The Insane Physics Behind Light

In March 2025, Jensen Huang, the head of Nvidia, did a piece of math out loud. A million GPUs. Six transceivers each. Thirty watts per transceiver. "180 megawatts of transceivers," he said. "They didn't do any math. They just move signals around."

A transceiver is a plug that turns a chip's electrical signals into light and back. In his math, each one costs about $1,000. And he said the part that matters: "energy is our most important commodity. Everything is related ultimately to energy."

The math is the cheap part

You'd think the limit on AI is the math itself, how fast a chip crunches numbers. But fetching one number from DRAM, the memory that feeds a GPU, takes about 6,400 times the energy of adding two numbers. When an Nvidia A100 trains an image model, 60 to 84% of its working energy goes into moving data. The math gets what's left.

Data centers, the buildings full of these chips, used about 415 terawatt hours in 2024. The International Energy Agency expects that to more than double by 2030. Every watt spent shuffling numbers is a watt that never goes into the math.

So what happens when those numbers travel as light? And can light do the math, too?

Copper leaks

Why does moving a number cost so much? Inside today's machines, data mostly travels as electricity through copper. Copper works like a leaky garden hose. Push ones and zeros through it faster and more of the signal turns to heat and fades.

A gigabit is a billion ones and zeros. At 200 gigabits a second, a copper lane carries a clean signal only about a meter. Past that, the data has to jump to light.

Photonics means building with light. In AI, light has two jobs. Job one is moving data between chips, called the interconnect, and that's shipping today. Job two is doing the math with light, and that's still in labs and pilots.

Nick Harris, Lightmatter's founder, told Forbes why light wins the first job. Electrical signals, he said, "just kind of dissipate," while "with light, you're able to transport data over extreme distances." His company had just announced an 800 Gbit per second optical fiber. "Your house has 1 Gbit per second internet connection. Each one of our optical fibers can power 800 houses."

Chips still think in electricity, though, so every signal has to switch to light somewhere.

A revolving door for light

So how does light on a chip carry data? Silicon photonics builds tiny light parts in the same factories that make computer chips. Silicon steers light well but is bad at making it, so a separate laser shines one pure color of infrared light. That light runs through a waveguide, a silicon wire for light about half a micron wide.

Then a modulator, a fast shutter, blinks the light to carry data. The classic Mach-Zehnder splits the light into two arms, slows one arm a tiny bit, and merges them again. Waves that line up make bright, a one. Waves that cancel make dark, a zero. It's reliable, but it's about a millimeter long.

The smaller option is a silicon ring, tens of microns across, sitting beside the waveguide. Nvidia picked it for its switches because it's tiny and uses very little power. Jensen showed one on stage and explained how the ring turns "this direct continuous laser beam into ones and zeros." His verdict: "that's the miracle."

That ring is a micro ring modulator. Think of it as a revolving door tuned to one color. When the ring matches the laser's color, it pulls that light in and the waveguide goes dark. Nudge it off tune and the light slips past.

The catch is heat. One degree of warming shifts the ring's color, so every ring needs a little heater to hold it on tune.

Because each ring grabs only its own color, a row of rings can share one fiber. Picture colored lanes on one highway: different colors of light don't mix in the glass. Lightmatter's design puts 16 colors on one fiber for 1.6 terabits a second.

Move the airport

Those plugs sit at the front of each machine, so the signal still crawls across tens of centimeters of copper. Co-packaged optics moves the light parts onto the chip's package, the little housing the chip sits in. It's like moving the airport from outside the city onto your office roof. The long copper drive almost disappears.

Nvidia's numbers from 2025 show the payoff: each connection drops from about 30 watts to about 9. And that's already for sale. Broadcom has been shipping Davisson, a co-packaged switch that moves 102.4 terabits a second, since October 2025. A switch is the traffic hub that lets thousands of chips act as one computer. Nvidia's Spectrum X photonic switch reaches 409.6 terabits a second, and Nvidia says it's in production as of mid 2026.

