Economies of Intelligence

For a century the dominant logic of industrial competition was scale. Build the largest plant, spread fixed cost over the most units, and drive unit cost below anyone who came later. The moat was capacity. The unit of value was the unit produced.

That logic remains, but it faces a new kind of advantage that potentially compounds faster than scale does.

I call it Economies of Intelligence. Competitive and cost advantage can be a function of how fast a business learns, and a business can be designed around learning in the same way it was once designed around throughput. Intelligence now has unit costs and scaling dynamics, just as capital and labor did. When the cost of a unit of learning falls, and the value of that learning carries into every subsequent cycle, the firm that generates the most learning per unit of time wins. It turns competitiveness on its head.

The thesis

An economy of intelligence exists when a business system is designed to create more learning opportunities per unit of time than its competitors, and to convert each of those opportunities into an improvement that carries forward into every cycle that follows.

Three things have to be true.

First, the system has to run in a closed loop. Every production or execution cycle produces data, the data updates a model of the system, and the model changes what the next cycle does. If any link in that chain is outsourced, slow, or political, the loop breaks and the learning rate collapses to the rate of the slowest link.

Second, the system has to experiment inside production, not beside it. The cycles that generate revenue must also generate learning. An R&D program that runs in parallel to the business learns at the rate of the R&D budget. A business that learns from its own operation learns at the rate of the business.

Third, learning has to transfer across the fleet. One unit learning something is an anecdote. Every unit inheriting it in the next cycle is a moat.

SpaceX

SpaceX is an excellent case.

In the early Falcon 9 years every launch was paid for by the cargo. The customer bought delivery of a payload to orbit, and once the payload separated the first stage was scrap. Everyone else in the industry treated it that way. SpaceX treated the return of that stage as a free experiment. The launch was already paid for. The re-entry and landing attempt cost almost nothing at the margin, and every attempt, including the ones that ended in the ocean or in a fireball on the barge, produced data that made the next attempt better. SpaceX did not waste the opportunity.

With Starship they went further and budgeted for learning regret explicitly. Vehicles were built with the expectation that they would be destroyed. Each one flew further into the envelope than the last. To an observer trained on the old aerospace model, where a single failure can end an entire program, it looked reckless. Seen as a learning system it was the fastest possible path.

None of this would have worked if SpaceX bought engines from one supplier, structures from another, and avionics from a third. They build the engines, the airframe, the avionics, and the software in house. When a landing fails, the lesson goes straight into the next design revision without a contract negotiation. The closed loop is what turns a failed flight into a better next flight.

Then the loop compounded. Reusability lowered the cost per launch. Lower cost meant more launches. More launches meant more learning per year. SpaceX now puts more mass into orbit each year than the rest of the world combined, at a fraction of the cost of any competitor. Each of those launches is another cycle of the same loop. A competitor with a larger factory and a bigger budget cannot catch that. The gap is not in capacity. It is in accumulated cycles and a system designed to constantly learn and improve.

The pattern elsewhere

SpaceX is one of many.

TSMC’s foundry model means it runs more wafers through more process nodes for more customers than anyone else. Every wafer is a data point on yield. TSMC learns faster because it sees more of the problem than any integrated device maker can, and because it never has to pause learning to defend a product line of its own. The result is a lead in process technology that Intel, with all its scale and capital, has not been able to close.

Tesla runs a fleet of millions of cars, each one a sensor for the self-driving stack. Every driver intervention is a labelled example. Every mile is a cycle. Competitors with better sensors and more careful engineering are learning from thousands of vehicles while Tesla learns from millions.

In each case the moat has the same shape. A closed loop optimizing system that learns from every production or execution cycle gets better with every cycle. Scale economics cannot compete with that, because scale is a stock and learning is a rate. A larger plant is bigger today. A faster learning system is better tomorrow, and the day after, and it does not stop. Over enough cycles the moat becomes unassailable.

Vertical Oceans

This is the thesis Vertical Oceans was built on.

The conventional model in aquaculture, and in most biological production, is large scale capacity built around a single set of operating parameters. You find a set point that works, build as much capacity around it as you can, and keep producing to that formula. It is a local optimum defended with capital. Any change is a risk to the whole facility, so nothing changes, or only marginally.

We went the other way. Our facility is a fleet of small batch production units, each an independent biological system running its own protocol. The choice was deliberate. Small batch multiplies the number of production cycles per year, and every cycle is an opportunity to learn.

We run the fleet with lambda greater than zero. In practice that means every cycle carries a deliberate exploration term. Most of the fleet runs near the current best-known parameter vector to protect output. A designed fraction runs at set points chosen precisely because we are uncertain about them. The system spends a little production to buy information.

Which set points to test, and what to do with the result, is not left to intuition. We maintain multiple reasoning models of the biological system: causal models of how water chemistry, feed, genetics, environment, and microbiome interact. Those models sit under an optimizer that scores candidate strategies against our objectives and decides which experiments are worth running and on which units. Each harvest updates the models. The updated models change the next cycle’s protocols across the whole fleet at once, not one tank at a time after a manager reads a report.

This is the SpaceX loop in hardware and biology rather than aerospace. Production pays for the experiment. The experiment improves production. The loop closes in house because we build the hardware, write the software, formulate the feed, and select the genetics.

When small batch wins

I want to be precise about the economics, because this is where the argument is usually misread.

Small batch does not win by default. Large capacity around a fixed set point exists for a reason. It usually has lower capital cost per unit of capacity and lower fixed cost per unit of output. A fleet of small units may carry a capacity cost premium.

The condition for small batch to win is that the learning curve is steep enough, and the loop fast enough, that the improvement in unit economics from learning outruns the capacity cost premium. That is an empirical question and it depends on the system. In a process already near its physical optimum there is little left to learn and scale wins. In a complex system that is far from optimal and poorly understood, which describes most biological production and a surprising amount of industrial production, the learning curve is steep and the fleet wins.

It is also not a forced trade. Small batch production can be capital cost competitive with large scale capacity, particularly when the units are standardized, manufactured rather than constructed, and improved through the same loop as the process they run. When that holds, the argument stops being a trade-off and becomes a dominance.

What comes next

The reason this matters now, rather than as a curiosity about a few unusual companies, is that AI is about to flood industry with things worth testing.

GNoME produced over two million candidate crystal structures in a single release, several hundred thousand of them predicted stable, roughly an order of magnitude more than had been identified in all prior work. The same is happening in proteins, catalysts, formulations, and process chemistries. The bottleneck is no longer generating candidates. It is converting a rapid stream of possibilities into something produced at commercial scale.

The large plant built around one set point cannot do that. It was designed to make one thing, very efficiently, forever. What is required is a production system that can take a new candidate, run it in a real production environment, learn from it, and either scale it across the fleet or discard it, on a cycle time measured in minutes, days or weeks. That is a fleet of small batch production units under a closed loop optimizer. It is the right design for the world AI is creating, and I do not think it is a coincidence that the companies already running loops of this kind are pulling away from their industries.

Rethinking the unit of value

Economies of scale focus on one question: what is the cost of the next unit? Economies of intelligence ask a different question: what did we learn from the last unit, and how fast did the whole system get better because of it?

That requires reassessing the unit of value in industry. It is no longer just the unit produced. It is also the cycle, and the learning each cycle generates. Capacity is still necessary, but it is no longer the moat. The moat is the loop: how many cycles you run, how much each one teaches you, and how quickly the lesson reaches every unit you operate.

Firms that understand this will design their businesses around learning the way the last century’s firms designed theirs around throughput. Everyone else may find that a competitor with a smaller plant but a faster loop has quietly become impossible to catch.