The Productivity Paradox Is Here. Now What?

Uber burned through its entire 2026 AI budget in four months. Not in experiments. Not in a skunkworks. In production — across the organization — with no measurable return to show for it.

Andrew Macdonald, Uber's president, put it plainly in an interview with The Verge in late May: "That link is not there yet... it's very hard to draw a line between one of those stats and 'Okay, now we're actually producing 25 percent more useful consumer features.'"

That is one of the most honest things any executive has said about AI in 2026. Billions spent. Budgets exhausted. And the honest answer is: we can't prove it worked.

Uber is not alone in this.

Amazon built an internal leaderboard called Kirorank — a system designed to track AI token consumption across teams, framed as a productivity signal. Employees optimized for the leaderboard. They burned tokens. The leaderboard measured exactly what it was supposed to measure. What it did not measure was whether any actual work got better. Amazon shut it down. (Financial Times, May 28 2026.)

GitHub Copilot, meanwhile, switched from flat-rate to token-based billing on June 1st. What used to cost developers $29 a month now costs some of them $750. One developer reported going from $50 a month to $3,000. The old pricing wasn't generosity — it was a subsidy. An unsustainable one that quietly masked how much compute was actually being consumed, and how unevenly.

And then there's the METR research finding that cuts deepest: AI didn't just fail to speed developers up — it slowed them down. Roughly 44% of tokens were spent fixing bugs that the AI itself had introduced.

The tools generated the problems they were supposed to solve.

I don't have the full picture, but I think there's a name for what we're watching: the Productivity Paradox. Not a glitch. A stage.

Gartner's 2025 Hype Cycle placed generative AI squarely at the Peak of Inflated Expectations. The Trough of Disillusionment typically follows one to two years later. We are now in that trough — not because AI is broken, but because this is how large technology transitions work. The pattern is not new. It is, if anything, running exactly on schedule.

I tried to map this timeline about six months ago in The Three Crucibles. I wasn't sure I had the sequence right when I wrote it. The data is starting to confirm it.

The book made a specific prediction: "We are pouring trillions into AI infrastructure — data centers, GPUs, energy grids. But in 2026, we will not yet see the corresponding explosion in GDP or corporate profit. Why? Because we do not know how to use it yet."

Uber's budget. Amazon's leaderboard. GitHub's billing shock. These are not aberrations. They are the cost of not yet knowing how to use it.

There's a scene in Contact — the 1997 film — where Ellie Arroway has spent years listening to radio static, burning through government funding, defending a project most people think is a waste. The machine gets built using plans nobody fully understands, funded by a billionaire who believes in something he cannot yet prove. The payoff isn't immediate. The proof, when it finally comes, isn't even the kind of proof anyone expected.

My take is that most organizations are in the radio static phase right now. They've turned the equipment on. They're listening. The costs are real and accumulating. The signal hasn't arrived yet in a form that can be shown to a board.

The mistake would be to shut down the equipment.

The deeper problem isn't the technology. It's the measurement.

Organizations built their AI strategies around an output model — tokens consumed, tasks automated, features shipped. That framing made sense when AI was a demo. It doesn't hold when AI becomes infrastructure. You don't measure the value of your database by how many queries it runs per hour. You measure what the business can do because the database exists.

As I wrote in the book: "The technology is not failing — it is gestating. It needs time to mature from demo to industrial backbone. But Stage Three markets have no patience for gestation."

The move now is not to spend more on tokens. The move is to build the substrate that makes tokens worth spending.

That means three things, in my experience.

Infrastructure first. Not models — protocols, integrations, data pipelines. The plumbing that lets intelligence actually flow through an organization rather than sitting in a chat window somewhere. Most companies have AI tools. Very few have AI architecture.

New processes, not new software. Bolting a copilot onto an existing workflow doesn't change the workflow — it just adds a layer of latency and a new failure mode. The organizations getting traction are the ones redesigning the work itself: what decisions happen where, who approves what, how information moves. Organizational physics, not just software procurement.

Change management as a technical discipline. The hardest part of the infrastructure phase is not the tooling. It is getting people to stop working the way they've always worked. That requires investment in rearchitecting how work flows, not just deploying new capabilities and hoping adoption follows.

"If 2025 was the year of the chatbot, 2026 will be the year of the infrastructure. This is when we stop marveling at magic tricks and start the brutal, tedious work of wiring intelligence into the actual nervous system of civilization."

That line felt slightly hyperbolic when I wrote it. It's starting to make more and more sense.

There is a tree called the serotinous pine. Its cones remain sealed for decades, sometimes longer — locked shut by a resin that only melts at extreme heat. The tree does not release its seeds in good conditions. It waits for fire. Only when the forest burns does the canopy crack open and the seeds fall into cleared, ash-rich ground. The new forest grows from destruction.

I think 2026 is the fire year for enterprise AI. The subsidies burning off. The leaderboards shutting down. The budgets running dry without ROI to show. It looks like failure. It is also exactly the condition required for something real to grow.

For companies that use this period to build infrastructure rather than defend last year's AI spend — the calculation in 2027 looks different. "In 2026, companies used AI to survive. In 2027, they use it to dominate."

That's not optimism. It's pattern recognition. As I wrote in the book: "Organizations that survived the First Crucible do not resemble their former selves. Crisis forced them to rip out their human 'middle layer' and replace it with the messy experimental AI plumbing. But by 2027, that plumbing is no longer messy. It is the Backbone."

The companies building their backbone right now — quietly, without press releases, in the trough — are the ones I'm watching.

"History will remember 2026 not as the year AI took over, but as the year of the Productivity Paradox."

I wrote that line as a prediction. It's starting to feel like a caption.

The disillusionment is not the end of the story. It is the prerequisite for the next chapter. The question worth sitting with is not whether to continue — it's whether you're building something during this phase, or just waiting for it to pass.

I wrote The Three Crucibles to try to map this sequence — the correction, the transformation, and what comes after. If you're in the thick of figuring out where your organization goes from here, it might be a useful frame.