I’ve had a laptop try to take off on my desk. Mid-video-export, fans screaming, the bottom hot enough that I had to slide it onto a book. If you’ve done anything graphics-heavy on an older machine, you know the exact moment I mean — the fans kick up a gear and the whole thing just gets hot.
For years I assumed that was just “computers working hard.” It’s actually something a bit more specific, and a bit more interesting: it’s copper wire hitting its limits.
Every chip is built on a purely electrical foundation. Billions of tiny transistors switch on and off, and equally tiny copper wires ferry the signals between them, like traffic lanes carved into the silicon. That setup got us the entire digital age. But push enough data through those lanes — which is exactly what today’s AI chips are trying to do — and the wires themselves become the traffic jam.
Researchers have been chasing an alternative for a while: photonics, using pulses of light instead of electricity. And a team at Caltech just made a genuinely big step in that direction — bringing the near-lossless performance of long-distance fiber-optic cable onto an actual silicon chip.
What’s a Waveguide? (Think: Fiber Optic Cable, Shrunk Down)
Before anything else, let’s deal with the jargon. A waveguide is just a tiny highway for light.
You’ve probably heard of fiber-optic internet. Laser light goes in one end of a glass strand, bounces along the inside walls, and comes out the other end miles later — even around bends — without leaking out.
A photonic waveguide does the same job, just etched directly onto a flat chip instead of stretched across the ocean floor. It keeps the light on a defined path so engineers can route it, split it, merge it, or filter it. Wire enough of these paths together and you get what’s called a Photonic Integrated Circuit, or PIC. Basically: shrinking a technology built for continent-spanning internet cables down to something the size of your thumbnail.
Why Copper Is Running Out of Road
Copper isn’t a bad conductor — it’s just being asked to do more than it comfortably can. Push high-speed signals through it billions of times a second, and three things start working against you:
- Resistance and heat. Copper resists the flow of current, and that resistance turns straight into wasted heat — the exact heat your fans are fighting.
- Capacitance and crosstalk. Wires sitting close together briefly hold onto charge like tiny batteries, which blurs crisp 1s and 0s into each other.
- A hard bandwidth ceiling. There’s only so much data a physical wire can carry before it simply runs out of room, no matter how clever the electronics around it get.
Light doesn’t have any of these problems. Photons don’t generate resistive heat, and different beams of light can pass right through each other without interfering — which is a genuinely strange thing to get used to if you’re picturing wires. Light can also multitask: a technique called wavelength-division multiplexing lets several different colours of light travel down the same physical waveguide at once, each carrying its own separate stream of data, the same way several radio stations can share the airwaves by broadcasting on different frequencies.
What Caltech Actually Built
The catch with putting light on a chip has always been propagation loss — light scattering and leaking away before it reaches where it’s going. Fiber-optic cable avoids this because it’s made from extremely pure, extremely smooth glass. A flat silicon chip, fresh off a standard manufacturing line, is comparatively rough at the microscopic level, and every bump scatters a little more light out of the path.
A team at Caltech, led by physicist Kerry Vahala, tackled this from two directions at once. First, instead of etching the waveguides straight into silicon, they built them out of germano-silicate — the same type of glass already used in fiber-optic cable — directly on top of standard 8-inch and 12-inch silicon wafers, using the same lithography equipment chipmakers already use. Then they used the fact that this glass has a relatively low melting point to their advantage: heating it just enough to let it “reflow” and smooth itself out, melting away the microscopic bumps until the surface approached atomic-level smoothness.
The second trick is one I genuinely like as someone who studied maths: rather than running the waveguide in a straight line, they coiled it into a spiral — the same idea as winding a fiber-optic cable around a spool, just compressed down to fit on a wafer through nanofabrication. That lets light travel a much longer effective path while the whole structure takes up almost no physical space. It’s a neat bit of lateral thinking — if you can’t make the chip bigger, make the light’s journey longer without moving it further from home.
Put together, the result outperformed the previous best silicon-based waveguides by a factor of 20 at visible wavelengths — a genuinely large jump for a field that usually improves in much smaller increments.
Why a Tiny Amount of Light Loss Actually Matters
You might reasonably ask: a chip is only a few centimetres across, so does losing a sliver of light really matter that much?
It does, because light on a chip doesn’t just travel from A to B once. A lot of photonic devices use ring resonators — tiny circular tracks, often just millimetres wide, where light loops around thousands of times a second. Loop something a thousand times and even a small loss per lap adds up fast; keep the loss low enough, and the light can circulate for a genuinely long effective distance while staying “coherent” — keeping its wave pattern smooth and predictable, rather than degrading into noise.
That coherence is what precision lasers, sensors, and quantum-computing components all depend on. The Caltech researchers found that this relationship isn’t gentle, either: every tenfold cut in loss produces roughly a hundredfold improvement in how long a laser stays coherent. In practice, lasers built on this platform stayed coherent more than 100 times longer than earlier chip-based designs.
How Light Could Actually Replace Electrical Connections
To be clear, nobody’s building a chip that runs entirely on light — transistors are still far better at doing the actual logic and math. The likely future is a hybrid: electronics handle the thinking, photonics handles the commute. A typical trip looks like this: an electrical transistor generates the raw data, a modulator converts it into pulses of light, a photonic waveguide carries that light across the chip at very high speed, a photodetector converts it back into an electrical signal on the other end, and the receiving processor takes it from there.
Why AI Is Making This Urgent
A single chip was never really the bottleneck AI companies are worried about — it’s the tens of thousands of processors in a data centre that need to talk to each other constantly. Moving that data back and forth already eats a serious chunk of a data centre’s total power bill, mostly as heat lost over long copper traces.
That’s the pressure behind “co-packaged optics” — mounting the light converters immediately next to the processor itself, so data switches over to light almost as soon as it’s created, instead of travelling any real distance as electricity first. Less distance on copper means less wasted heat and more usable bandwidth, which at data-centre scale adds up to a genuinely large amount of saved electricity.
So Is Copper Going Away?
No — and probably not for a long time. Your laptop’s motherboard will keep using copper for power delivery and plenty of everyday wiring; it’s cheap, simple, and perfectly good over short distances.
What’s more likely to shift is the connections between chiplets — the small, specialised pieces that modern processors are increasingly built from instead of one giant slab of silicon. Wiring chiplets together with optical waveguides lets them exchange data almost as if they were sitting on the same piece of silicon in the first place, which matters more and more as chips get broken into smaller, more specialised parts.
Where Else This Could Show Up
Because Caltech’s waveguides handle both visible and infrared light well, the applications reach well past data centres:
- Atomic clocks — ultra-precise timekeeping that depends on very specific, stable laser frequencies.
- Quantum computing — controlling fragile qubits with coherent, stable light paths.
- Medical and environmental sensors — lab-on-a-chip devices that use light to detect specific chemicals, gases, or biological markers.
The Bigger Picture
For decades, “faster computer” mostly meant “smaller transistors, more of them.” That race isn’t over, but it’s no longer the whole story — increasingly, the question is how efficiently you can move an answer once you’ve calculated it, not just how fast you calculated it.
What Caltech has shown is that the same basic idea behind global fiber-optic internet — keep the light smooth, keep the loss low — works just as well shrunk down to the size of a fingernail. Future chips almost certainly won’t drop copper altogether. They’ll more likely run two systems side by side: electrical wiring for short, local jobs, and optical highways for the heavy, long-distance traffic. Watching a “wire” and a “beam of light” end up solving the exact same engineering problem, just at wildly different scales, is the kind of thing that makes me stop and actually appreciate the machine that’s currently trying to melt my desk.