Tomorrow Is Coming to Collect

~$7,500 per employee. Per month. On AI.

That’s what the top 1% of companies are spending right now, according to the Ramp AI Index. And that number is growing 14.1% month-on-month.

The median company? ~$11.

That’s a 680x gap. Let that sink in for a moment.

Companies have been consuming AI like there is no tomorrow. Throwing tokens at every problem. Building leaderboards to gamify who can burn the most compute. Firing employees to “refocus on AI” while 84% of CFOs can’t point to a single dollar of return.

Well. Tomorrow is here. And it’s coming to collect.

The Binge

You know that feeling when you wake up after a party you probably shouldn’t have gone to? That moment when you check your bank account and wonder what exactly happened between midnight and 3am?

That’s where enterprise AI is right now.

Uber gave 5,000 engineers unlimited access to Claude Code in December 2025. By February, 63% were using it. By March, 84%. By April? The entire annual budget was gone. Four months. Done.

One executive racked up $1,200 in a single two-hour coding session. One session!!

I’ve written before about how this dependency pattern looks a lot like addiction. Turns out, when the bills hit, it looks like one too.

Tokenmaxxing: When You Measure the Wrong Thing

Here’s where it gets really interesting. And by interesting, I mean completely absurd.

Companies didn’t just accidentally overspend. They actively incentivized overconsumption.

A Meta employee built a dashboard called “Claudeonomics” that tracked token consumption across 85,000 employees. In 30 days, Meta employees consumed 60 trillion tokens. The top user? 281 billion tokens. Someone estimated it cost $50,000 per employee on AI tokens.

The term for this phenomenon? Tokenmaxxing. Treating a cost metric as a success metric. IBM called it out explicitly: “Token counts make sense as a cost metric. They are a disastrously poor proxy for business value.”

It’s the AI-era version of measuring developers by lines of code. We abandoned that metric decades ago. And somehow we thought “tokens consumed” would be different?

This is the same invisible technical debt I’ve been warning about. Except now it’s not hiding in your codebase. It’s hiding in your finance department.

The Jevons Paradox (Or: Why Cheaper Doesn’t Mean Cheaper)

Here’s the part that should genuinely worry you.

Token prices have dropped 98% since late 2022. Ninety-eight percent!! By any normal logic, enterprise AI bills should be collapsing.

Instead, they tripled. Enterprise AI bills rose 320% over the same period. The average enterprise AI budget went from $1.2 million per year in 2024 to $7 million in 2026.

William Stanley Jevons figured this out in the 1860s studying coal and steam engines. Make something cheaper, people use more of it, not less. Total spending goes up.

Three forces are driving this:

1. Agentic workflows consume 10 to 100x more tokens than chat
2. Longer context windows bloat every single request
3. Vendors ended their subsidies on frontier model pricing

Per-developer token consumption rose 18.6x in nine months. Not 18.6%. 18.6 times.

Goldman Sachs projects global token usage will multiply 24x by 2030.

You can’t cost-cut your way out of a consumption addiction.

The ROI That Never Showed Up

Let’s talk about what all this spending actually produced.

Read those numbers again. $2.52 trillion in spending. 95% returning nothing. And this isn’t 2024 early-days experimentation. This is mid-2026, over three years after GPT-4 launched.

The defenders will tell you it’s a “J-curve.” That ROI is coming, you just need patience. But here’s the thing: that 57% number (enterprises where ROI fails to outpace spend) is unchanged since 2025. The J-curve has been flat for a full year.

At what point do we stop calling it a J-curve and start calling it a flatline?

Meanwhile, They’re Firing People

And this is the part that genuinely makes my blood boil.

While 95% of AI spending returns nothing, companies are firing humans to fund the AI that isn’t working.

• Cloudflare reported a 34% revenue increase and announced 1,100 layoffs on the same earnings call
• Meta posted $26.8 billion in Q1 net income and still cut 8,000 jobs
• Block eliminated 40% of its employees to “refocus on AI” while projecting $12 billion in gross profit

And here’s the kicker. Gartner found that companies reporting AI-driven workforce reductions saw no correlation to higher ROI. They cut jobs. It didn’t help. The AI still didn’t pay for itself.

Fire the humans to fund the machines. The machines don’t deliver. Now you have no humans and no ROI.

Brilliant strategy.

The next casualty of this revolution won’t just be the people who got fired. It’ll be the institutional knowledge that walked out the door with them.

