In November, Uber's 5,000 engineers were tokenmaxxing on an internal leaderboard. By March, they had burned through the entire 2026 AI budget. By May, Uber capped employees at $1,500 a month. Microsoft killed most of its direct Claude Code licenses. Amazon retired KiroRank. Gartner says 70% of companies are about to slash AI budgets. Scott Galloway is calling this 1999 again.
In a matter of four months, we went from "use AI at all costs" to "ROI or die." This is the pendulum swinging wildly based on anecdotes and discourse on X, not reality.
The Paranoia Pendulum
The pendulum first swung all the way to the one side, driven by paranoia, but under the guise of ambition & progress. The narrative said every job was about to become obsolete and if you weren't on the train, you were dead. So many companies went all-in, spending fast and hard. They launched internal leaderboards for token consumption, and made it a status game. It was AI or bust, but it was a movement based on fear.
Predictably, that was an overreaction. Usage spoiled, causing spend to balloon, and while the leaderboards lit up, teams had nothing to show for it. It was Goodhart’s Law in action.
So, the counter reactionary measure swung the pendulum all the way to the other side. The narrative that drove it claims token spend is the new dot-com bubble, and any CFO who doesn't lock it down is asleep at the wheel. Enterprises are responding by capping spend, demanding ROI, and pretending austerity is really discipline and smarts. It’s the same effect just in reverse.
While both sides have a degree of correctness the paranoia echo chamber of X, and Silicon Valley discourse, means we swing the pendulum at full amplitude left & right, and at greater frequency.
The productivity paradox
This pattern has a name in economics. In 1990, Paul David documented what happened when factories adopted electric dynamos in the 1880s: productivity didn't move for twenty years. Factory owners swapped the steam engine for an electric motor and kept the same belt-and-shaft layout, the same floor plans, the same workflows. The gains only arrived when they redesigned the entire factory around distributed power — individual motors per machine, single-story layouts, new production flows. The technology was available for decades before the organizational transformation caught up.
David called this the productivity paradox. The companies cutting AI budgets today because they haven't seen ROI are the factory owners who bought a dynamo, kept the belts, and concluded electricity was overhyped. The companies spending wildly without a thesis are the ones buying more dynamos hoping volume solves the problem.
Neither is paying tuition. Neither is redesigning the factory.
So which one is right?
So here‘s the thing: of course, you need to spend; and yes, it must be disciplined. But the real cause of the wild amplitude and frequency of the pendulum swings is the lack of a leadership thesis.
When the top of the org lacks enough tacit literacy to construct its own view, the org follows whichever direction the discourse is blowing; in today’s reactionary environment, that just means everyone gets whiplash.
Neither extreme of “investment vs. austerity” is a strategy. The orgs that win this moment are the ones with a thesis about what they’re trying to learn.
It’s the age of Token Tuition
Tuition is the word I use internally when teams ask whether the spend is justified. Most people read it as a fancy synonym for "expense." It isn't.
Tuition is spend with an explicit hypothesis about what you're learning.
Tokenmaxxing isn't tuition. Burning tokens to win a leaderboard is gamified activity. These aren’t compounding activities.
At the same time, underfunded pilots aren’t tuition either. If the experiment is too small to falsify a real hypothesis, you paid for the privilege of confirming nothing.
Real tuition has four properties:
- A hypothesis per experiment: What do you believe will or won't work, and why?
- A falsifier: What result would change your mind?
- Capability accumulation as the KPI: Not features shipped. What can the team do this quarter that it couldn't do last quarter? This is what climbing the AI Adoption Continuum actually looks like inside an org.
- Learning velocity as the cadence: How fast are you closing the gap between what you knew last week and this week?
If your AI program lacks all four, you aren't paying tuition. You're either tokenmaxxing or you're doing nothing dressed up as fiscal discipline.
The Math Problem with ROI
The austerity end of pendulum has a math problem, in that demanding ROI on transformation is incoherent.
In optimization theory, premature convergence is what happens when a system settles into a local optimum because the gradient funneled it there before it could explore the full solution space. It's the cousin of Knuth's premature optimization (same family of error, different domain).
Install a brutal ROI loss function on your AI program and teams will converge on the nearest, most measurable efficiencies. Faster boilerplate. Quicker marketing copy. Summarized emails. These are real wins. They are also incremental improvements on legacy workflows, which means they live on a local minimum. The supercomputer never turns on.
The global minimum, where AI-native processes actually live, sits on the other side of an inefficiency valley. Crossing that valley costs tokens, costs time, and produces failed experiments. If your CFO needs 3x return on every dollar by next Tuesday, the org will never cross. You will be hyper-efficient, fully optimized, and irrelevant.
That is the real cost of premature ROI. Not the dollars you saved. The basin you got stuck in.
Tuition in Practice
Resolving the two ends of pendulum swings looks like this in practice.
Run a portfolio, not a moonshot: Most experiments will fail. That's the innovation math. Many small bets with capped downside beat one big bet with implied certainty. Decentralize the probing.
Cap individual experiments, not the program: A $1,500-per-engineer cap is the wrong knob. The right knob is a per-experiment budget tied to a hypothesis. Once the hypothesis is falsified, the experiment ends. Once it's validated, it earns more.
Measure the capability ledger: Every quarter, ask the team a clean question. What do we know how to do now that we didn't know how to do last quarter? Write it down. That's the actual return on tuition. It compounds in a way feature velocity doesn't.
Make the thesis explicit and public: If leadership can't articulate what the org is trying to learn this year about AI, the org will swing with the discourse. Your job at the top isn't to authorize the spend or veto the spend. Your job is to define what learning the spend is buying.
Earn the Right to Be Efficient
The endgame is hyper-efficiency. Lean, targeted, ROI-positive deployment. Everyone wants to land there, but you can't start there.
You have to pass through the messy phase to build the cognitive scaffolding that makes lean operation possible later. Skip the messy phase and you'll spend the same money on strategy theater (e.g. vendor demos, readiness assessments, and AI-flavored slide decks that make everyone feel productive without changing anything). That's the most expensive form of doing nothing.
The companies that win the next decade won't be the ones that spent the most tokens or the fewest. They'll be the ones that knew, every quarter, what they were trying to learn.
Pay the tuition and pay it intentionally.
