model invented, the security hole, the rewrite, the production incident at 2 a.m. He watches the trailer and greenlights the movie without reading the script. And because he is the CEO, his happy-path delusion becomes your performance metric, your mandate, and eventually your layoff.

This is the mechanism the entire moment runs on. The people closest to the work — the engineers cleaning up the model’s messes, the support reps watching quality crater, the COOs discovering governance holes the CFOs won’t fund — knew the truth. The people furthest from the work made the decisions. Boston Consulting Group is already documenting the predictable result: organizations watching their own critical skills erode as judgment, problem-framing, and creative thinking atrophy under AI overuse, with more than 60% of surveyed leaders calling de-skilling a material threat within five years. We are hollowing out the people who actually understand the systems, on the orders of people who don’t, to chase a productivity gain that — by the executives’ own admission — mostly hasn’t arrived.

They fired your coworkers to pay for it

The money to build all this — the data centers, the GPUs, the frontier-model contracts — came out of payroll. Yours and your coworkers’. That isn’t a metaphor; it’s the cash-flow statement.

TechCrunch’s running tally of 2026 layoffs where companies named AI as a factor reads like a roll call. Oracle cut roughly 30,000 positions, the single largest event of the year. Meta laid off about 8,000 while shoving some 7,000 survivors into AI roles many of them reportedly hate. Intuit cut 3,000. PayPal announced plans to shed north of 4,500. Coinbase cut 700 and started experimenting with one-person teams. Block, under Jack Dorsey, hacked headcount from 10,000 to under 6,000 and told shareholders that intelligence tools had changed what it means to build and run a company. Cloudflare cut 1,100 people in a record-revenue quarter — $639.8 million, up 34%, the best in its history — and CEO Matthew Prince was almost admirably blunt about who got the axe: the “measurers,” the people in finance and legal and internal audit whose job was to count things. Accenture’s Julie Sweet drew the line hardest of all, telling staff that “those we cannot reskill will be exited.”

Add it up and more than 142,000 tech workers lost their jobs by the middle of 2026, a pace running well ahead of the post-pandemic bloodletting of 2023. And here is the detail that should end any argument about whose interests were being served: across those same months, Amazon, Microsoft, Alphabet, and Meta committed somewhere between $700 and $740 billion in AI capital expenditure — roughly 70% more than the year before. They were firing people and buying chips with the savings. That isn’t my interpretation. Challenger, Gray & Christmas said it flatly: companies are shifting budgets toward AI at the expense of jobs.

And a great deal of it wasn’t even real automation. Sam Altman — not a hostile witness — conceded there was “some AI washing,” companies blaming AI for layoffs they’d have done anyway. Marc Andreessen called AI “the silver bullet excuse” for cuts firms wanted to make regardless. Wharton’s Peter Cappelli was sharper still: companies announce that AI will cover the work, but they haven’t actually done it — “they’re just hoping.” Oxford Economics concluded that firms aren’t, in fact, replacing workers with AI at any meaningful scale. So in a large share of cases the play was simply this: fire the human, blame the machine, redirect the salary into compute, and order whoever’s left to burn that compute hard enough that the adoption numbers could justify the whole thing on the next earnings call. The worker was the input on both ends — the cost that got cut and the meter that had to spin.

Machines of loving grace, pointed at the workforce

Step back and remember what this was all supposed to be for, because the gap between the pitch and the product is the most damning thing in the entire story.

In October 2024, Dario Amodei — the CEO of Anthropic, the company that makes Claude — published a 15,000-word essay called Machines of Loving Grace. It is a genuinely beautiful document. In it, he argues that powerful AI could deliver a “compressed 21st century”: a hundred years of biological progress in five to ten. He lists the deliverables without flinching. The reliable prevention and treatment of nearly all infectious disease. The elimination of most cancer. Effective cures for genetic disease. The prevention of Alzheimer’s. The cure of most mental illness, including PTSD, depression, and addiction. And, as the headline act, the doubling of the human lifespan to roughly 150 years. This was the promise. This was the stated reason for the capital, the data centers, the urgency, the disruption. We were going to cure death.

