Silicon Photonics and the Future of AI Scaling — John Bowers

ELI5/TLDR

The wires inside a data center are running out of road. Copper can only carry so much data before it overheats and starts losing the signal, and modern AI chips need to move staggering amounts of data — far more than copper can handle. The fix is to send the data as light instead of electricity, using tiny laser-and-fiber circuits built right onto silicon chips. John Bowers spent forty years figuring out how to put a working laser on silicon — which silicon, on its own, is famously terrible at — and that quiet achievement is now becoming the backbone of how AI computers talk to each other.

The Full Story

The problem: copper hit a wall

Start with the thing everyone forgets when they picture a data center. It is not a warehouse full of separate computers. It is one enormous computer, spread across a warehouse, where a million chips all have to talk to each other constantly. When you type a question into an AI, the work gets split across hundreds or thousands of those chips, and they spend most of their effort not computing but coordinating — shuttling data back and forth.

That shuttling is the bottleneck. A modern AI chip needs to move tens of terabits of data on and off itself. (A terabit is a trillion ones-and-zeros per second. One such chip moves more data than the entire early internet.) Today, chips sitting next to each other on a shelf are connected with copper wires. And copper, it turns out, is a leaky bucket.

Here is the physics, gently. To push data faster down a copper wire, you flip the voltage on and off faster — higher “clock speed.” But a copper trace at very high speeds behaves less like a clean pipe and more like a wire that fights you. The faster you push, the more signal leaks away as heat. Bowers gives a brutal number: you can lose 95% of your signal — what engineers call 20 decibels of loss — over just one meter of copper at these speeds.

“You might literally have 20 dB of loss over just even a meter of of copper.”

And it gets worse, because pushing harder costs power, and that power is proportional to how much wire you are charging up. More wire, faster flipping, more electricity burned as heat. In a data center already venting kilowatts of heat per rack, that is a dead end.

Why light wins

Now the magic of doing it with light. When you send a signal as photons down a glass fiber, the loss barely changes whether you travel one millimeter or ten kilometers. Think of it like the difference between shouting across a room (your voice fades fast, and the bigger the room the more you strain) versus shining a flashlight (the beam reaches the far wall just as easily as the near one).

“Phetonics, as you know, the loss is the same whether you go a millimeter or 10 kilometers.”

And the energy cost is fixed per photon — roughly one electron-volt to make a particle of light, the same whether it travels a hair’s width or across the building. So light lets you connect not just neighboring chips but chips kilometers apart, all at the same low cost. That is why Bowers thinks every new data center build will be “photonically integrated” — light woven directly into the chips, not just the long-haul cables.

Silicon’s embarrassing flaw

Here is the central irony of Bowers’ whole career. Silicon is the perfect material for guiding light — you can carve channels in it (waveguides) that lose almost nothing, and you can make them in the same ultra-precise factories that already stamp out billions of computer chips. But silicon is hopeless at making light.

The reason is a quirk of its atomic structure called an “indirect band gap.” You do not need the details; you need the consequence. Inject electricity into silicon hoping for light, and you get one photon for every million electrons. It is like a light bulb that converts almost all its electricity into nothing useful. Other materials — the so-called “III-V” compounds like gallium arsenide and indium phosphide — are over 90% efficient. They are brilliant emitters. But they cannot be mass-produced cheaply the way silicon can, and they break if you grow them in the wrong place.

So Bowers’ field faced a choice: make silicon emit light (decades of effort, mostly failed) or somehow marry good silicon to a good light-emitter. Bowers bet on marriage from the start.

“It was clear from the beginning that you had to have a direct band gap material.”

The marriage trick: heterogeneous integration

The solution he helped pioneer, and which Intel turned into a billion-dollar product, is called heterogeneous integration. The word just means “gluing unlike things together,” and the trick is elegant.

You grow a wafer-thin layer of the good light-emitting material (indium phosphide) on its own native crystal, where it grows clean and defect-free. Then you flip it onto a silicon wafer and bond it down — chemically activating both surfaces with an oxygen plasma so they stick — and dissolve away the original substrate. What is left is a one-micron sliver of perfect light-emitting crystal sitting on top of silicon. Now you finish everything else in the silicon factory: lasers where you need light, modulators, detectors, all of it.

The clever economics: the good emitter material is expensive, and most of a photonic chip — the waveguides, the multiplexers, the fiber connections — does not need it. So Intel only places it where light actually has to be made. They take a handle wafer, stick 5,000 tiny chiplets of emitter on it, flip and bond them all at once. Because every later step is etched in alignment with the underlying silicon wafer, the placement does not have to be precise — within tens of microns is fine.

“They’ll take a handle wafer and put literally 5,000 chiplets on it and flip it over and bond it all at once.”