Jensen spelled out what that buys. "In a data center, we could save tens of megawatts," he said. "6 megawatts is 10 Rubin Ultra racks." By his math, 60 megawatts is 100 of them.

A data center can only pull so much power from the grid, so the network and the chips share one fixed budget. My read: energy is the real limit on AI. More energy means more compute, and more compute means smarter AI.

Who's building it

Lightmatter started out building chips that do math with light. This July, Harris said the bottleneck for computing has shifted. "It's no longer about compute. It's really about how you move data and how you create light that allows you to do that."

Lightmatter's Passage M1000 links chips with light at 114 terabits a second. For now, it's a developer kit. Ayar Labs, valued at $3.75 billion this March, puts a small light chip inside the GPU's own package, and partners are testing it.

Celestial AI aimed at the links between GPUs themselves. Co-founder Dave Lazovsky said on Bloomberg in March 2025 that Nvidia's NVLink, today's gold standard, runs over copper. "With the photonic fabric, we are changing that," he said. "We're moving information optically at a fraction of the energy and at much higher bandwidths and lower latency."

The chip company Marvell bought Celestial AI in a deal that closed February 2nd, 2026, for about $3.25 billion upfront. Light is winning the moving job, and each new product brings it closer to the GPU.

Math through stained glass

Job two is doing AI's core math with light. That math is called matrix multiplication, and a matrix is a big grid of numbers, the kind an AI model learns in training.

Harris laid out why light fits. "The equations that govern light and the physics of light, they're called Maxwell's equations," he said. "And if you break down what they're really doing, they're doing adds and multiplies."

Think of light passing through a stained glass window. Each pane dims the light by its own amount, so whatever comes out the other side is already the answer. On a photonic chip, the panes are a mesh of tiny light splitters, each set to one number from the grid.

It works in the lab. In April 2025, the journal Nature published two light-based processors on the same day. Lightmatter's ran small versions of well-known AI models at close to a normal chip's accuracy. Lightelligence's chip, PACE, used a 64 by 64 grid and finished each step in about 5 billionths of a second.

But these are lab chips and early pilots. The startups chasing light-based math are targeting 2028 to 2030 for products at scale.

The walls

Size is the first wall. A GPU does more than 100 times the parallel math of a 64 by 64 light grid.

AI also needs steps that bend the numbers, like turning every negative into zero. Plain light can't do that, so the signal hops back to electricity. There's no light-based memory either. The grid's numbers are held by tiny heaters, and each one burns power nonstop.

And light works in analog: each answer is a brightness on a smooth scale. That tops out around 8 to 10 bits of precision, a few hundred to about a thousand levels it can tell apart. PACE measured 7.61 bits. That's enough to run a trained model, but it's hard to train one that way.

Then every answer crosses back to digital through a converter, a circuit that turns brightness into a number. Think of a customs checkpoint at every border. The light flies across the country, then waits in line at each crossing. Converters burn energy on every number, so light needs much bigger grids to pull ahead, hundreds to thousands wide in most analyses. Today's lab chips are 64 and 128 wide.

The lasers turn only about 10% of their electricity into light, and Lightmatter itself posted that lasers are the next bottleneck. Copper still wins the shortest hops, too. Nvidia's NVL72 rack wires its GPUs with more than 5,000 copper cables.

So for now, light moves data between machines, and electricity does the math.

Where the saved watts go

That split is still a big deal, because moving data ate most of that A100's working energy. Less power per answer means cheaper compute.

My bet is that cheaper compute won't sit idle. When something useful gets cheaper, people use far more of it, the way coal use took off once steam engines got more efficient. So every watt saved on the wires gets soaked up by more AI. My read is that AI still doesn't have nearly enough compute.

Go back to Jensen's plugs, the parts that "didn't do any math." In the newest switches, light has moved onto the chip package in their place, and the power those plugs burned goes back into AI.

Light already won the job of moving AI's numbers between machines. Doing the math is the next fight, and it's still in the lab.