Case Studies: What Shipped

Ten modules of physics deserve one module of accountability. Billions of dollars have flowed into photonic computing over the past decade; this lesson asks the investor's blunt question — what actually shipped? — and reads the answers through everything the course has built. Five stories, chosen because each stakes out a different bet: a photonic-AI star that pivoted, an optical-I/O company that didn't need to, a quantum sampler that made headlines, a fab-scale quantum wager that refuses interim products, and a piece of photonics so successful it runs quietly in production without a press release. Watch for the pattern; it will assemble itself by the end.

Lightmatter: the compute company that became an interconnect company

Lightmatter, spun out of MIT in 2017, was the flagship of the photonic-AI thesis: its Envise accelerator put MZI-mesh matrix engines beside electronic control on one package — precisely the architecture of Module 5. Envise was demonstrated, benchmarked in marketing material… and never conquered the market. The company's centre of gravity shifted to Passage, a photonic interposer: a wafer-scale slab of programmable waveguides that other people's electronic chips sit on, giving dies optical bandwidth to their neighbours that copper traces cannot match. On the strength of Passage — interconnect, not computing — Lightmatter reached a multi-billion-dollar valuation. Read the pivot with this module's eyes and it is overdetermined: Envise carried the full burden of the conversion tax, the calibration treadmill and the benchmarking gauntlet, while Passage sells photonics' uncontested strength — moving bits — into the exploding GPU-to-GPU bandwidth crisis. Same waveguides, same fab, different physics being asked to do the work.

Ayar Labs: optical I/O, no pivot required

Ayar Labs never flirted with optical computing. Its product, TeraPHY, is an optical I/O chiplet: a small die speaking a standard electrical protocol on one side and driving optical fibre with microring modulators on the other, designed to be co-packaged beside a GPU or switch ASIC in the manner of Module 8's co-packaged optics. A separate module, SuperNova, supplies multi-wavelength laser light from outside the package — the lasers, the least reliable and most temperature-sensitive components, live where they can be cooled and replaced. The company assembled an investor list that reads like the chip industry's org chart (Intel, NVIDIA, AMD among them) and has spent its funding on the unglamorous work of qualification: reliability hours, yield, packaging. Note what the strategy concedes and what it keeps. It concedes computing entirely — every TeraPHY bit is digital at both ends, so there is no precision problem, no calibration-versus-accuracy trade, no benchmark ambiguity: just picojoules per bit and terabits per second, metrics a buyer can verify with a power meter. It keeps photonics' killer application. Ayar Labs is what betting only on Module 8 looks like.

Google's optical circuit switches: production, quietly

The least publicised case is the most deployed. For years, Google's datacenter fabrics and TPU supercomputers have used in-house optical circuit switches (OCS): boxes of MEMS-actuated micromirrors that physically steer light from any input fibre to any output fibre. This is Module 8's switching-fabrics lesson made corporate: a circuit switch, reconfigured in tens of milliseconds when a training job is scheduled or a topology re-shaped — never per packet. Because the mirrors just redirect photons, the switch is transparent to data rate and wavelength: the fabric bought years ago carries each new generation of transceivers unchanged, a capital efficiency no electrical switch can offer, alongside lower power and cost than the packet-switched spine layer it replaced. Published retrospectives credit OCS with double-digit percentage savings on fabric cost and power at datacenter scale. Nobody calls an OCS a computer. It performs no arithmetic whatsoever — and it is, by deployed unit count and by revenue-weighted usefulness, arguably the most successful photonic information machine ever built. That sentence is this lesson's thesis in miniature.