There is a defect-tolerance bonus too. One kind of light-emitter, quantum dots — picture tiny islands of crystal a few atoms across — actually freezes defects in place rather than letting them spread, so quantum-dot lasers survive abuse that kills ordinary lasers. A recent Sandia test found them ten times more radiation-resistant than the best commercial alternatives. Good news if your data center is in space, or near a nuclear test.

A surprise gift: the world’s quietest lasers

The conversation takes a turn into something Bowers clearly delights in. Semiconductor lasers have always had a reputation for being noisy — jittery in their frequency, like a slightly out-of-tune note that keeps wobbling. Gas and solid-state lasers were thousands of times steadier. For precision work, nobody would touch a chip laser.

Then his group, working under a DARPA contract that demanded everything fit in a tiny package, was forced to skip a component everyone normally includes (an isolator, which keeps stray reflections out of a laser). With no room for it, they soldered the laser right up against a high-quality glass resonator — a tiny loop where light circles around and around. And something unexpected happened.

A whisper of light scatters backward out of that loop and re-enters the laser, and that faint feedback locks the laser steady. It is called self-injection locking. The intuition Bowers offers: if the loop stores an enormous amount of light energy (these loops have “Q factors” — a measure of how long light lingers — in the billions), it utterly dominates the small laser, dragging it into line.

“The energy in that cavity is so huge compared to the energy in the gain region … that it just dominates everything.”

The numbers are staggering. A chip laser that started life jittering at a megahertz of linewidth got pulled down, step by step, to one hertz. That is a wobble of one part in hundreds of trillions — the longer and lower-loss you make the light’s path (the sweet spot is around four meters of path, spiralled into a centimeter of chip), the quieter it gets. There is a trade-off, charmingly inverted: the narrower (quieter) the laser, the bigger the chip area it needs.

Rebuilding the optical lab on a chip

This is the deeper theme. For thirty years, the gear of a precision physics lab — ultra-stable lasers, reference cavities, frequency combs, nonlinear crystals — lived on vibration-isolated optical tables the size of dining tables. Bowers’ field is quietly shrinking all of it onto silicon.

Take the frequency comb — one of the show’s harder ideas, so go slowly. Imagine a single laser, but instead of one pure color it emits dozens or hundreds of colors, all perfectly evenly spaced, like the teeth of a comb laid across the spectrum. Why is that useful? In a data center, you currently bolt together many separate lasers to send data on many colors at once (more colors = more parallel data down one fiber). Managing 64 separate lasers — each temperature-controlled, each spaced just right — is a nightmare. A single comb gives you all 64 colors from one source, perfectly spaced, for free.

The comb is built from solitons — and Bowers gives the cleanest possible definition. A soliton is a pulse that refuses to spread out. Two forces fight inside the loop: dispersion wants to smear the pulse, nonlinearity wants to pull it together. At exactly the right power they cancel, and the pulse holds its shape — the same physics as a tsunami that crosses an ocean without dissipating. The breakthrough was making this turnkey: old comb setups needed a PhD twiddling knobs to coax the system into a clean single-soliton state. Self-injection locking makes it snap into that state the instant you switch it on.

Once you can make these ultra-quiet combs and lasers on a chip, the applications spill out. The best atomic clocks in the world are now optical, not microwave — and Bowers expects them shrunk onto inexpensive chips. Why care? Because knowing time precisely lets you know position precisely (velocity times time equals distance). A chip-scale optical clock plus an optical gyroscope means a car could navigate accurately even with no GPS — which matters for self-driving, and matters a great deal somewhere like a GPS-jammed war zone.

He is equally bullish on LiDAR — the laser “eyes” that let a Waymo see the road. Today they are big and expensive. Put the whole thing on a photonic chip and every car could carry several, measuring the velocity of objects looming out of fog 300 meters away.

The big architectural shift: switching with light

The thread that ties back to AI. In the old design, a data center’s top-level switches were electronic — meaning every signal had to be converted from light to electricity, switched, then converted back to light on a fresh laser. Wasteful. Optical switching skips the conversion entirely: the light just gets routed, optically, with only a couple of decibels lost.

Bowers founded a company, Calient, on exactly this idea — in 2000, during the telecom bubble, when 60 companies were chasing optical switching. The bubble burst; by 2002 only three survived. He calls himself “25 years too early.” But the argument was right, and Google has now ripped out that whole electronic switching layer and replaced it with optical switching across its data centers.

The reason it matters for AI is failure. With a million processors, even a tiny failure rate means several chips and transceivers die every single day. An AI training run needs all its processors present the whole time. Optical switching lets you instantly route around a dead chip and swap in a spare — without a technician walking over to pull a card.

“You need to be able to switch out failed units. And that’s what the optical switching can do.”

This is also, he notes dryly, the real problem with the fashionable idea of putting data centers in space for free cooling. A million processors means a million times the chance of failure — and nobody is flying up to swap a card. (An interviewer volunteers for the job. Bowers warns he might get bored.)

