The Half of AI That Electricity Cannot See
From the AI is the New Oil collection
For most of the past decade, the dominant metaphor for AI has been electricity. AI is like the power grid, the argument goes -- a universal utility that makes everything run better once you plug it in. Clean, invisible, democratically available. The metaphor has shaped investment theses, corporate strategy decks, and the way an entire generation of technologists thinks about what AI is and where it is going.
The limits of the electricity metaphor
It captures something real. But it also hides something important -- an entire layer of the AI industry that has no equivalent in electricity. Seeing that layer requires a different metaphor. Not a better one. A more complete one.
The metaphor is oil.
What electricity does
Electricity powers appliances. You plug in a toaster, it makes toast. You plug in a motor, it turns. The electricity does not become toast or torque -- it triggers a function that was already designed into the appliance. This is a genuine and important capability, and it maps cleanly to one layer of the AI industry: the layer where AI powers applications. A smarter search engine. A faster document summarizer. A more patient tutor. These are appliances. The AI provides energy. The appliance performs its designed function, only better than it could without the power source.
This is what people mean when they say "AI is the new electricity," and they are not wrong. That utility layer is real, it is large, and it is valuable.
But electricity has no raw material. It has no refinery. It produces no derivatives. And this is where the metaphor breaks -- not because it is incorrect, but because it is incomplete.
What oil does differently
Oil does everything electricity does -- and something more. Oil is an energy source: gasoline powers engines, diesel drives generators. At this level, oil and electricity serve the same function. But oil has a second structural layer that electricity lacks entirely. Crude oil enters a refinery and comes out not only as fuel but also as polyethylene, nylon, synthetic rubber, pharmaceutical intermediates, and thousands of other materials that bear no resemblance to the crude that entered the process. These materials are genuinely new. They did not exist in nature. They did not exist in the crude. They were created by the refining and synthesis process itself.
Oil is an energy source and a raw material for creating new substances. That second layer -- the refinery, the derivatives industry -- is what makes the oil metaphor structurally more complete. And the AI industry has precisely this dual structure, if you know where to look.
Models as crude oil
A frontier model is crude oil. GPT-4, Claude, Gemini -- each contains vast latent capability, the way crude oil contains thousands of latent molecular compounds. In its raw form, a frontier model powers applications, just as crude oil in its simplest use powers engines. But the parallel runs deeper. Crude oil produces nothing useful until it enters a refinery. A frontier model produces nothing beyond generic output until it is guided through a structured analytical process -- the right method, the right domain expertise, the right sequence of operations.
The refinery is where the transformation happens. And the transformation is not incremental. It is categorical.
The refinery layer
The refinery produces genuinely new intelligence, not faster versions of existing work. This is the distinction the electricity metaphor makes invisible. When AI powers an application -- a chatbot, an autocomplete engine, a recommendation system -- it performs an existing function more efficiently. The output is a faster version of something that already existed. This is the electricity layer: energy applied to appliances.
But when a frontier model is guided through a structured process -- when domain expertise acts as the catalyst, when analytical frameworks direct the synthesis -- the output is something categorically different. Analysis that did not exist anywhere in the world before the question was asked. Frameworks that no individual had assembled. Connections across domains that no textbook had drawn. These are not faster versions of existing professional functions. They are new intellectual substances -- as different from raw model output as polyethylene is from crude oil.
This is the refinery layer. It exists. It is growing. And most industry analysis cannot see it, because the only available metaphor has no vocabulary for it.
The parallel extends to the specific operations inside the refinery. Distillation separates what already exists in the crude -- lighter fractions from heavier ones, each useful but none genuinely new. The AI equivalent is structured prompting that elicits a model's existing knowledge: extracting and condensing, but not yet creating. Cracking breaks complex compounds into simpler, more usable components. The AI equivalent breaks dense model knowledge into structured, accessible pieces. Synthesis -- polymerization, alkylation -- combines simpler molecules into entirely new materials. The AI equivalent is guided analysis that produces intelligence no one has produced before: original frameworks, novel hypotheses, cross-domain connections that constitute new knowledge.
Each of these processes happens inside the refinery. But distillation and synthesis produce fundamentally different outputs. Conflating them -- treating all AI output as the same kind of thing -- is like treating gasoline and polyethylene as the same product because they both came from crude. The line between extraction and creation is the most important distinction in the AI industry that almost no one draws.
The derivatives layer
And then there are the derivatives.
The derivatives layer will eventually surpass the crude -- just as petrochemical derivatives surpassed fuel. The petroleum industry's history contains a structural prediction. For decades, oil's primary value was fuel. Then the derivatives industry -- plastics, pharmaceuticals, synthetic materials -- grew until it became more valuable than the fuel business itself. The substances created by the refinery surpassed the crude that fed them.
The AI industry is approaching the same inflection. Today, most attention and most capital flow toward the crude -- the frontier models themselves -- and toward the fuel layer, the applications those models power. But the derivatives layer is forming. Original knowledge products. Validated analytical frameworks. Domain-specific intelligence systems. These are the polyethylene and the pharmaceuticals of the AI industry: new intellectual materials that bear no resemblance to raw model output and that could not have been produced without the refinery process.
The resource curse applies here too. In petroleum, countries with abundant crude but no refining capability -- Nigeria is the canonical example -- captured a fraction of the value available to those that built refineries. Norway had the same resource and built a sovereign wealth fund. The difference was not access to crude. It was refining capability.
The same structural pattern is forming in AI. Access to frontier models is becoming abundant. The cost of capable AI is falling fast. Having AI is no longer a differentiator -- it is a commodity input. The value is migrating to the refinery: to the process, the domain expertise, the analytical structure that transforms raw model capability into intelligence that never existed before.
The other half of the story
The electricity metaphor is not wrong. It captures one real and important truth about AI -- that it powers things. But it is half the story. The oil metaphor reveals the other half: the refinery, the derivatives, the new substances that the process creates. Anyone building a strategy for the AI industry with only the electricity lens is seeing the fuel business and missing the petrochemical revolution forming underneath it.