Now look at what actually got built and shipped to the workplace. Not a cure for cancer. A keystroke logger. The same wave of AI that was supposed to compress a century of medicine into a decade arrived at your desk as bossware — employee-surveillance software that, by various 2026 surveys, now runs at something like 78% of companies and monitors around 80% of remote workers. A New York Times investigation found that eight of the ten largest employers in the country track individual productivity, often in real time. Roughly 61% of companies, by ExpressVPN’s count, now use AI-powered analytics to evaluate employee performance. And the AI layer is the whole point: these systems log every keystroke, capture your screen, run sentiment analysis on your messages to flag stress or “dissatisfaction,” build predictive-attrition models to guess who’s about to quit, and in some deployments track heart rate and eye movement. Microsoft Teams can now detect which room of the building you’re sitting in. Cisco’s Spaces platform has digitized eleven billion square feet of office into 24.7 trillion location data points. Google added the ability to archive employees’ end-to-end-encrypted messages so compliance teams can read them. The Electronic Frontier Foundation’s read on the genre is blunt and correct: most of it is punitive, built to penalize.

So hold the two images side by side. The promise was a machine that loves us — that ends disease, extends life, frees us to become whatever we choose. The delivery, for the ordinary worker, is a machine that watches us: counts our tokens, scores our keystrokes, reads our chat logs for signs of disloyalty, ranks us against our colleagues on a dashboard, and feeds the number to the manager deciding who to cut. The technology sold as humanity’s liberation is being deployed, right now, as the most comprehensive apparatus of workplace control ever assembled. And the grim punchline is that the research is unanimous it doesn’t even work: Arizona State and Cornell both find that heavy surveillance lowers productivity and raises quit rates, around 42% of monitored employees are already planning their exit, and workers simply learn to perform “productivity theater” for the algorithm — which is, of course, exactly what tokenmaxxing was. We built the surveillance state of the office to chase a productivity gain that the surveillance itself destroys. Machines of loving grace, indeed.

The sinkhole: the financial disaster nobody wants to name

Here’s the part that should unsettle anyone whose retirement account is index-weighted toward seven companies. The two firms at the center of all this — OpenAI and Anthropic, the model makers everyone is feeding — are not profitable. They are not close. They are, by the standards of corporate history, on fire.

Take OpenAI. In a single recent quarter, per Microsoft’s own 10-Q filing — actual audited accounting, not a leak — OpenAI’s losses ran around $12 billion, which annualizes toward $46 billion against run-rate revenue near $20 billion. By various analyses the company spends somewhere between $1.69 and $3.30 for every single dollar it takes in. Its internal documents projected roughly $14 billion in losses for 2026 and cumulative losses well past $100 billion before any hope of profit in the 2030s. Deutsche Bank, studying the trajectory, wrote that “no startup in history has operated with losses on anything approaching this scale,” and called the situation firmly uncharted territory. HSBC estimated the company still needs to raise something like $207 billion beyond everything already committed.

Then there’s the number that broke the spell for a lot of serious people: $1.4 trillion. That’s the scale of compute commitments OpenAI signed — to Microsoft, Oracle, Nvidia, Broadcom, AMD — against 2025 revenue of about $13 billion. Harvard economist David Laibson did the arithmetic out loud: $13 billion in revenue, $1.4 trillion in commitments, and a reasonable worry that the anticipated profits might never be large enough to justify the spend. The financing underneath it is a circle drawn by someone hoping you won’t trace the line. Microsoft invests in OpenAI; OpenAI commits a quarter-trillion dollars back to Microsoft’s Azure. Nvidia pledges up to $100 billion to OpenAI; OpenAI’s own CFO, Sarah Friar, acknowledged the money “will go back to Nvidia” as chip purchases. When Microsoft disclosed in late January that 45% of its $625 billion cloud backlog was tied to OpenAI’s IOUs, its stock fell 12% in a day and roughly $440 billion in market value evaporated. A venture executive described the whole structure to The Economist as “the WeWork story on steroids.” Jeff Bezos called it “an industrial bubble.” Goldman Sachs’ David Solomon said plainly that a lot of deployed capital simply won’t deliver returns. Even Altman admitted out loud that “people will overinvest and lose money” — and when the investor Brad Gerstner pressed him on how a company with $13 billion in revenue makes $1.4 trillion in commitments, Altman’s answer was to snap that if Gerstner wanted to sell his shares, he’d find him a buyer. Enough. That is the sound of a man who knows the math doesn’t close.

