Photonic compute

Light Standard Technology · Lightmatter · Lightelligence (Shanghai Xizhi) · Neurophos · Arago · CogniFiber · StrAilight · Celestial AI · LightSolver (Quantum Interconnects division) · GlobalFoundries (SiPh) · Lessengers · Xscape Photonics · Lucidean · Drut Technologies · aiXscale Photonics · PhotonicX AI · Advanced Micro Foundry · Mixx Technologies · Aloe Semiconductor · Aeluma · Ayar Labs · LightSolver

• 21 Jul 2026A second company independently reaches the same physical limit: you cannot make an optical component much smaller than the wavelength of the light in it, so the usual route to improvement is closed. They add a constraint nobody else raised, how completely you can block light, and a question worth asking anyone in the field: there were silicon photonics booms in the eighties and nineties that went nowhere, so what exactly have you cracked that they had not?
• 21 Jul 2026The clearest reason photonics is entering the server rack now: at 200 gigabits a second, a copper cable only reaches about a metre, and doubling the speed halves that again, so the cables are simply too short to wire the machine together. Also a number worth knowing, the laser is a quarter to half the cost of an optical link.
• 21 Jul 2026The strongest argument for computing with light, from a founder doing it: the conversion cost people object to is fixed, but the useful work grows with the square of the problem size. On that maths bigger models suit optical better, not worse, which is the reverse of the usual assumption.
• 21 Jul 2026The counter to the optical-computing case, from someone building optical networking rather than optical compute: nine tenths of the cost of running a model is fetching data from memory, so making the arithmetic faster speeds up the part that was never slow. If that is right, the whole category is optimising the wrong thing.
• 21 Jul 2026The founder of one of the few companies still computing with light gives the number that bounds the whole field: past about three bits of precision you cannot trust the answer coming out. His response is to design around it rather than fight it, and that is also why his company avoided the pivot everyone else made.
• 21 Jul 2026Three companies at three different layers of the stack independently name the same obstacle, and it is not performance: proving the parts still work after three to five years in a rack. The industry measures it as failure-in-time, and the answer is redundancy rather than perfect components. Worth asking about before any speed claim.
• 21 Jul 2026The company that builds the design tools for photonic chips says the real obstacle is not the maths, it is size: a component cannot be much smaller than the wavelength of the light it carries, and that is physics rather than engineering. On that view no amount of manufacturing progress closes the gap with a conventional chip.
• 21 Jul 2026An investor who did deep diligence on Lightmatter gives the same verdict from the opposite direction: nearly every company that set out to compute with light has quietly moved to moving data with light instead. The reason is that too many steps still have to drop back into electronics, and that hand-off sets the pace for everything else.
• 21 Jul 2026The counter-argument, from an investor backing an all-optical chip company: trying to build a halfway machine is the riskier path, because Nvidia is adding optics to its own chips and will be a hundred times faster within two generations. On that view you either go for the full paradigm change or you lose the race. He also thinks the memory problem that killed earlier attempts is solvable, and points at a use for light-based chips nobody else raised: they are immune to electromagnetic interference, which matters in vehicles, satellites and on a battlefield.
• 21 Jul 2026Every well-known photonic computing company has ended up selling data movement rather than computation. A founder in the field explains why: to beat an Nvidia core you have to win on whole-system energy, not on the chip in isolation, and nobody has managed it. The gap is concrete. Photonic chips handle matrices of 8 to 32 elements where an Nvidia core handles 256.
• 3 Jun 2026A photonic take on probabilistic computing is emerging (several new photonic concepts at the 2026 probabilistic-computing conference): quantum-optical entropy instead of magnetic or electronic noise, argued on controllability and manufacturability rather than raw speed. The more convincing differentiation sits in the compute fabric and compiler (flexible topology, native higher-order interactions, constraint-handling without penalty terms), not the entropy source, and as with optical inference, no one yet publishes a system-level rather than component-level benchmark.
• 1 Jun 2026Another team building general AI inference in indium-phosphide photonics concedes the architecture is the same as everyone else's and competes only on time-to-market, with no shipping system-level benchmark yet (the one full-system analysis is a 2022 MNIST simulation that itself shows roughly 6x conversion-and-control overhead). That keeps intact my call that most optical AI inference is a crowded trap, and that the durable value sits in the enabling layer (modulators, nonlinear materials) and the workload-specific wedges, not the full-stack processor.