The Open Loop

Enterprises didn’t get a worse AI. They bolted a bigger engine onto a car with no steering.

There’s a comfortable story going around that AI “works for small teams and fails at scale.” It’s true. It’s also explained by all the wrong reasons.

The usual one is some hand-wave at network effects — more people, more data, more surface area, so the big org should win and somehow doesn’t. That gets the law exactly backwards. Network value scales with the square of the nodes; if AI rode that curve, the giants would be extracting the most from it, not the least. The people invoking it are borrowing the shape of a curve while flipping its sign, and hoping nobody checks.

So let’s check. Then let’s trace the chain to where it actually leads, which is not economics and not networks. It’s control theory. And it’s less comforting than the version where you win just by being small.

The thing that scales is the bill

In January 2026, Uber handed Claude Code to roughly five thousand engineers and built an internal leaderboard ranking teams by how many tokens they burned. Adoption went from a third of engineers to most of them in two months. By April the entire annual AI budget was gone. Power users were running two thousand dollars a month.

The reflex read is “AI got expensive.” Wrong again. Listen to what the operations chief actually said when the budget cratered: he couldn’t draw a line from any of the usage stats to “now we’re producing 25 percent more useful consumer features.” Not “too expensive.” Can’t see the output.

That’s the whole story, and the cost was never the point. A company will happily pay ten times more for something it can measure returning twenty times more. The bill only becomes a crisis when the value is invisible — because then any number looks unjustified, since there’s nothing to justify it against.

So the first correction: this isn’t a cost problem wearing a cost costume. It’s an attribution problem. The enterprise multiplied something. It just can’t find what.

Multiply a zero

Here’s the mechanism. AI is a multiplier. It doesn’t have a direction of its own — it amplifies whatever direction you were already pointed.

Picture an organization as a field of vectors, one per person, each the direction that person is actually pulling. The useful output of the whole thing isn’t the sum of everyone’s effort. It’s the sum of everyone’s effort after the cancellations — the displacement left over once the people pulling against each other have annihilated each other’s work.

Now turn up the multiplier.

If the vectors mostly agree, you get more displacement in the right direction. Amplification becomes acceleration. If the vectors mostly cancel — which is the resting state of any large organization, for about a million reasons all of which reduce to misaligned incentives — you don’t get displacement. You get heat. More motion, more spend, more activity, and a net travel of approximately zero.

Five thousand engineers, each amplified flawlessly, each pointed a few degrees differently. The multiplier did its job perfectly. It multiplied a zero. The leaderboard measured the energy and called it progress, which is exactly the mistake of measuring an engine by its temperature instead of the car’s position.

This is why the same per-head spend is a triumph at a kitchen table and a five-alarm fire in a tower. It was never the size of the bill. It’s whether there’s a coherent vector for the multiplier to land on.

You are your own denominator

At a team of one, the vector field has one arrow. There is nothing to cancel. Whatever the multiplier touches converts directly into displacement, and — this is the part that matters — the displacement is attributable by construction, because there is exactly one place for the result to land. Your bank account. Your shipped thing. Your P&L.

The CEO can’t draw the line from tokens to value because between the token and the value sit fourteen layers, nine reorgs, and a few thousand people whose vectors point at their own promotions. The solo operator draws the line without trying, because the operator is the line. Sensor, decision, and outcome are the same node.

Which means the token math everyone’s arguing about — subsidized consumer plans, metered enterprise seats, who’s profitable at what tier — is a sideshow. Even a tenfold price hike is a rounding error against any business that’s actually producing value, because the value is visible enough to weigh against the cost. The enterprise’s much smaller per-head spend is a catastrophe for the opposite reason: it’s weighed against fog.

Cost was downstream the whole time. The variable is attribution, and attribution is a function of how many vectors you have to disentangle to see the result. At one, you don’t disentangle anything.

Mass

“But the solo operator is confidently wrong,” says the enterprise, reasonably. “We hedge. Five thousand people pulling in five thousand directions means someone’s aimed at the right thing. Your lone genius bet everything on one vector and it was the wrong one.”

True, and it doesn’t save you, because of mass.

A small team has almost none. A wrong move is cheap. A correction is cheap. An experiment is cheap. The cost of changing direction is the cost of one mind changing its mind. A large org’s cost of changing direction is the cost of re-aligning every node with every other node — and that re-coordination scales with the square of the headcount while the headcount scales linearly. Mass outruns size. By the time you’re large, turning is a quadratic expense paid in a currency you don’t have: agreement.

So the small team doesn’t win by predicting the right direction. It wins by converting the direction problem from a prediction problem into a search problem. You don’t need foresight if correction is cheap enough to iterate. You navigate by trial, not by oracle. The supertanker has to be right in advance because it can’t afford to be wrong; the speedboat can be wrong eight times before lunch because each wrong is a fender bender, not a shipwreck. Less mass isn’t just less momentum. It’s a higher search rate, and search beats foresight whenever the territory is unknown — which, for anything worth doing, it always is.

The sensor nobody reads

Here’s where the enterprise’s last consolation dies.

You’d think the giant at least has sensing — five thousand people means thousands of eyes, surely someone sees the iceberg. And someone does. The engineer three layers down knows the product is wrong. The signal exists.

