When people talk about AI data centers, the picture that comes to mind is processors: Nvidia's chips, Google's TPUs, AMD's rivals. But a large AI model doesn't train on one processor. It trains on tens of thousands of processors working together, constantly exchanging data. If the network connecting them is slow, the world's most expensive processors sit and wait. This is the arena almost no one talks about, and it is one of the most important in the race.
Why the network matters so much
In an ordinary computer, the processor does most of the work alone. In training an AI model, work is split across thousands of processors, and every few moments they need to synchronize results. As models grow, the time spent on communication grows too. That is why cloud companies now spend enormous sums not only on processors but on switches, network cards and cables. Whoever sells the network sells part of every data center being built.
The Israeli part: Mellanox
This is where Israel comes in. In 2019 Nvidia announced the acquisition of Yokneam-based Mellanox, in a deal of about $7 billion that closed in 2020. At the time it looked like a large and odd acquisition for a graphics card company. Today it is clear it was one of the smartest deals in the industry's history: Mellanox technology, including InfiniBand and later Ethernet solutions for AI, turned Nvidia from a processor vendor into a vendor of an entire system. The development center in Israel is still responsible for a central part of that technology.
The rivals: Cisco, Broadcom and Arista
Nvidia is not alone. Cloud giants dislike depending on a single vendor, certainly not the same vendor that sells them the processors. So they also buy from Cisco, Broadcom and Arista, which push open Ethernet networks. This week Cisco showed how hot this arena is: its networking revenue jumped 43% to $11.1 billion, and it raised its forecast for AI infrastructure revenue from hyperscalers this fiscal year to $9 billion, from $7.5 billion.
The battle over the standard
Behind the competition is a technical question with enormous business stakes: which standard future data centers will run on. InfiniBand, developed largely at Mellanox, was long considered the fastest standard for connecting supercomputers. Ethernet, the world's most common standard, is cheaper and more open, and Nvidia's rivals are trying to make it fast enough for AI too. Nvidia, seeing the direction, developed Ethernet solutions of its own. Whoever wins this battle will sell equipment to every data center in the world.
The next part: light instead of electricity
The next stage of the battle is already here. As speeds rise, moving data electrically over copper becomes hard and wasteful. The solution is to move it with light, through optical fibers and chips that convert electricity to light and back. Here too Israel has a role: Migdal HaEmek-based Tower Semiconductor makes silicon photonics chips, and its revenue from the field reached about $240 million last quarter, almost half of all revenue. Tower expects annual revenue of $2 billion.
Why it's good for Israel
Unlike software companies hurt by AI, infrastructure companies benefit from it directly. Nvidia's development center in the north, Tower in Migdal HaEmek, and dozens of startups working on networking components are all on the side selling the "shovels" in the gold rush. That also explains why news of the Nvidia affair stirred concern here: the company is the largest employer in the field, and any change in its leadership is felt.
What could go wrong
Three risks. The first: a slowdown in data center investment, which would hit the whole chain at once. The second: export controls, which already affect sales to China and could tighten after the Huang affair. The third: competition from within. Cloud giants are developing their own chips, like Google's TPUs, and may develop their own networking components too, reducing dependence on outside vendors.
What to watch next week
In Nvidia's results on Tuesday, it is worth looking not only at processor revenue but at networking revenue, built largely on technology from Israel. If it keeps growing fast, that is a sign Nvidia is holding its position against Cisco and Broadcom. If it slows, that may be the first sign the rivals are taking a bite.
The bottom line
In the AI race, processors are the stars, but the network is the nervous system. Whoever controls it controls part of every dollar invested in data centers. And Israel, through Yokneam and Migdal HaEmek, has a far larger presence in this arena than its size would suggest. It is one of Israeli tech's strategic assets, even if almost no one talks about it.
