The capacity dividend.

Intelligence changes the location of scarcity

A factory can have a better model and still ship the same number of parts. A data centre can have its chips installed and still wait for an energised connection. An engineering team can run more simulations and still miss its production deadline.

These are different problems with a common structure: improving one capability does not necessarily improve the system’s economically useful output. As AI makes certain forms of reasoning, design exploration and coordination cheaper, the constraint can move elsewhere—to a transformer, a qualification process, a production changeover, or the person authorised to approve a decision.

My working thesis is that an important class of deep-tech companies will earn its place by moving those constraints. I call the resulting gain the capacity dividend: more useful output, or a shorter path to that output, from a physical system whose expansion is expensive, slow or otherwise constrained.

The engineering instinct is to ask what the technology can do. The investment question adds three further tests: does it change the limiting factor, can a customer realise the benefit, and can the supplier deliver it repeatedly at attractive economics? An impressive answer to the first question does not settle the other three.

A megawatt is valuable where it can be delivered

The energy branch makes the distinction unusually clear. The IEA’s 2026 central projection puts data-centre electricity consumption at 950 TWh in 2030, up from 485 TWh in 2025. That is a projection for all data centres, not AI alone. Improving energy efficiency per task can coexist with rising aggregate consumption as adoption and workload intensity increase. [1]

Yet annual electricity volume is an incomplete description of the customer’s problem. A developer needs power at a particular location, at the required voltage and reliability, on a date that supports its commercial plan. Additional generation elsewhere does not automatically satisfy that requirement. Nor does a renewable procurement agreement guarantee continuous physical delivery.

Grid optimisation can therefore create value alongside new generation and network construction. The IEA estimates that dynamic ratings, topology optimisation and advanced power-flow control could enable up to 330 GW of additional generation, storage and demand to connect without reinforcement. This is conditional potential across networks, not a pool of universally available spare capacity. [2]

The mechanisms matter. A dynamic rating estimates safe loading under actual conditions; topology optimisation changes network configuration; power-flow control redirects flows. None creates electricity. The commercial milestone is capacity an operator will actually recognise in dispatch or a connection agreement, including contingencies and safe fallback. A more accurate model is valuable only insofar as it changes an approved decision.

This leads to a less obvious underwriting issue: the beneficiary and the buyer may be different. A connecting customer values earlier revenue; the operator carries reliability obligations and works within a regulated spending framework. The company must solve that purchasing and approval problem as deliberately as the technical one. A large social benefit is not, by itself, a repeatable sales motion.

In a factory, the unit of progress is accepted output

Manufacturing is where I find this thesis most actionable. It connects engineering, equipment and operating decisions in a setting where performance can ultimately be measured against a physical product. The customer buys a part that meets specification, delivered when needed—not a model’s benchmark score or a robot’s most successful demonstration.

I separate the opportunity into six customer outcomes: faster engineering and qualification; cheaper deployment and changeover; reliable execution of variable tasks; higher first-pass yield; greater asset availability; and better operating decisions. These are distinct routes to value. They should not be added together indiscriminately: preventing a defect and avoiding downtime may describe overlapping contributions to the same extra production.

Consider a hypothetical line where machining can produce 120 units per hour, inspection can accept 100, and packaging can handle 110. Raising machining capacity to 150 does not raise shipped output while inspection remains the constraint. Improving inspection to 115 moves the limit to packaging at 110. Even that gain creates incremental contribution only if demand, inputs and working capital support it. This is an illustration of constraint migration, not a forecast for an actual factory.

For a founder, the implication is to define the improvement at the level where the customer gets paid. An inspection product may protect quality without increasing throughput. Predictive maintenance may reduce risk without producing measurable additional units during a short pilot. Those benefits can still justify a purchase, but they require the right baseline and a commercial argument that does not depend on imaginary output.

The deployment learning curve may matter more than the demo

The most interesting manufacturing wedge may sit between a technically capable system and an economically repeatable installation. Variation in parts, fixtures, tolerances, processes and safety requirements can turn each deployment into an engineering project. Better AI creates an opportunity to reuse more of that work; the business must demonstrate that reuse rather than assume it.

Public product descriptions already show different architectural approaches. PhysicsX argues for connecting engineering workflows beyond accelerating a single simulation. Augmentus describes scanning real parts and generating adapted robot paths, including retrofits to existing cells. Sunrise Robotics describes integrated dual-arm cells with perception and simulation-based training. These are examples of product scope, not independent evidence that one company has solved generalisation or achieved superior unit economics. [3] [4] [5]

The productive comparison is what happens at the second and third installation within a defined application. Do engineering hours fall? Does commissioning become more predictable? How much of a new part variant can be handled with existing tooling, process knowledge and validation? Does reduced installation effort survive once exception recovery, operator training and subsequent field support are included?