The quantum bets: Xanadu and PsiQuantum

Photonic quantum computing splits into a show-something-now camp and a show-everything-later camp. Xanadu is the first: its 2022 Borealis machine — squeezed light pulsing through time-multiplexed loops, 216 modes — performed Gaussian boson sampling beyond any classical simulation of its day, a genuine quantum-advantage milestone you could rent over the cloud. The fine print from Module 9 applies: GBS is a sampling demonstration, not a useful computation, classical spoofing algorithms have chipped at the margin ever since, and Xanadu's own roadmap (its Aurora prototype networks modules toward fault tolerance) implicitly concedes that Borealis was a proof of physics, not a product. PsiQuantum is the second camp, run to an almost opposite philosophy: no intermediate devices, no cloud demos — a straight bet on fusion-based quantum computing at fault-tolerant scale, manufactured in a commercial semiconductor fab (GlobalFoundries), funded by billions in private and government capital on the argument that only a million-qubit machine matters, so only fab-scale photonics is worth building. One camp ships demonstrations and iterates in public; the other ships nothing by design and asks to be judged at the finish line. Both are coherent strategies; they cannot both be the right one, and watching which converges first is the field's best spectator sport.

The pattern

Player	The bet	Status	The lesson
Lightmatter	Photonic AI compute (Envise)	Pivoted to Passage interconnect	Conversion + calibration taxed the compute thesis; bandwidth paid
Ayar Labs	Optical I/O chiplets only	Qualifying with major chipmakers	Digital ends → verifiable metrics, no precision problem
Google OCS	MEMS circuit switching	In production at scale	Zero arithmetic, transparent to rate — pure photonic strength
Xanadu	GBS advantage now, fault tolerance later	Borealis demonstrated; product pending	Advantage ≠ usefulness; sampling is not yet a market
PsiQuantum	Fab-scale fusion QC, no interim product	Building; judged at the finish line	The purest long bet in the industry

The pattern: interconnect ships; compute pivots; quantum waits. And it is not an accident of personalities — it falls straight out of this course's physics. Moving bits asks photonics only for what it is superlative at (bandwidth, distance, transparency) and keeps the endpoints digital, so the benchmarking story is clean and the conversion tax is the product rather than a parasite: a transceiver is a DAC-to-ADC pipeline, priced honestly in pJ/bit. Computing with analog light, by contrast, drags in precision, calibration, and a fight against a digital incumbent that improves every year. The market found the seam this module has been drawing all along — it just found it with money instead of equations.

Every company above has produced a genuinely impressive demonstration; only some have produced products, and telling the two apart is a skill this course can sharpen. A demo optimises for one number under lab conditions: hand-picked chips, a graduate student per phase shifter, accuracy measured after the fact. A product must hit every number at once — yield, reliability hours, temperature range, cost, software — under the checklist discipline of the previous lesson. The photonic-computing literature and press are overwhelmingly populated by demos, and the gap is widest exactly where this module's taxes bite: a mesh that ran one model at 92% accuracy on the bench says nothing about ten thousand meshes holding calibration for three years in a hot aisle. Reliable tells of a real product: named paying customers, volume-fab partners, spec sheets with wall-plug denominators, and boring qualification milestones announced instead of benchmark records. Reliable tells of a demo wearing a product's clothes: "up to", "equivalent TOPS", and a roadmap slide where the interesting column is always next year's.

Veterans of Module 7 will feel a draught of déjà vu. In 1990, Bell Labs unveiled the first digital optical processor, built on SEED devices, to enormous press attention — and Module 7's post-mortem recounted how that programme collided with cascadability, power and a CMOS industry improving on an exponential. The 2017–2025 photonic-AI wave rhymes: a genuine physical advantage, spectacular demonstrations, and a grinding discovery that the surrounding system — conversion, control, memory — sets the price. But the rhyme is not a repeat. The 1990 wave left behind components (modulators, detectors, integration techniques) that became the telecom and datacom photonics of today; the current wave is likewise depositing a layer of capability — foundry PDKs, co-packaging, wafer-scale waveguides — that the interconnect business is monetising immediately, whatever becomes of the compute thesis. Optical computing keeps failing upward: each hype cycle's wreckage becomes the next decade's infrastructure. There are worse fates for a field.

Where this goes next

Five verdicts from the market; one remains — yours. The capstone hands you the spreadsheet: you will spec a photonic accelerator end to end, budget every line item this module has priced, and decide — with your own numbers — whether you would build it.