Anthropic — again, the maker of the model you may be using to read a summary of this — is in better shape than OpenAI, and it’s worth being fair about that. The Wall Street Journal’s reporting on both companies’ documents has Anthropic cutting its cash burn toward a third of revenue this year and projecting break-even around 2028, while OpenAI’s losses keep widening. Anthropic’s revenue reportedly went from about $1 billion at the end of 2024 to north of $5 billion by late 2025, and it expected a profitable quarter in 2026. But “better than the most cash-incinerating startup in history” is a low bar to clear. Anthropic still raised on the order of $65 billion this year at a valuation pushing $965 billion — near a trillion dollars for a company with a few billion in revenue and its own enormous compute obligations — and it, too, has tempered any expectation that profitability will hold while it spends to grow. Both companies filed confidentially to go public in early June. Sit with the timing: the analyst class openly says current growth rates are “the fastest they will ever be,” the very enterprises feeding these models are starting to cap their spend, and the smart money’s read is that the model makers want retail investors holding the bag before the curve they can already feel bending bends in public. The IPO isn’t a victory lap. It’s an exit.

And when it pops, you’ll pay for that too

There’s one more move in this game. It hasn’t happened yet, but the smart money is already pricing it. Ask what happens when $1.4 trillion in commitments collides with $13 billion in revenue and the gap doesn’t close on schedule.

The structure is fragile by construction. Financial analysts — including warnings echoed at the Bank of England — have flagged that the circular financing knitting OpenAI, Microsoft, Nvidia, Oracle, and a dozen infrastructure players together could produce a 2008-style cascade if a single link breaks, because everyone is, in effect, collateral for everyone else. Harvard’s Paulo Carvão framed the trap precisely: the buildout may be genuinely necessary for AI in the long run, but the near-term risk is the gap between that long-term vision and whether revenue shows up fast enough to pay for it. We are wiring a continent’s worth of data centers, on debt, for demand that the buyers are — at this very moment — trying to cap. If that demand softens even a little, the contracts don’t get honored, the vendors don’t get paid, and the cascade the economists keep describing stops being a thesis and starts being a Tuesday.

And here is the only part of that scenario that is certain in advance: when it unwinds, the executives will not be the ones who absorb it. They never are. The same year this bubble inflated, employers cut more than 1.2 million jobs across the economy — the highest total since the 2008 crisis — and named AI as the reason. That was the inflation phase: workers paying for the buildout with their livelihoods while the valuations climbed. The deflation phase has a tell, and you already know it, because you’ve watched the move twice this year — the layoff announced as “discipline,” the right-sizing that funds the “pivot,” the survivors told to do more with less. When the AI trade reprices, the cost will not land on the CEO who signed the trillion-dollar commitment or the board that rushed the IPO. It will land, as it always does, on the people who were told to burn tokens to keep their jobs — first by taking the job, and then by eating the correction. You will have paid for the boom and the bust, and nobody will have asked you about either.

And now the bill is your problem too

Which brings us back to the whiplash. The bill landed, and the people who’d cheered the combustion discovered that combustion is expensive.

Nvidia’s own VP of applied deep learning, Bryan Catanzaro, told Axios the quiet part: “the cost of compute is far beyond the costs of the employees.” Read that again, from Nvidia — the company selling the shovels conceding the digging now costs more than the diggers. A 2024 MIT study had already found AI automation was economically viable in only about 23% of the roles it examined; in the other 77%, the human was cheaper. One company reportedly ran up a $500 million Claude bill in a single month. Ramp’s data pegged the most AI-saturated firms at roughly $7,500 per employee per month. None of the promised unit economics arrived, because consumption climbed faster than per-token prices fell — by one TechCrunch-sourced figure, per-developer token use rose about 18.6 times in nine months.

