Why AI runs on light and glass

We are creatures of light. 

Every organism on the surface of our planet, from Sapiens to slime moulds, has an intimate relationship with it. Light gave life its start, and powers it still.

And light gives sight. Almost everything that lives — or has ever lived — can respond to it through growth or movement. You can thank evolution for perfecting a sense that can feel, for us, miraculous.

Follow a cosmic arc through the atmosphere of our small blue planet, past galaxies to the edge of everything. Cup our universe in your hands if you can. What you’d see is light stitching together the fabric of space itself, setting a universal speed limit, and regulating the transformation of matter into energy. And we still don’t fully understand it. 

Then there’s glass. It occurs in nature, and we’ve been manufacturing it for millennia. Picture a towering cathedral window filled with images composed with stained glass. The illustrations are luminous and vivid, and light spills to make pools of colour on the polished stone floor at your feet. Like light, glass has its own special weirdness.

Those panes of cathedral glass may have stood for a thousand years, but the glass itself has never been motionless. Slowly flowing downhill, the panes are now thicker at their bottoms than their tops. The solidity of glass is an illusion.

Light and glass enjoy a symbiotic relationship. Glass changes the properties of light as it moves through it. And light influences the nature and experience of glass. Entwined, dancing, spooning — scintillating. No other material comes close to glass for sheer mutability and range. 

Enter Corning

We’d argue that no business on the planet understands this better than Corning Inc. They manufactured the first lightbulbs for Edison, put windows in every crewed NASA spacecraft, and built the mirrors for every NASA space telescope. That’s more than diversification. That’s a worldview. 

Corning already had more than a hundred years of innovation behind it when it invented the world’s first low-loss optical fibre in 1970. Laser light — the most ‘together’ light known to science — could be pulsed into the fibre, channeled to exactly where it was needed, following curves and covering long distances without amplification. It was a game changer for telecommunications. Apologies to George Lucas, but this was the original industrial light and magic. Optical fibre, like computing, like the Internet protocol, became foundational to the information revolution. 

To make an optical fibre, you start with incredibly pure glass, then add back trace materials that tune its physical and optical characteristics. The aim is to let light travel further and faster — with minimal distortion, scattering, or attenuation. The glass also needs to be strong under tension, but flexible enough to bend without breaking. That part is pure, Nobel-standard materials science.

Picture a pencil — a cylinder of graphite at the core, surrounded by a cladding of wood. The arrangement is the same in an optical fibre, but core and cladding are both glass, each with slightly different optical properties.

Both go into a “preform” for manufacturing. Think of the glass pencil, but at the scale of a drainpipe. The preform enters a drawing tower where the tip is heated to 2000°C, and a hair-thin strand of core and cladding is drawn continuously downwards from the softened tip — its diameter measured and controlled in real time. Before the fibre can be scratched or contaminated, it gets a polymer coat, cured with UV light and spooled. That part is high-tech manufacturing.

One third glass

Roll forward to 2026 and AI, the totemic technology of our age

Want a recipe for an AI’s brain that a barista could memorise? One third compute, one third memory — and one third glass. 

What about software, training? What about energy?  Yes, there’s more to an AI brain than that. But there’s an essential truth in the recipe too.

Without glass — optical fibre — there is no neural network. GPUs make the headlines, but optical fibre makes the connections. A GPU without connections is just a cluster of brain cells floating in a petri dish — marvellous in their own way, but practically useless.

In operational AI datacentres today, GPUs communicate at 400 Gbit/s (around 400 to 4000 times faster than home broadband). That’s fast. But 800 and 1600 Gbit/s (or 1.6 Tbits/s) are already emerging, and 3.6 Tbit/s is on the horizon. 

Those bandwidths are achieved with parallel lanes of fibre into every GPU — 8 fibres to transmit, 8 more to receive — each fibre carrying a fraction of the total. Every GPU needs to connect to lots of other GPUs to form a neural network. That’s one huge, extremely dense mesh. 

You can’t build that mesh by taking fibre off the spool, splicing on connectors and making links strand by strand. Modern optical fibre is as much about systemisation as science. So much glass. If you had a superhero’s x-ray eyes tuned only to glass, you’d see plains, rivers, waterfalls of it.

Optical fibre itself has become thinner and far more flexible: it can be wrapped around a regular pencil and still perform — not that you would. Connectors have become smaller and smarter, fitted at the factory, not the field. Cables carry many thousands of fibres in shrinking outer dimensions. The sprawling mesh becomes a practical geometry that humans can actually manipulate and manage.

Closer to the heat

But it won’t stop here. Imagine if you could lift the roof off a datacentre and pour in glass that organised itself — finding its own connections, weaving its own mesh. We’re not there yet. But we’re closer than you’d think, and the direction of travel is unmistakable: more glass, denser glass, glass closer to where the heat happens.

Take multicore fibre — more than one light-carrying core inside a single strand. It’s already a reality, and the first real step toward that self-organising vision: glass doing more with less, boosting bandwidth while saving space, materials and cost. Co-packaged optics is next, taking glass beyond the front of the server and onto the chip itself — eliminating the copper connections and the conversions between them, saving a lot of energy in the process. All of this in service of machines learning to think.

We started as creatures of light. It turns out our machines are too.