Standard Oil and NVIDIA

The comparison people reach for when a company gets very large is usually an accusation. That is not what this is. Standard Oil and NVIDIA are separated by a century, two legal regimes, and entirely different conduct — and setting them side by side is still one of the most useful exercises available, because the shape of how each came to dominate is nearly identical, and that shape is a pattern you can learn to recognize before it finishes forming.

Here is the short version of the pattern: neither company won by making the thing everyone was excited about. Both won by controlling the boring layer underneath it.

What Rockefeller actually sold

The romance of early oil is the wildcatter — the gusher, the boomtown, the fortune made in an afternoon. John D. Rockefeller, who founded Standard Oil in 1870, mostly stayed out of that business. Drilling was a lottery. He went into refining, which was dull, and then into the parts of the industry that were duller still.

He built his own barrels rather than buying them. He bought timberland to supply the barrel works. He acquired tank cars, and then pipelines, so that oil moved on assets he owned. He chased by-products with unusual seriousness — the residues other refiners discarded became lubricants, waxes, and paraffin, which meant his cost per gallon of kerosene fell below what rivals could reach even before scale was counted. And he ran the whole thing on cost accounting far more rigorous than his competitors used, which let him know, refinery by refinery, exactly where a fraction of a cent could be removed.

Then there was transport. Because Standard shipped in enormous, predictable volume, it could negotiate rates from the railroads that smaller refiners could not match — rebates on its own shipments, and in some arrangements payments on rivals’ shipments too. This is the part of the story that became genuinely notorious, the part Ida Tarbell documented in 1904, and it is where the modern comparison stops being clean: some of those arrangements were coercive in ways that no serious person defends and that the Sherman Act was written to reach. But strip the coercion out and the structural insight underneath still stands, because it is the insight that mattered most. Rockefeller understood that whoever controls the chokepoint between production and the customer controls the industry, no matter who owns the wells.

By the early 1880s Standard was refining the overwhelming majority of American oil. Kerosene, meanwhile, got dramatically cheaper for ordinary households — which is why the case has been argued about for a hundred years and will be argued about for a hundred more.

What NVIDIA actually sold

NVIDIA spent its first decade as a graphics card company in a competitive market with thin differentiation. The decision that changed its trajectory was made in the mid-2000s and looked, at the time, like a strategic error: it built CUDA, a software layer letting developers use graphics processors for general computation, and then it funded that ecosystem for years — libraries, tooling, university programs, documentation — against a market that had not yet appeared.

The payoff arrived from outside. When deep learning broke through in the early 2010s, researchers discovered that the training workloads they cared about ran extraordinarily well on hardware that happened to be sitting in NVIDIA’s catalog, with a mature software stack already built around it. A decade of unglamorous investment turned out to have been laid across exactly the path the industry was about to walk down.

What compounds from there is the same logic as the barrels and the pipelines. The moat is not primarily the silicon — competitors ship capable chips. It is that a generation of researchers learned on CUDA, that the frameworks and libraries assume it, that the accumulated code and institutional habit sit on that side of the fence. Then came the acquisition of the networking layer, because at data-center scale the interconnect between chips became the bottleneck, and then full systems rather than components. Each step moved the company further from selling a product and closer to owning the layer everything else must pass through.

Which yields the cleanest sentence in the comparison. Rockefeller did not need to own the wells because he owned the pipelines. NVIDIA does not need to own the models because it owns the substrate the models are trained on. In both cases the visible glamour industry — wildcatting, frontier AI labs — is the volatile, capital-hungry, winner-uncertain part. The durable position was one layer down.

Where the parallel breaks, and why that matters more than where it holds

Three differences are worth holding onto, because they are the ones that determine what happens next.

NVIDIA does not own its own production. Standard Oil integrated backwards until it controlled nearly every physical step. NVIDIA is fabless — it designs, and manufacturing happens at foundries it does not own, concentrated in a geography with its own risks. Rockefeller would have found this arrangement unacceptable. It is a genuine structural vulnerability with no analogue in the 1880s.

The lock-in is voluntary. Nobody is contractually prevented from leaving CUDA. The barrier is switching cost — retraining, porting, rewriting — which is real and large and also entirely different in kind from an arrangement that penalizes a rival for shipping. Modern antitrust interest in large platforms is a live matter and worth watching on its own terms, but the conduct question and the structural question are separate, and conflating them is how this comparison usually goes wrong.

Standard Oil’s dominance was already eroding before the government touched it. This is the part almost everyone forgets. The great strikes in Texas and elsewhere at the turn of the century put enormous new supply in the hands of new entrants, and Standard’s share was falling for years before the 1911 breakup. Its position was built on controlling a specific geography’s logistics, and the discovery of oil somewhere else was the thing the moat could not answer. The general lesson: the moat is defined against the world as it currently exists, and gets bypassed by a change in the terrain rather than a frontal assault. The equivalent question here is not whether a rival ships a faster chip. It is whether the shape of the workload changes enough that the accumulated advantage stops being the thing that matters.

So how do you spot the next one

Four questions fall out of the pattern, and they are more useful than any list of hot companies.

Who sells to every participant regardless of which one wins? During a gold rush the miners are a lottery and the supplier is a business. Rockefeller sold refined product no matter which well came in; NVIDIA sells compute no matter which lab ships the best model.

What unglamorous thing is somebody funding years before there is a market for it? By-product chemistry and cost accounting; a software ecosystem for a use case that did not exist. These investments look like poor discipline right up until they look like foresight, and they are visible in advance if you are reading carefully.

Where is the chokepoint, and is it moving? Refining, then transport. Chips, then interconnect, then systems. Dominant firms tend to notice the bottleneck shifting before the market does, and follow it.

What would make the terrain change? Not who might compete, but what discovery, regulation, or shift in workload would make the moat irrelevant. Spindletop was not a better refiner.

And then the caution that has to be attached to all of it, because this is a publication about the gap between being right and being paid. Correctly identifying an empire is not the same as earning a return from it. Investors who recognized Standard Oil’s dominance and bought at peak enthusiasm did considerably worse than investors who recognized it and paid attention to price. The pattern above is a lens for understanding how industries settle. It is emphatically not a buy signal, and the moment it starts feeling like one is the moment it has stopped working.