The Line Nobody Can Draw — Yet

“That link is not there yet”

In late May, a single sentence from an Uber executive explained the AI economy better than any earnings call. The company dangled an internal leaderboard to push its engineers onto AI coding tools, and they responded so enthusiastically — adoption leapt from a third of its 5,000 engineers to 84% — that Uber burned through its entire 2026 AI budget in four months, some engineers running $500–$2,000 a month. Then its president, Andrew Macdonald, was asked the only question that matters: had all that spending actually produced better software for riders and drivers?

Uber has since capped engineers at $1,500 a month. Hold that sentence, because it cuts against everything else you’ll read. Spend a week with the headlines and you’d conclude everyone is winning at once: Anthropic “passed OpenAI,” Microsoft’s AI is “a $37B business,” OpenAI is “worth $852 billion.” Each line is real — and each is doing a job: keeping capital flowing toward a trillion-dollar build-out. That isn’t a conspiracy; it’s how a gold rush gets financed. But it means the loudest number is rarely the most honest one.

Where the money is really pooling

If the famous labs barely earn, why are they worth so much? Because while value-out stays unproven, money-in is pouring in two places — and only one of them is profit. The first is capital. In one 30-day stretch this spring, Alphabet raised $80B, Anthropic raised $65B (at a $965B valuation), and DeepSeek lined up a reported ~$7B. Then there is the building — capex, money sunk into things a company will own — running at a scale tech has never seen. The real divide isn’t how much each spends; it’s whether you own what the money buys.

Underneath every bar is electricity: each dollar is a bet on power plants and multi-year grid connections — the real ceiling on growth. The platforms write the biggest checks and keep the hardware; the labs spend a fraction and own almost none of it. The one certain winner isn’t on the chart — NVIDIA, which sells the chips nearly all of them depend on.

The second pool is quieter and stranger: money going in circles. Nvidia invests in OpenAI; OpenAI commits to buy Nvidia chips and signs a ~$30B Oracle cloud deal; Amazon, Nvidia, and SoftBank fund OpenAI’s $122B war chest; the clouds then book the labs’ spending as their own revenue.

Each arrow is money. The same dollars circle the ring — so one company’s “demand” is another’s spending.

When a slice of the “demand” is the same dollars circling between a few mutually invested companies, it lifts everyone’s valuation on round-trip money rather than proven outside demand, and it concentrates risk instead of spreading it. The load-bearing piece is OpenAI’s path to profit; if that one company can’t get there, the loop tightens the way the telecom-financing circle did when it snapped in 2001. The giant valuations rest, to a real degree, on conviction and inflows — not on money already made. It only turns dangerous the moment you mistake the inflows for earnings.

A rollout is a purchase, not a verdict

Here is the trap that catches even careful readers, and it’s the most important idea in this piece. “Company gives AI to 300,000 employees” tells you a contract was signed. It says nothing about whether those people use the thing, like it, or get more done. The gap between bought and used is enormous — and it’s where the real story lives. Start with the product held up as enterprise AI’s great success: Microsoft Copilot.

Copilot adoption from independent trackers. The tell: Microsoft itself wired Anthropic’s Claude into Copilot “to alleviate its dependency on OpenAI” — hedging its own flagship.

So how can Microsoft brag about a “$37 billion AI business” if its flagship AI app is half-ignored? Because most of that $37B isn’t Copilot. The bulk is Azure AI — other companies (OpenAI very much included) renting Microsoft’s cloud to run their models (Motley Fool, GeekWire). Copilot is only ~$7B of it.

Then comes the part that never makes the press release: the bill itself. Because AI is sold by the token — you pay per word the model reads and writes — a flat software fee becomes a meter that spins faster the more useful the tool gets. Uber’s reckoning was the first big one; by spring it was a pattern with names attached.

Microsoft made the point loudest, against its own interest. Inside its Experiences + Devices division — the engineers who build Windows, Office, and Teams — Claude Code was adopted so heavily that usage hit 84–95% and per-engineer token bills ran $500–$2,000 a month; in May, Microsoft began revoking the licenses, winding the tool down by June 30 and steering staff back to its own Copilot CLI (TheStreet, Windows Forum). Read that twice: every dollar spent on Claude was also a dollar not proving Copilot.

Days earlier, GitHub had moved Copilot itself to usage-based billing; developers watched a month of credits vanish in hours and revolted, some projecting their bills jumping from $29 toward ~$750 (TechCrunch, The Register). The meter has stopped being a Silicon Valley anecdote and started arriving in everyone’s invoice at once.

Now hold that against the headline everyone celebrated. When KPMG put Claude in front of all 276,000 of its employees in May — the most sweeping AI bet any Big Four firm has made — it circled the globe as a triumph. It is a genuine, bold move. But today it’s a purchase; the usage curve and the invoice arrive later. If Uber, a company of engineers with the clearest use case there is, still couldn’t connect spend to value, the honest play is to watch what KPMG’s people and KPMG’s bill do over the next few quarters — not to score the banner.

Zoom out and the firm sources agree, even after you throw out the viral one. Ignore “95% of AI projects fail” — it traces to one disputed pilot where “failure” didn’t even mean zero return. The quieter, sturdier numbers tell a sharper story: HBR Analytic Services finds AI clearly lifts productivity (64% of firms) and efficiency (58%) but new revenue only 30% of the time — it cuts costs, it doesn’t yet grow the top line. Gartner expects over 40% of agentic-AI projects cancelled by 2027, and notes that of thousands of “agentic AI” vendors only ~130 are real — the rest “agent-washing.” That’s the line, half-drawn: AI reliably makes work cheaper; it has not yet been shown to make most businesses more. My bet against the consensus: the first ten-figure AI failure won’t be a model that didn’t work — it’ll be a deployment that worked beautifully and never paid for itself.

The hollowing middle

The deepest shift of all: the thing that felt scarce and magical two years ago — the model itself — is becoming a cheap commodity, fast. DeepSeek will sell near-frontier capability for pennies, give the software away (open-weight, anyone can run it), and cut its flagship price ~75% in a single move. Cost to generate ~750,000 words of AI output:

A ~100× spread for the same job — and it’s sliding down. What you pay tracks the moat the seller is protecting, not how smart the model is. OpenAI and Anthropic charge a premium for fame and trust; Google undercuts because it owns the factory; DeepSeek charges almost nothing because making “almost nothing” the going rate is its plan.

And the erosion is not hypothetical — it is already vaporizing real companies. When Anthropic shipped Claude Design in April, Figma’s stock fell ~16% that month, dragging Adobe, Wix, and GoDaddy with it.

Why I keep calling it the wrong scoreboard

Here is the force the whole field is transfixed by, and the reason I keep insisting it’s the wrong scoreboard. AI has started to accelerate its own development. Anthropic’s own account is the clearest yet: more than 80% of the code it ships is now written by its own models, the length of tasks those models finish unattended is doubling roughly every four months, and the human role, it argues, is narrowing to taste — choosing which problems matter, which results to trust, when an approach is a dead end. This is the engine under the superintelligence fear, and it is real enough to take seriously.

But notice what it does to the argument rather than to the imagination. If the curve holds, intelligence gets cheaper and more abundant still — which doesn’t settle the value question, it sharpens it. When the model is nearly free and nearly superhuman, the only thing left to compete on is the line from capability to useful work shipped per dollar. The scarce skill becomes taste — knowing which problems are worth pointing all that capability at. So the more powerful AI gets, the more the moat I’ve described becomes the whole game, not less. That is the genuine reason to watch this curve: not because it ends the value question, but because it raises the stakes on answering it — and on who is still steering.