Where it’s all heading

The frontier now is higher integration — pushing the optics ever deeper inside the chip. The progression: pluggable optics on the board edge → optics ringing the switch chip → optics inside the chip package (Broadcom and Intel are shipping this, 50- and 100-terabit chips with light built in) → eventually light carrying data between regions of a single chip, even across distances as small as 100 microns, because past a certain point light just wins everywhere. Capacity keeps roughly doubling every couple of years, by some combination of more lanes, faster devices, and more colors of light.

A telling business aside: Bowers favors keeping the fast electronics on one chip and the photonics on another, bonded together — because photonic devices are large (light’s wavelength is a whole micron, versus nanometre-scale transistors), so building them in an expensive cutting-edge process wastes money. Run the photonics in an older, cheaper 45-nanometre factory and it costs a hundredth as much per square millimetre. Those aging fabs, abandoned by the latest processors, get a new reason to exist. The true tipping point, he says, will come when TSMC — the giant that makes most advanced chips — decides to integrate lasers. “Everyone will be doing it at some point.”

Key Takeaways

Copper has hit a physical wall. At modern clock speeds (100+ GHz), copper traces lose ~95% of the signal over a meter, and the power needed scales with wire capacitance. AI chips moving tens of terabits cannot be fed by copper.
Light’s killer feature is distance-independence. Optical loss is nearly the same over a millimeter or ten kilometers, and the energy cost is fixed per photon. This lets you connect chips across a whole warehouse at one chip’s cost.
A data center is one computer, not many. Modern AI work fans out across thousands of chips that spend most of their energy coordinating — which is precisely the traffic photonics carries best.
Silicon guides light beautifully but emits it terribly (indirect band gap → 1 photon per million electrons). The fix is heterogeneous integration: bond a thin layer of efficient III-V emitter onto silicon, then finish in a standard chip fab.
Heterogeneous integration scales because you only place the expensive emitter where needed — Intel bonds 5,000 chiplets at once, alignment tolerant to tens of microns because later steps register to the silicon wafer.
Quantum-dot lasers tolerate defects (they freeze dislocations in place) and are ~10x more radiation-hard — useful for harsh environments, and they need no isolator because they shrug off reflections.
Self-injection locking turns noisy chip lasers into the world’s quietest — a high-Q resonator’s faint back-reflection drags the laser to as low as 1 Hz linewidth. The catch: narrower lasers need bigger chip area (~cm² for a 4-meter folded path).
Frequency combs replace racks of separate lasers — one comb gives dozens of perfectly-spaced colors (WDM channels) from a single source. Made turnkey via self-injection locking that snaps the system into a clean single-soliton state.
A soliton is a pulse that won’t spread — dispersion and nonlinearity cancel at the right power; same physics as a tsunami.
Optical switching is finally winning after 25 years. Google removed the electronic spine-switch layer; routing light directly (no light→electric→light conversion) saves power and lets you swap out dead processors — critical when a million-chip cluster loses several units daily.
Chip-scale optical clocks + gyroscopes enable GPS-free navigation — relevant to self-driving and jammed/denied environments.
Divide and conquer beats monolithic integration. Keep fast electronics and large photonics on separate chips bonded together; running photonics in an old 45nm fab costs ~1/100th per mm². The real inflection comes when TSMC integrates lasers.

Claude’s Take

This is a dense, generous, genuinely expert conversation — the interviewers (from 632nm, a photonics-nerd channel named after the helium-neon laser wavelength) clearly did their homework, and Bowers answers like someone who has nothing left to prove and enjoys explaining. There is almost no hype and no salesmanship; when he says he was “25 years too early” with Calient he is laughing at himself, not spinning the failure.

What makes it worth the time is that it threads a real arc: the same underlying win — putting a good laser on silicon — quietly powers four different futures (AI interconnects, ultra-stable clocks, LiDAR, quantum). That is the mark of a foundational technology rather than a point solution, and Bowers is credible because he has the products and the billion-dollar Intel line to back the claims, not just slides.

The one thing to keep in perspective: this is an optimist describing his own field, so the timelines (“every car will have several LiDARs,” “TSMC will integrate lasers”) are aspirational, and the harder applications (optical computing, chip-scale quantum) are explicitly hedged — even he admits optical switching took 20 years to arrive partly because electronics kept refusing to stand still. The honest, recurring lesson buried in the chat is that silicon CMOS is a moving target that has killed many “obviously better” alternatives by simply improving faster. That is the BS filter applied to his own enthusiasm, and to his credit he applies it himself.

Score: 8/10. Substantive, accurate, well-explained, and unusually candid about failure and timing. Loses a couple of points only because it is long, occasionally drifts into shop-talk (etch chemistry, contact metals) that a general listener will skim, and offers more breadth than depth on the AI-scaling angle the title promises.