A signal is worth nothing unless it reaches someone who can act and is paid to. And that conjunction — can-act and will-act — doesn’t hold flat as the org grows. It decays. Every layer between the eye and the wheel attenuates the signal, curates it, launders it into something that won’t get the messenger fired. The person who can turn the ship is structurally insulated from the people who can see where it’s going, and the people who can see are rewarded for staying quiet.

So count it honestly. Raw sensors go up with size. Connected sensors — sensors wired to a hand that can and will move the wheel — go down with size. Effective sensing is maximized at small, and it’s maximized for the dumbest possible reason: at a team of one, the eye, the brain, and the hand are the same organ. Loop length is zero. No attenuation, no laundering, no incentive gap, and the one watching is the one whose own money is bleeding. The small team’s sensing edge was never more sensors. It’s no distance between sensing and steering.

And the one scalar the giant actually steers by? Share price. Which isn’t merely a lagging sensor — it’s a correlating one. It collapses thousands of independent internal signals into a single number, and every competitor reads the same number. So large organizations don’t just stampede together. They’re blind together, off the same eye. Everyone blitzscaled at once, everyone laid off at once, everyone is pouring into the same AI capex at once — not because thousands of sensors independently concluded the same thing, but because they’re all squinting at the same tape. The diversity that was supposed to be the hedge lives at the bottom of the org and is forbidden from ever touching the wheel.

The current AI bubble is that pathology in real time: a whole sector steering by “the market rewards AI announcements” instead of “is this producing attributable value” — which is the Uber gap, herded.

High gain, broken loop

AI multiplies the gain. Whether that becomes control or chaos depends on whether the loop is closed — and loops close tightest at one.

That’s the law the whole thing collapses into, and it isn’t a metaphor, it’s the literal frame:

• AI is gain — it multiplies whatever the actuator does.
• Alignment is the setpoint — the direction you’re steering toward.
• Mass is the turning radius — and the size of the crater. It grows quadratically.
• Loop length is whether the loop is closed at all — and it opens up as you scale.

A small team is a tight control loop: sense, decide, act collapsed into one node, near-zero latency, near-zero loss, one aligned incentive. A large org is a broken loop: rich sensing, lossy transmission, misaligned actuation, steering by a single correlated lagging scalar.

Now turn up the gain on both.

High gain on a tight loop is responsive control. High gain on a broken loop is oscillation — overshoot, instability, thrash. Which is exactly what enterprise AI looks like from outside: blow the budget, panic, yank the licenses, herd to the next thing, repeat. They did not buy a worse model than you did. They bolted high gain onto an open loop and got the textbook result that any first-year control systems student could have predicted on a napkin.

Enterprises can still turn. But only crisis-gated — the signal has to get loud enough to punch through every layer and flip the incentive so the person at the wheel finally has their own neck on the line. That only happens at catastrophe. The giant turns on a heart attack; the small team turns on a hunch over coffee. The one exception is domains where the law mandates an incentive-immune channel from sensor to actuator — aviation safety reporting, nuclear stop-work, medical adverse-event systems. Notice the boundary: those exist only where failure means physical death, so society forced the loop closed by statute. For “we are building the wrong thing,” no regulator builds you that channel. You get the open loop by default.

The part that’s actually uncomfortable

If you’re small and reading this feeling vindicated, stop. The same logic that frees you indicts you.

Low mass makes the turn cheap to execute. It does nothing to make you decide to turn. The very coherence that made you fast is what deleted the dissent that would’ve told you you’re wrong. You have a frictionless wheel and no one’s hand near it but your own. Cheap correction is worthless if the trigger never fires — and a one-person org is the easiest place in the world for the trigger to never fire, because there’s no one in the room to say “this is two signups for a reason.” You won’t crash because you couldn’t turn. You’ll crash because nothing told you to, and you were too aligned with yourself to notice.

And the lightness you’re counting on isn’t a property you were handed for being small — it’s a discipline you have to keep choosing. Mass isn’t just headcount. It’s sunk capital, signed contracts, public promises, and the founder’s ego fused to the idea. Take the money, name the vision on a stage, fall in love with the plan, and you can carry enterprise-grade momentum at a headcount of three. You’ll have all the turning cost of the supertanker and none of its sensors. The worst of both.

So the edge was never “small.” Small is just the cheapest way to buy a closed loop, and the loop comes half-built — you get the tight steering for free and you have to install the sensors yourself, on purpose, against the grain of being small and sure. Customer reality piped straight to your face. Advisors paid in candor. A standing discipline of hunting the evidence that proves you wrong. You are manufacturing the dissent your structure deleted, because the structure that made you fast made you deaf.

No neat ending

There isn’t a tidy conclusion, because the law cuts in every direction at once. The giant has the sensors and can’t reach the wheel. The solo operator has the wheel and unplugged the sensors. AI hands both of them a bigger engine and walks away.

The ones who win this aren’t the small and they aren’t the large. They’re whoever runs the tightest loop they can while paying — deliberately, against their own nature — for the one half of it their size left out. The giant that buys an incentive-immune channel to the wheel. The solo operator who buys back the dissent that being alone deleted.

Almost nobody pays for the half they’re missing. The giant won’t, because the salary of everyone between the sensor and the wheel depends on the signal never arriving. And the solo operator usually won’t either — because installing the thing whose entire job is to tell you you’re wrong is the least fun purchase in entrepreneurship, and being small means no one can make you.

The multiplier is already in everyone’s hands. It’s just that most people pointed it at the engine and never looked at the wheel.