The Public Landscape

Topics We'll Cover

• How to Read a Company's Claim
• Lightmatter
• Lightelligence
• Q.ANT
• Newer Compute Start-ups
• Encryption, Research Labs and Universities
• Interactive: What a Claim Implies
• The Interconnect Companies
• What Has Been Shown, and Where
• The Pull Towards Interconnect
• Questions to Ask
• What to Take Away

How to Read a Company's Claim

Most of what is public about optical AI companies comes from the companies: press releases, product pages and funding announcements. This deck quotes them, links them, and labels each as company claim. A few companies have also published peer-reviewed papers, labelled peer-reviewed. Nothing here has been independently benchmarked by this series.

Coverage is a snapshot as of October 2026, drawn from public sources found by search; companies not listed may simply have published less. Funding figures are included only where a public announcement states them.

Lightmatter

• Reading the paper's figures: 65.5 trillion operations per second over 79.6 W (electrical plus optical) is about 0.82 TOPS/W, or about 1.2 pJ per operation for the whole system (this deck's arithmetic). The precision is a block floating-point format reached with analogue cores, which is why the paper reports accuracy on real models rather than an ENOB alone.
• The interactive on slide 07 starts from these figures.

Lightelligence

• The pattern: its peer-reviewed result targets latency on an iterative solver, the niche of OptAI 03, slide 08, and it also sells optical interconnect inside the package. Both Nature papers of April 2025 (this and Lightmatter's) appeared in the same issue.

Q.ANT

• What is not public: no peer-reviewed paper on the processor was found for this deck, and the releases do not state the baseline, precision or workload behind the 30× and 50× figures. They are the company's own and are best read as targets until measured independently.
• Thin-film lithium niobate gives fast, low-loss phase modulation without free carriers (thin-film lithium niobate), a route to fast weight and input writing.

Newer Compute Start-ups

• None of these has a peer-reviewed system paper found for this deck, and the figures accompany funding or launch announcements. They are targets and positioning, stated by the companies.
• The interactive on slide 07 shows what any such claim must assume about array size, precision and converters to hold together.

Encryption, Research Labs and Universities

• Academic figures are the authors' own too, but peer review means the method and the definition of an operation are on record and can be checked. That is the main difference from a press release.

Interactive: What a Claim Implies

Enter a claimed throughput and power, and an assumed array size and precision. The checker converts the claim to energy per MAC (two operations per MAC) and compares it with what the converters and the shot-noise laser floor alone would need, using the model of OptAI 02. Defaults are the throughput and power of Lightmatter's Nature paper (65.5 TOPS, 78 + 1.6 W) and its 128 × 128 cores; the 8-bit precision and the converter figure of merit are illustrative assumptions, since the paper's 16-bit format is not an ENOB.

Try a claim of 100× better than a GPU: a GPU at about 0.7 pJ per MAC (OptAI 02, slide 11) becomes 7 fJ per MAC, and see what array width and precision it demands. Limits: converters and the photon floor only; the DAC is taken at half the ADC's figure of merit, laser efficiency 10% and 3 dB of loss.

The Interconnect Companies

These are interconnect, not compute: the devices and economics are those of Optical Technology for Data Centres. They belong here because several of them began in optical computing.

What Has Been Shown, and Where

• The strongest published compute result, Lightmatter's, works out at about 1.2 pJ per operation for the whole system (slide 02): real models at useful precision, but not yet below a modern GPU's board-level energy per operation. Its chief executive's “likely a decade” is consistent with that.
• Large-factor claims (30×, 50×, 100×) come from announcements without published methods; the interactive shows they require large arrays, low precision and very little energy outside the converters.

The Pull Towards Interconnect

• A visible pattern: of the companies here that began with optical compute, Lightmatter and Lightelligence now also sell interconnect, and the largest transaction in this list, Marvell's acquisition of Celestial AI, is for optical interconnect (slide 08). Luminous Computing, a photonics start-up that raised $105 million in 2022 to build what it called the world's most powerful AI supercomputer (press release), has made no public announcements since.
• Why, on the arguments of this series: interconnect moves bits digitally, with no analogue precision to defend, no weights to store, and a clear energy target set by copper and pluggables (OptDC 02). Optical compute must beat a digital MAC whose energy falls every process generation, while paying for conversions, static power and memory (OptAI 03).
• This is an observation, not a forecast. Compute companies continue to publish and ship, and the peer-reviewed results of 2025 are the most capable optical processors yet shown.

Questions to Ask

1. What workload, at what batch size, and is it a full model or a selected kernel?
2. What precision at the output, and what accuracy against the digital reference on the same model?
3. What does the power figure include: optical power, the photonic chip, converters, lasers, memory, host?
4. How many conversions per MAC: how large is the array, and how often are weights reloaded?
5. How much static power holds the weights and the calibration, and how well is the engine utilised?
6. Where do the weights live, and what does fetching them cost?
7. Measured or projected, and against which digital baseline, at which process node?
8. Is it peer-reviewed, or independently benchmarked?

Each question maps to a term in the models of OptAI 02 and OptAI 03, and to the interactive on slide 07.

What to Take Away

• Peer-reviewed optical AI systems now exist: Lightmatter's runs ResNet and BERT at near-electronic precision (65.5 trillion operations per second at 78 W plus 1.6 W optical); Lightelligence's solves Ising problems at up to 1 GHz; Microsoft's and Tsinghua's target inference, optimisation and vision.
• Most large-factor claims (Q.ANT's 30–50×, Lumai's 50×, Neurophos's 100×) are the companies' own, stated in announcements without published methods.
• Several optical-compute companies have moved towards interconnect, where the engineering case is clearer; Marvell's acquisition of Celestial AI is the largest deal listed.
• Read every claim with the questions of slide 11, and check it against the conversion floor (the interactive).
• The series in one line: optics computes linear algebra almost for free, and pays for it in conversions, precision, static power and memory; it wins where those are small against the work done.