Nvidia andMellanox

The inference from hindsight

The common inference from what happened next is that Nvidia foresaw the AI boom. The announcement supports a narrower claim. Jensen Huang, Nvidia's chief executive, said that “addressing this demand will require holistic architectures that connect vast numbers of fast computing nodes.” The demand he referred to was the one created by AI and data science, which were already growing in 2019. The release describes the deal as one that “will unite two of the world's leading companies in high performance computing” (joint press release). Nothing in it predicts a particular application or a particular scale.

What Mellanox sold

Mellanox made high speed networking hardware, mainly for a technology called InfiniBand. It is not internet networking. It connects machines inside one facility, where delay is measured in microseconds and consistency matters more than distance. Its customers were largely scientific computing centers, such as national laboratories, weather services and genomics institutes.

Why that equipment sets the value of other equipment

A computing job too large for one machine is divided across many. Those machines cannot work entirely alone, because each periodically needs results the others have produced, so they stop, exchange data and continue. If the connection is slow or inconsistent, every machine waits for the slowest exchange.

A hypothetical example shows the size of the effect. Suppose a facility has a thousand processors and the network keeps each of them waiting for ten percent of the time. The operator is paying for a thousand processors and getting the output of nine hundred. A faster network that removes the waiting adds the equivalent of a hundred processors without buying one. The customer therefore compares the network's price with the cost of idle processors, not with the price of other cables, and that comparison is why a supplier of what looks like commodity hardware can charge well.

Why buy and not partner

The strongest objection is that the two companies already worked together. Mellanox's chief executive, Eyal Waldman, referred in the announcement to “our longstanding partnership.” If the partnership worked, a purchase paid a premium for something Nvidia already had.

Ownership gave Nvidia something a partnership did not, which was control of the roadmap for a component that determined how its own products performed in the largest installations, and protection against a competitor owning that component. The evidence that would decide the matter is how Nvidia sold the combination afterward. Selling networking and accelerators together as one system, which Nvidia later did, is consistent with the ownership argument. It does not prove the premium was necessary.

What a reader should take from it

Separate what an acquirer said it was buying from what happened after. The announcement is adequately explained by what Nvidia's customers were doing in 2019, and attributing the result to foresight makes the decision look less repeatable than it was. The repeatable version is to ask what determines how much use a customer gets from equipment it has already bought, and to look for suppliers of that component.