That is a deployment learning curve. It is different from a model learning curve: accuracy can improve while every new site still requires substantial bespoke work. A startup becomes more compelling when its accumulated knowledge reduces the cost of delivering the next accepted outcome. The useful boundary might be one family of finishing tasks or one class of components. It need not be arbitrary factory automation to support a substantial business.

I am particularly interested in tasks with enough variation to make rigid programming expensive, but enough recurrence for engineering and learning to amortise. Too little variation weakens the reason to use AI; too little recurrence weakens the economics. Finding that middle ground requires process knowledge, not simply a large manufacturing market estimate.

Creating capacity and capturing value are different achievements

A solution that saves a customer money can still be a poor business for its supplier. The relevant margin includes hardware, integration, inference, warranty, replacement parts and field service. Recurring revenue is not enough if recurring delivery absorbs it.

Software on existing assets, hardware-software retrofits and integrated autonomous systems allocate these costs differently. Integration gives the supplier more control over performance, but can also bring inventory, acceptance and support obligations. A retrofit preserves installed capital but inherits interfaces and equipment variation. Software avoids some hardware exposure while potentially carrying heavy implementation work. I favour control of the hardware when the outcome requires it; I do not treat that control as proof of superior returns.

Contracts matter as much as architecture. Equipment sales, subscriptions, leasing and output pricing distribute financing and performance risk differently. An output-based agreement can align incentives while transferring utilisation risk to a startup that does not control the customer’s demand. The diligence question is who pays when the machine is available but there is no work, and who pays when it is not available at all.

There is also a potential advantage in making performance easier to agree and verify. A supplier that can define acceptance criteria, establish a baseline and demonstrate repeatable improvement may reduce the buyer’s uncertainty enough to unlock a budget. That is my investment inference, not a universal moat: the same clarity can help customers compare alternatives and negotiate away some of the value.

Three places I would look first

Engineering-to-production workflows in a narrow component family. The attraction is linking design exploration to manufacturing constraints and test evidence. I would look for an engineering leader with an expensive iteration loop and a product that shortens time to a validated design. The thesis weakens if faster inference merely moves work into manual preparation or validation, or if conventional solvers improve enough to erase the proposed advantage.

Adaptive automation for recurring, variable industrial processes. Finishing, coating or inspection-to-rework are useful areas to investigate because part geometry and condition can vary while process expertise remains reusable. The proof is lower engineering effort per accepted deployment, dependable task completion and repeat purchases. The thesis weakens if each installation still needs a new integration project, or if human recovery costs overwhelm the productivity benefit.

Operator-approved access to constrained electricity networks. Here I would start with a specific network constraint and an identifiable route to procurement and cost recovery. The proof is approved capacity or an operating improvement the customer can monetise, rather than a better dashboard. The thesis weakens if the released headroom exposes another binding constraint, if the operator cannot credit it, or if the buyer cannot retain enough of the benefit to pay.

These are starting points for research, not a claim that every company in these categories fits venture economics. Each requires a separate view on market size, deployment time, customer concentration, capital needs and the alternatives likely to exist when the product reaches scale.

A defensible business must learn something it can keep

“More deployments create more data” is an incomplete explanation of defensibility. The company needs permission to use that data, coverage of relevant failures, reliable labels or feedback, and an architecture that transfers learning across customers. A growing installed base can be a service obligation before it becomes an advantage.

The stronger question is what becomes systematically easier with every installation. It might be qualifying a process, predicting an exception, calibrating a sensor, generating a fixture, or commissioning a familiar equipment configuration. If that knowledge is reusable and difficult to reproduce, it can improve delivery economics and customer confidence together. If it remains in individual engineers’ heads, growth may still depend on proportionate hiring.

The competitive benchmark must also move forward. Incumbent equipment suppliers have distribution, service networks and trusted workflows. Foundation models and engineering tools will improve. A startup needs an advantage that survives those changes: a proprietary measurement capability, a qualified process, a tightly integrated control loop, or accumulated deployment knowledge with demonstrable economic value.

What would make me change my mind?

I would become less convinced if deployment effort stayed linear with revenue, if customers enjoyed pilots but did not expand, if the improvement disappeared under full operating conditions, or if the supplier’s contribution margin deteriorated as it assumed more responsibility. I would also reconsider a product that solved yesterday’s constraint after the customer’s process or competing technology had moved on.

Over the next five years, my expectation is that more industrial AI products will be judged on responsibility for an outcome: reaching an accepted design, bringing a cell into production, producing a conforming part, or making usable power available. This is a hypothesis to test, not a promise that all industrial workflows will become autonomous.

The attraction of the physical economy is that intelligence meets a consequential test. A part either meets specification or it does not. A connection is usable under agreed conditions or it is not. The companies I want to understand are those that turn improving intelligence into a measurable increase in what the world can reliably deliver—and learn to do it again with less friction.