There’s a number I can’t stop thinking about.
For every dollar a company spends on AI coding tools right now, roughly 18 cents turns into software that reaches a real user. The other 82 cents disappears into bug fixes, rewrites, reverts, and review cycles, much of it cleaning up code the same tools just generated. That figure comes from EntelligenceAI, which aggregated data across more than 2,000 companies using advanced AI coding tools.
Hold that next to the headline of the year: Anthropic is going public at a valuation approaching one trillion dollars, having blown past OpenAI to an annualized revenue run-rate around $45 billion, growing faster than any company in recorded history.
Both of these are true at the same time. The product is selling like nothing we have ever seen. And the people buying it, some of the most sophisticated engineering organizations on earth, are quietly admitting they can’t yet prove what they got for the money.
That contradiction isn’t a scandal. It’s a signal. And if you build, invest, or operate in this space, reading it correctly is worth more than any model benchmark.
Let me put the evidence on the table, because the specifics matter more than the vibes.
Uber rolled out Claude Code to its engineers in December 2025. By March, adoption jumped from 32% to 84% of its roughly 5,000-person engineering org. Around 70% of committed code now originates with AI. Sounds like a triumph, until you learn the company burned through its entire 2026 AI coding budget in four months. “I’m back to the drawing board,” the CTO said, “because the budget I thought I would need is blown away already.”
Then came the part nobody at a vendor wants quoted. Speaking in late May, Uber’s COO Andrew Macdonald was asked whether all that spend was translating into better products for riders and drivers. His answer: “That link is not there yet.” The usage stats, he admitted, “make your head explode,” but he couldn’t draw a line from tokens consumed to features customers can feel.
Microsoft went further. It began revoking some engineers’ access to Claude Code, moving them onto its own cheaper Copilot CLI. The framing was “tool consolidation.” The subtext was the bill.
Meta built an internal leaderboard ranking 85,000+ employees by token consumption. Nvidia’s Jensen Huang has floated the idea of giving engineers token budgets as part of their compensation. Somewhere along the way, “how much AI did you use” quietly became a goal in itself, which is the exact moment a metric stops measuring anything real.
And the now-legendary data point: one developer running an autonomous agent framework racked up a $1.3 million monthly bill, 603 billion tokens across 7.6 million requests. The bill wasn’t the result of waste. It was the result of the tool working, running flat out, doing exactly what it was built to do.
This is what I’ve started calling the tokenmaxxing trap: when consumption becomes the scoreboard, you will always be able to consume more. The ceiling isn’t usefulness. It’s budget.
Here’s the part most hot takes miss. This is not an AI problem. It’s a productivity problem that economists have watched for a century.
There is a vast, treacherous gap between task-level productivity and economic productivity. AI can absolutely make the task of writing code faster. That is real and measurable. But shipping a product that customers value is not one task, it’s a chain: deciding what to build, building it, testing it, reviewing it, integrating it, maintaining it, and not breaking the eleven things connected to it.
Speed up one link in that chain and you don’t automatically speed up the chain. Sometimes you slow it down. One analysis of over 10,000 developers found that high-AI-adoption teams took on 47% more pull requests per day, and spent the gains on context-switching, orchestration, and reviewing machine-generated code. At the median organization studied, 44% of all engineering output was reactive, fixing or maintaining existing code rather than building anything new. As AI pushes more code out faster, reverts climb faster than output. Each revert spawns a bug-fix, each bug-fix adds to the reactive pile, and the loop compounds.
Steam power took decades to show up in productivity statistics, because factories had to be physically redesigned around it before the gains appeared. Electricity was the same. The technology arrives first; the reorganization that makes it pay off comes later, and it’s the hard part. We are living in exactly that lag right now, except it’s compressed into quarters instead of decades, which is why it feels like whiplash.
This is where I part ways with the doom crowd. The “82 cents is wasted” framing is catnip for people who want AI to fail. They’re misreading it.
When a genuinely general-purpose technology lands, electricity, the PC, the internet, a huge amount of early activity should look like waste. Thousands of people poke at it, try dumb things, build stuff nobody wants, and burn money learning what works. That experimentation isn’t the bug. It’s the price of discovery, and it’s how every general-purpose technology in history found its killer app.
