Inside the quiet handover from silicon to photonics

For fifty years, computers have moved information the same way: electricity down copper wires. Inside a chip, across a circuit board, between servers in a data centre. Small distances, huge speeds, and — until recently — a manageable amount of heat.

That equation has quietly stopped working. The AI boom needs chips that talk to each other constantly, at speeds that make the copper itself the bottleneck. Pushing electrons a few centimetres across a board now burns almost as much power as the actual computing. Engineers have known this was coming for years. In 2024, the industry finally admitted it.

The replacement is light. Instead of shoving electrons down copper, you send pulses of laser light down glass — the same idea as the fibre-optic cables that carry the internet, but shrunk down and moved inside the machine itself. It's called silicon photonics. Light travels faster than electricity through a wire, generates almost no heat, and doesn't care whether the destination is one centimetre or one kilometre away.

The company leading the shift is Nvidia. At its 2024 developer conference, it showed off networking switches with the lasers built directly into the chip package, instead of plugged into the front like a cable. That sounds like a small engineering detail. It isn't. It's Nvidia telling the entire industry that the next generation of AI data centres will be wired with light.

Around Nvidia sits a small cast of companies that saw this coming a decade ago. Ayar Labs, backed by Intel and Nvidia themselves, makes tiny optical chips that sit next to a processor and let it talk in light. Lightmatter, valued at $4.4 billion last year, is building something even more ambitious: a base layer for chips that routes light between them the way a motherboard routes electricity. In Belgium, a research lab called imec has become the place every startup goes to get their first prototypes built.

The strange twist is that the biggest name in chips, Intel, chose to walk away. After twenty years of internal work on silicon photonics, it sold the whole division in 2023. Its bet was that a standalone optics business couldn't fund itself. What has replaced Intel is a supply chain: Taiwan's TSMC, a handful of specialists in the Netherlands and Japan, and the compute giants themselves.

The customers driving all of this are the hyperscalers — Microsoft, Google, Meta, Amazon. When you spend forty billion dollars a year on data centres, you notice when a third of the power budget is being burned on shuffling data instead of computing it. Light fixes that. It also, not incidentally, lets you build data centres bigger than the electrical grid around them would otherwise allow.

None of this looks like a revolution from the outside. The parts are millimetres wide, the acronyms are ugly, and the timeline runs past 2030. But the direction is settled. The next leap in AI won't happen over copper, and the engineers rebuilding the wiring underneath the machines are, for now, being left alone to finish the job.