The Hangover Begins

The good news, if you can call it that, is that reality is setting in.

Microsoft is cancelling Claude Code licenses across its Experiences & Devices division. The company that invested $13 billion in OpenAI is telling its own engineers “actually, this costs too much.”

Uber slapped a $1,500/month cap per employee per tool. AT&T pulled back on GitHub Copilot. Accenture is restricting routine AI usage. Walmart capped its in-house AI agent.

Chris Reed, Priceline’s Senior Director of IT Finance, put it perfectly:

“It’s like the crack-cocaine epidemic. They let you try it to get you hooked on it, and now you’re kind of beholden to it.”

And the FinOps Foundation’s J.R. Storment:

“In April and May, I started hearing from companies: ‘Oh my god, we are 3x over our entire 2026 token budget and it’s only April.’ The whole conversation shifted from tokenmaxxing and ‘go fast’ to ‘we need guardrails.’”

The conversation used to be “What can AI do?” Now it’s “Why is our bill so high and what did we get for it?”

What Actually Sustainable AI Looks Like

I’m not here to tell you AI is useless. It isn’t. I use it. You probably use it. But there’s a massive difference between targeted, measured adoption and an all-you-can-eat binge with no accountability.

The median firm at ~$11 per employee per month? They might actually have the right idea. A ChatGPT or Claude seat per employee. Specific use cases. Measurable outcomes.

Jellyfish’s research found something telling: engineers who used the most tokens were about 2x as productive as those who used AI less. But they spent 10x the tokens to get there. That’s a 5:1 cost-to-benefit ratio that would make any CFO weep.

Their recommendation? “The best ROI comes from moving the broad middle from low to moderate usage, not pushing heavy users higher.”

Companies using AI to amplify workers rather than replace them are seeing 2x the cash flow margin expansion of their peers. The ones firing people to fund AI? Flatline.

Here’s what I think sustainable actually looks like:

1. Know what you’re spending and why. Sounds obvious. But most companies literally cannot tell you which team is consuming what, on which model, for which business outcome. If you can’t attribute the cost, you can’t measure the value. Full stop. And while you’re at it, treat your AI provider with the same engineering rigor you’d apply to any critical infrastructure dependency.

2. Set budgets before you hand out access. Uber gave 5,000 engineers unlimited access and acted surprised when they used it. That’s not an AI problem. That’s a procurement problem. You wouldn’t give every employee an unlimited corporate Amex and then be shocked when the bill arrives.

3. Measure outcomes, not consumption. Code shipped. Tickets resolved. Time saved on actual workflows. Not tokens consumed. Not “AI adoption rate.” Not leaderboard position. The moment you incentivize usage rather than results, you get exactly what you measured: usage. And nothing else.

4. Start narrow, expand with evidence. Pick three workflows where AI demonstrably saves time or improves quality. Measure them ruthlessly. Then expand. The companies getting 2x cash flow margin expansion aren’t doing everything with AI. They’re doing the right things with AI.

5. Accept that “not every problem needs AI” is a valid engineering decision. Sometimes a bash script is the answer. Sometimes a human is the answer. The best tool for a job is the one that solves it at the right cost. Not the trendiest one.

None of this is revolutionary advice. It’s basically how you’d approach any technology investment. But somehow, in the AI gold rush, companies forgot all of it. They forgot that tools need to justify their cost. That budgets exist for a reason. That measuring activity is not the same as measuring value.

The Bill Always Comes Due

Here’s my take on all of this.

We’re watching the same pattern we saw with cloud adoption play out again, except faster and with bigger numbers. First comes the gold rush. Then comes the bill shock. Then comes the painful maturation where companies actually learn to use the thing responsibly.

The difference this time? The numbers are insane. $2.5 trillion with a 95% waste rate. Companies burning through annual budgets in four months. Employees gaming leaderboards to consume tokens for the sake of consuming tokens.

Tomorrow was always going to come. It was always going to collect. The question was never “if” but “when” and “how much.”

Well. It’s here. And the tab is enormous.

The companies that will come out of this well are the ones who treated AI like a tool. Not a religion. Not a leaderboard game. Not an excuse to fire people. A tool. One that needs to justify its cost, prove its value, and earn its place in the budget.

For everyone else? I hope the party was worth it. Because the hangover is going to be spectacular.

I would be very interested to hear your thoughts or comments, so please feel free to ping me on Twitter/X or LinkedIn.