So the same finance departments that treated rising token counts as a victory lap now treat them as a grease fire. Amazon shut down its developer leaderboard in late May with the internal line “don’t use AI just to use AI” — a sentence that would have earned you a performance flag two months earlier. Meta pulled its leaderboard within days of it leaking. GitHub Copilot moved to usage-based billing in June, pushing the volume-versus-value question straight onto individual developers’ invoices. Uber, after its CTO admitted the budget he thought he’d need was already gone, rolled out tiers starting at $1,500 a month. Everyone suddenly discovered “model routing” — the radical notion that you shouldn’t fire a frontier model at a task a cheap one can handle — an idea so advanced that, by Glean’s CEO’s estimate, something like 95% of enterprise AI usage is still pointed at frontier models anyway. The identical metric — tokens consumed — went from the thing that proved your loyalty to the thing your CFO is trying to drive to zero, in roughly one quarter, with no apology to anyone fired for “low adoption” or burned out gaming a dashboard.

Read the incentives, not the slogan

Here’s what I want you to take from all of this, as someone who builds and ships agentic systems for a living and pays the token bills he’s describing. Pull the threads together and a single shape emerges. The executives mandating AI mostly couldn’t use it. The technology sold as a cure for cancer shipped as a keystroke logger. The companies at the center are losing tens of billions a year on financing that looks like a magic trick. And through all of it, exactly one group of people has been adjusted, repeatedly, to make the story work: you.

None of these reversals were ever about you. Not the mandate, not the layoffs, not the cap, not the surveillance. Every move optimized for one thing — the income statement — and lately, the IPO. They adjusted you down to fund the buildout. They adjusted you up, with leaderboards and performance flags, to prove adoption. They wired you to a dashboard to prove control. And now they’re adjusting you down again to protect the margin and the multiple. At no point in any of those moves did anyone touch the executive’s incentive. They touched yours. The CEO who told you to torch $250,000 in tokens to keep a $500,000 seat, the CFO now telling you $1,500 is your ceiling, the board rushing to IPO before the curve bends — those are the same priorities, and they were never, for one second, aligned with yours.

That’s the thing about a leaderboard, and a layoff, and a surveillance dashboard: none of them were ever measuring your worth. They were measuring someone else’s narrative. When the story was we are an AI-first company, your token count was the proof and you’d better make it climb. When the story became we have disciplined unit economics, your token count was the liability and you’d better make it fall. When the story is we need to justify the buildout, you’re the line item that gets cut. Same person, same work, the whole time — only the tale the executives needed to tell their board changed, and your job was always to produce whatever number that tale required.

So no, there are no backsies for the people laid off in March to fund a capex line, or for the survivors who spent their spring climbing a leaderboard that’s since been deleted from the wiki, or for anyone whose keystrokes are now scored by a model that was supposed to be curing disease. But the executives get all the backsies they want. They get to call the mandate “innovation” when it suits them and “tokenmaxxing run amok” when the bill arrives. They get to fire you for AI and then concede, off the record, that it was washing. They get to promise machines of loving grace and ship machines of loving surveillance. They get to IPO on the way down.

The work was never the tokens. Anyone who actually ships knew that the entire time — the leaderboard was theater, the mandate was a loyalty test, the surveillance is a confession of distrust, and the only metric that ever counted was whether the thing you built solved a real problem and survived contact with production. So build for the outcome. Learn the technology more deeply than the executives mandating it ever will, because that knowledge is the one asset they can’t reprice. Hold your own line on what good engineering is, because the company’s line will move the instant it costs them a basis point. And the next time a memo lands telling you to do something “or else,” read the incentives underneath it, not the slogan stamped on top. The slogan is disposable; they’ve now proved that twice in one year. Their priorities are real and durable and — this is the part worth keeping somewhere you’ll see it every morning — they are not, and never were, you.

About the Author

Gal Ratner is the founder and CTO of Inverted Software and WhiteStar Labs and Chief Architect at Prana Entertainment, a Las Vegas–based enterprise software and AI consultancy. He has spent nearly thirty years shipping on the Microsoft and .NET stack for clients including Microsoft, Sony, Rockstar Games, 2K Games, Best Buy, and Allegiant Air, and his current work is squarely in production agentic AI — MCP servers, the Microsoft Agent Framework, SQL Server 2025 vector search, RAG pipelines, and the PLogger observability framework. He writes as “the .NET AI guy that ships,” which means he pays the token bills he’s describing and has sat on both sides of the budget meeting. When he isn’t shipping, he co-hosts the motorcycle-focused Edge Grip Podcast, trains Brazilian jiu-jitsu in Las Vegas, and is the author of the novel The Archive of Lost Suns.