In 2019 NVIDIA agreed to pay about $7 billion for Mellanox, an Israeli company that made networking equipment for data centers. The price raised eyebrows. NVIDIA sold graphics processors, and Mellanox sold the cables, switches and adapters that move data between machines. Intel had been circling the same target. The deal closed in 2020, and for a while it looked like an expensive detour.
It was the moment NVIDIA stopped being a chip company.
What the company sells today is closer to a machine room than a component. Its current systems arrive as racks: dozens of processors wired together with NVIDIA’s own interconnects, cooled by a design NVIDIA specified, running NVIDIA’s software, and treated by the buyer as a single computer. The chip is in there somewhere. It is no longer the product.
Why the fastest chip does not win
There is a persistent assumption in technology that the best component wins the market. It is usually wrong, and in AI hardware it is wrong for two specific reasons.
The first is software. NVIDIA released CUDA in 2007, a set of tools that let programmers use graphics chips for general mathematics rather than for drawing pictures. For roughly five years this looked like a niche product for academics with unusual problems.
Then, in 2012, three researchers in Toronto entered an image-recognition contest with a neural network trained on two consumer graphics cards bought for gaming. It won by a margin that embarrassed every other entry, and the field of machine learning reorganised itself around the result within about eighteen months. The cards were NVIDIA’s. So was the software everyone copied afterwards.
What CUDA did, in other words, was accumulate. Every research paper, every library, every graduate student who learned to write for NVIDIA hardware added a small amount of weight to one side of the scale, and none of it was planned. By the time machine learning became the most valuable workload in computing, essentially all of it was already written for one company’s chips. A competitor offering faster silicon is asking a customer to rewrite work that took the industry the better part of two decades to pile up.
The second reason is the one Mellanox answered. Training a large model does not happen on a chip. It happens on thousands of them at once, and the limiting factor is often not how fast each one computes but how quickly they can tell each other what they have computed. A cluster where the processors wait on the network is a cluster running at a fraction of what it cost. Owning the interconnect means owning the number that customers actually care about, which is the throughput of the whole installation rather than the specification of any part of it.
Put those together and the competitive question changes shape. A rival can build a faster processor and still lose, because the customer is not buying a processor.
What everyone else builds
NVIDIA does not manufacture anything. The processors are fabricated by TSMC in Taiwan, on production lines that took decades and extraordinary sums to build and that no competitor has replicated. TSMC’s position is arguably the harder of the two to attack, since it rests on physical plant and accumulated process knowledge rather than on software habits, and it is concentrated on an island whose political situation is the single largest uninsurable risk in the industry.
The memory stacked beside each processor comes largely from SK hynix and Samsung in Korea, and from Micron. Broadcom and others supply switching silicon. Vertiv and its competitors build the cooling. Eaton and Schneider supply the electrical apparatus that keeps a building full of these machines running.
This is worth stating plainly because the popular version of the story has one company in it. The AI build-out is an industrial supply chain in the old sense, with several irreplaceable links, and the company at the front of it has arranged to sit at the point where the margin is highest rather than the point where the capital is heaviest. Fabricating chips requires tens of billions of dollars of plant. Designing them and specifying the system around them does not.
The parts of this that could break
Two things could loosen the position, and both are already visible.
The first is that NVIDIA’s largest customers are also the companies most capable of replacing it. Google has been designing its own AI processors since 2015 and now runs a substantial share of its work on them. Amazon builds its own. So does Microsoft. None of these efforts has to beat NVIDIA on performance to matter. They only have to be good enough for the internal workload that a company already understands, at a cost that does not include someone else’s margin.
There is a limit to how far that goes. A chip a company builds for itself does not get sold to anyone else, which means it never accumulates the outside software, the trained engineers or the second-hand market that made NVIDIA’s position sticky in the first place. In-house silicon shaves the bill. It does not create a rival ecosystem, and so far nobody has tried very hard to make it one.
The second is that software advantages decay. CUDA’s hold rests on the assumption that the difficult work of porting to other hardware stays difficult. That is an engineering problem, and engineering problems attract money when the prize is large enough. Several efforts are underway to make the leading AI frameworks run well on non-NVIDIA silicon. They have been slower than their backers hoped. They are not obviously doomed.
The shape of the bet
The reason infrastructure companies are worth watching is not that they are exciting. It is that infrastructure gets built once and rented for a long time afterwards. Railways, electrical grids, undersea cable, cloud regions: in each case the returns went disproportionately to whoever controlled the layer everyone else had to pass through, and they kept going there for decades after the initial construction was finished.
Whether AI computing settles into that pattern is not yet decided. It could equally follow the path of PC manufacturing, where the hard part was commoditized within a decade and the profits moved somewhere else entirely. The difference between those two outcomes is worth more than any individual quarter’s results, and it will be visible first in unglamorous places. Watch whether the big buyers keep expanding their in-house chip programmes or quietly wind them down. Watch whether the software that runs on other hardware gets meaningfully better.
For now, the company that started by selling graphics cards to video game players has made itself the toll gate on the most expensive construction project in the technology industry. It got there by buying a networking company nobody thought it needed.