The proof it can work already exists. Look at the one company that’s clearly winning: Anthropic shipped 120+ features in the first 90 days of 2026, more than one per working day, across Claude Code, its API, and its models. That’s the loop closing: more shipping creates more usage, more usage creates more feedback and revenue, which funds more shipping. The 18% problem isn’t that the tools don’t work. It’s that most organizations haven’t built the system that turns tool output into shipped value. Anthropic has. That’s the whole difference.
So the question isn’t “bubble or not.” The question is sharper and far more useful:
Where does the value actually land?
Here’s the contrarian bet I’d put money on.
Most people assume that if AI makes software cheaper to build, we’ll get better versions of the software we already have, faster apps, slicker websites, smarter ad targeting. I think that’s mostly wrong, and it’s the most important thing to get right.
The incumbent software industry, the social networks, the e-commerce stacks, the SaaS giants, is in many ways a mature industry. Like steelmaking or the internal combustion engine, the core problems have been solved. Making those products 30% cheaper to maintain is nice, but it doesn’t create a trillion dollars of new value. It mostly compresses margins and quietly eliminates some jobs. That’s why Uber’s COO can’t find the link: he’s looking for AI to improve a product that is already near its local maximum.
The trillion-dollar value shows up somewhere else: in software that was never written because it was too expensive to justify.
Think about every piece of software that didn’t get built because the addressable market was too small to pay a team of engineers. The custom tool for a 40-person logistics firm. The niche internal system for a single hospital department. The hyper-specific app for a community of 10,000 people. The personalized software that serves exactly one company’s weird workflow. For the entire history of the industry, this “long tail” of software was economically impossible, the engineering cost exceeded the value created.
AI doesn’t make Instagram better. AI makes the impossible software possible. When the cost of building drops by an order of magnitude, an enormous category of software that was previously underwater suddenly floats. That’s not a feature on an existing app. That’s a new industry, and it’s where founders, not incumbents, will win.
This is why I’m not worried about the 18% number, and why I’m long on this technology while being short on the hype around the current use of it. The waste is real. The reckoning on token spend is real and overdue. But the discovery happening underneath the waste is the most exciting thing to happen to builders in twenty years.
If you’re building or running a company right now, here’s how I’d translate all of this into decisions:
1. Kill consumption as a metric. Today. The moment “tokens used” appears on a dashboard as a goal, you’ve created a leaderboard for waste. Measure cost-adjusted outcomes instead: cost per shipped feature, cost per resolved ticket, cost per completed workflow, human time saved after review and correction.
2. Put a hard cap before you scale, not after. Uber found its budget gone in four months. A runaway agent loop with no ceiling is a seven-figure invoice waiting to happen. Set spend limits at 120% of expected budget on day one.
3. Route, cache, and batch ruthlessly. Use frontier models only where reasoning genuinely demands them; route everything else to cheaper tiers. Cache repeated context. Send anything latency-tolerant to batch inference at half the cost. For most agentic systems this cuts spend 40 to 70% with zero quality loss.
4. Build the loop, don’t just buy the tool. Anthropic’s edge isn’t access to a better model, it is the model, and it still wins on shipping discipline. The advantage is the system around the tool: clear specs, codebase standards, fast review, tight feedback from production. Buy the tool, but invest in the loop. The loop is the moat.
5. If you’re a founder, hunt the impossible software. Don’t build a thinner wrapper on an existing category. Find the software that couldn’t economically exist before this year. That’s where the next decade of value is hiding, and the incumbents structurally can’t chase it.
AI didn’t make software free to build. It made a whole universe of previously impossible software suddenly cheap enough to attempt.
The companies measuring how many tokens they burned are asking the wrong question. The ones asking “what could we build now that we never could before?” are about to run the table.
The tokens were never the point. The systems are.
If this was useful, I write about AI, startups, and what’s actually working (vs. what’s just loud) for operators and founders. Follow along, and forward this to the one person on your team who keeps quoting the token leaderboard.
— Abhishek Soni