The Intelligence Inside the Power Station

Before a city becomes a skyline, it becomes a load curve. Before a hospital becomes a place of healing, electricity must arrive at the right voltage and remain there. Before a model answers a question, a power system has already answered a harder one: can the energy be delivered now?

The AI revolution depends on that quiet achievement. It may also help improve it.

My long-term expectation is that AGI will accelerate energy science and engineering profoundly. Today's evidence supports a narrower, still consequential claim: specialized AI already helps optimize parts of the computational and physical systems through which energy is used and studied. We should distinguish those achievements from a future in which general intelligence transforms the whole sector.

The breakthrough can be a better control decision

Fusion attracts grand language because the prize is enormous. Yet useful scientific progress often arrives as a better way to control a difficult system.

DeepMind and collaborators reported reinforcement-learning control of plasma in the TCV tokamak in 2022. The system learned to operate magnetic control coils to produce different plasma configurations. This was an experimental control achievement, not a demonstration of commercial fusion electricity.

Why should someone in Nairobi care about a plasma experiment in Switzerland? Because it shows a general pattern: an AI system can search a complex control space and help operate a physical process under constraints. My inference is that similar methods, carefully adapted and validated, may support other industrial control problems. That does not mean a fusion controller can simply be copied into a Kenyan grid.

Each physical system has its own stability limits, sensors, failure modes, and responsibilities. A good simulator is valuable; it is not permission to discover every mistake on live equipment.

Waste is also an energy resource

Another route is to make useful work require fewer resources. Google's AlphaEvolve reported a scheduling improvement that recovered, on average, 0.7% of its fleet-wide compute resources. That is a compute-utilization result. It should not be relabeled as an identical percentage reduction in electricity consumption.

The broader opportunity is my own extrapolation: if more computation can be delivered from an existing fleet, the economics of a service improve, and some capacity expansion may be deferred. But demand can grow in response to lower cost. Efficiency alone does not guarantee that total electricity use falls.

For a proposed Nairobi neocloud, the immediate research question is concrete. How many useful customer tasks can be completed per kilowatt-hour under realistic latency requirements? That measure connects model quality, serving efficiency, cooling, and customer value more directly than accelerator utilization alone.

Three horizons for an African energy strategy

The first horizon is operational. Improve load forecasting, maintenance prioritization, storage dispatch, and anomaly detection using well-defined data and independent evaluation. These are proposed applications; their value must be demonstrated against current practice. A model that predicts a fault after the maintenance team already knows about it has added little.

The second horizon is engineering design. Use AI to propose better components, explore materials, and optimize constrained systems. Here the validation cost matters. A promising battery material must be synthesized, characterized, cycled, manufactured, and assessed for safety. A simulated result is a candidate, not a product.

The GNoME materials-discovery paper is an important reference for large-scale computational searches for stable crystals. Stability predictions can expand the candidate space; they do not establish that a candidate is an affordable, manufacturable energy technology. That distinction is exactly where a local experimental laboratory can contribute.

The third horizon is scientific discovery. More general systems could help generate hypotheses, design experiments, and connect ideas across fields. This is where the AGI thesis becomes most powerful and least certain. We should invest in instruments and laboratories now so that better hypotheses can meet reality when they arrive.

A country that only imports finished equipment may miss some of this learning. Universities, utilities, laboratories, and manufacturers need shared projects that leave behind local data, technical skill, and maintainable systems.

The grid refuses to be a metaphor

There is a temptation to describe intelligence as though it dissolves every physical constraint. It does not. Transmission lines take time to build. Transformers need materials. Storage has losses. Contracts determine who gets paid. A technically excellent control policy can fail commercially if the market gives it the wrong incentive.

Kenya's opportunity therefore requires systems thinking. A data center can become a valuable customer for reliable electricity, but its presence does not automatically improve access for nearby households. The power contract, grid upgrades, operating pattern, and local economic contribution determine that outcome.

If a developer claims flexible AI demand will support the grid, ask how much load can actually be interrupted, for how long, and under which customer commitments. A training job and an emergency medical inference service have different flexibility. The details are the policy.

Build the instruments that intelligence can use

I am optimistic because energy progress can compound. Better simulation can improve experiments. Better experiments can improve models. Better models can improve design and control. A discovery engine connected to reliable physical feedback could shorten many expensive searches.

The word “could” belongs in that sentence. It is a research and industrial thesis, not a report that AGI has already solved energy.

Abundance begins when a promising idea survives contact with a meter, a machine, and a community that needs the power.

That is the energy future I want Africa to help build: ambitious enough to pursue new science, disciplined enough to measure what reaches the socket.

Editorial note: This essay expresses the founder’s ideas, developed with AI-assisted research and writing and reviewed against the linked sources on 29 October 2026. Convictions, proposals, and forecasts are identified in the text; scenes framed as imagined are illustrative.

Cover photograph / illustration: Matthew T Rader. CC BY-SA 4.0. Original source. Illustrative image: Transmission towers at sunset in East Texas.jpg. Credit: Matthew T Rader; CC BY-SA 4.0. This image does not document the research or proposed facility described.