Harnessing the Light: How AI Actually Works, If You're Extra Curious

Start with a light switch. Flick it up, the bulb is on. Flick it down, it is off. On or off, and nothing in between. Now imagine a switch far too small to see, so small you could fit tens of billions of them on a chip the size of a fingernail, and imagine each one flipping not once but billions of times every second. That is a computer. The whole towering business of artificial intelligence is built on that one humble idea: a switch, repeated at a scale that stops feeling physical and starts feeling like thought.

You do not need to know any of this to put AI to work. This is the extra-curious tour, the one for people who like knowing what is actually happening inside the box. And the reason it is worth taking is that the honest version is more interesting than the hype, and it quietly tells you a great deal about what good AI should cost and why.

The switch that thinks is electrical

The tiny switch has a name: a transistor. It is a switch for electricity, with no moving parts, that turns a current on or off when you tell it to. On is a 1. Off is a 0. That is the entire alphabet a computer has to work with, and everything else, this article, a photograph, a language model that can hold a conversation, is built out of enormous piles of those two numbers.

The reason a computer feels clever is not that any single switch is clever. It is speed and quantity. A modern processor packs tens of billions of transistors and flips them billions of times a second, and by arranging those flips just so, it can add numbers, compare them, and follow instructions. Stack enough simple decisions on top of each other, fast enough, and you get something that can recognise a face or draft a paragraph. It was never one big brilliant switch. It was always a mountain of dumb ones, switching in concert.

Here is the part worth being precise about, because plenty of people get it wrong: today, the thinking is electrical. The switching that does the math is electrons moving through silicon, not light. If someone tells you your laptop computes with light, they are ahead of the science. But light is very much in this story, and its role is growing fast. It just starts somewhere else: not as the thinker, but as the courier.

Light is how the answer reaches you

When you send a message to an AI, your words do not stay in the room. They travel, often thousands of kilometres, to a data centre where the real work happens, and then the answer travels all the way back, usually in less time than it took you to lift your finger off the key. What carries them is light.

Almost all long-distance data moves as pulses of light down hair-thin strands of glass called optical fibre. Picture a lighthouse blinking a message in code, on, off, on, but blinking billions of times a second, and instead of a beam across the harbour it is a ray of light racing down a glass thread. Light in fibre travels at roughly two-thirds of its speed in a vacuum, around 200,000 kilometres every second, which is why a question can cross a continent and come back before you have taken a breath. Your prompt becomes light, flies down the glass, becomes electricity again at the far end, gets thought about by all those switches, and the answer makes the return trip the same way.

For most of computing history, light stayed outside the machine, in the cables between buildings and cities. That line is now being crossed, and it is one of the most important shifts in AI infrastructure happening right now.

The new part: light is moving in among the chips

Modern AI runs on thousands of specialised chips working in parallel, and the bottleneck has quietly stopped being the chips themselves. It is getting the data between them fast enough, without melting the power budget. Copper wires, the traditional way, run hot and lose steam over distance. Light does not.

So the industry is now putting the light much closer to the silicon. NVIDIA, whose chips run most of the world’s AI, has been rolling out what is called co-packaged optics, where the light-handling parts sit right next to the processor instead of in a plug on the edge of the board. The company presents this shift as essential to scaling the AI data centres now being built, not an optional upgrade, and points to roughly 3.5 times lower power for moving the same data compared with the older electrical approach. That power figure is the whole story in a number. At the scale AI now runs, the cost of an answer is largely the cost of the electricity to produce and move it, and light moves it for a fraction of the energy.

This is the same theme as where your AI actually runs and the day I burned a week’s budget, seen from the physics end: the eye-watering efficiencies and the eye-watering bills both come from infrastructure choices, not from anything mystical about intelligence.

And the frontier: light that does the math

Here is where it gets genuinely futuristic, and where I will be careful to mark what is shipping today versus what is still on the bench.

It turns out light can do more than carry information. It can compute with it. The core operation inside a neural network is a particular kind of multiplication, done billions of times, and it happens that when beams of light pass through a carefully designed optical system, the way they interfere and combine performs exactly that multiplication on its own, at the speed of light, using almost no power for the calculation itself. The physics does the arithmetic for free.

This is no longer just theory. In April 2025, the company Lightmatter published work in the journal Nature demonstrating a chip that runs real AI workloads with the core math done in light, steering hundreds of light beams through photonic circuits to classify images and generate text. First-generation photonic computers are expected around 2026 and 2027, with the broad view being that wide adoption is still years out, into the early 2030s. So this is a frontier, not a product you can buy for your business tomorrow. But the direction is clear enough that the biggest names in chips are pouring money into it: the switch that thinks may, in time, become a switch made of light.

Why a business owner should care at all

You can run a thriving operation on AI and never think about a single photon, and that is completely fine. So why tell you any of this?

Because it dismantles the two unhelpful stories people carry about AI. The first is that it is magic, a black box that either works or does not, beyond understanding. The second, the reaction to the first, is that it is all smoke, a trick with no substance. Neither is true. AI is electricity switching and light travelling, arranged with enormous care by people who know what they are doing. It is physics and engineering all the way down.

That matters to you as a buyer, because everything governed by physics and engineering can be measured, budgeted, and improved, and can also be done badly. The reason one AI system costs ten times what another costs to do the same job is almost never the laws of nature, which charge everyone the same. It is the quality of the engineering wrapped around them. This is the identical point we make in why AI is not the one-click tool it looks like: the effortless surface is real engineering underneath, and the engineering is the part that decides whether the thing is quick, cheap to run, and trustworthy, or none of those.

What the light is really telling you

So the next time you type a question and an answer appears as if from nowhere, you can picture what actually happened. Your words became pulses of light and raced down a glass thread at 200,000 kilometres a second. They arrived somewhere far away and became electricity again. Tens of billions of tiny switches flipped billions of times to turn them into an answer. That answer became light once more and flew back to you, all faster than you could blink. Nothing about it was magic. Every step was somebody’s careful engineering decision.

That is the quiet lesson under all the physics. The intelligence you rent rides on light and electricity that someone shaped on purpose, and the difference between a system that delights you and one that drains you is not the light or the electrons, which behave the same for everyone. It is the judgment of whoever built the thing. Electrons do the switching, photons do the carrying, and engineering decides what it all adds up to. That last part is the only one you get to choose, and it is the one that matters most.