On November 11, CollPlant Biotechnologies announced that its subsidiary LightSolver will give the first public demonstration of its Laser Processing Unit — the LPU, an all-optical computer that solves mathematical problems with interacting laser beams — at SC26 in Chicago this week. The demonstration runs at booth #5102 during the exhibition, November 16–19, with the conference itself running November 15–20. LightSolver says the LPU will solve systems of linear equations, differential equations, and complex optimization problems live, using only light.

That phrase — first public demonstration — is the whole story. Until now, every claim about the LPU has arrived through company announcements, a peer-reviewed paper built on an emulator, or paid customer programs whose results are not public. Nobody outside LightSolver has watched the machine run. A trade-show demonstration does not settle whether the machine is fast, accurate, or economical. But it is the first time the claims become checkable in public, on problems the audience can see being solved. That is a different kind of evidence from a press release, and a weaker kind than a benchmark. Hold both halves at once.

This is the fourth LightSolver datapoint this desk has covered in eighteen days — after the paid Defense Ministry framework (October 29), the ~$1.8 million Innovation Authority grant (November 5), and the OPHI ticker going live (November 13). A customer paying, a government panel funding, and a board rebranding are three different kinds of evidence; none proved the machine works. A public demonstration does not prove it either — but it is the first to put the machine itself in front of strangers.

What was actually announced

The facts come from the November 11 release, issued by CollPlant via GlobeNewswire. LightSolver will unveil the LPU’s capabilities in a first public demonstration at SC26, the International Conference for High Performance Computing, in Chicago, at booth #5102 during the exhibition, November 16–19. The demo will show the LPU solving “real mathematical and computational problems — including systems of linear equations, differential equations, and complex optimization problems — using only interacting light.”

The release positions the LPU as a machine designed to operate alongside conventional GPUs, “with the potential to significantly accelerate certain computationally intensive workloads while offering the potential for reductions in energy consumption” — a double “potential” that claims neither a measured speedup nor a measured energy saving on the physical hardware.

Dr. Ruti Ben Shlomi, LightSolver’s co-founder and CEO, called the demonstration “a major milestone for LightSolver as we work to move from a working prototype toward a commercial product.” The phrase “working prototype” is doing honest work: the machine exists and runs; a commercial product is still ahead, targeted for 2028, according to the company. The release’s own forward-looking-statements section flags “the risk that the LPU does not perform as expected in live demonstrations.” A live demo is a risk the company is choosing to take in public. A smooth demo on vendor-chosen problems is the expected outcome of a rehearsed demonstration, not evidence about performance on a customer’s problem.

One administrative note: CollPlant’s shares began trading as OPHI at the open on November 13, 2026, with the CUSIP unchanged. The name change to OptiPhi Computing Ltd. goes to a shareholder vote on Tuesday, November 24, and takes effect only if shareholders approve and the Israeli Registrar of Companies consents.

How the machine is supposed to work

The LPU is a closed-loop, all-optical system: a computational problem — a set of equations, an optimization — is mapped onto the physical dynamics of a network of interacting lasers, coupled so that light from each spills into its neighbours and the system settles into a state that encodes the answer. The iterative loop stays in the optical domain rather than shuttling intermediate results back into electronics at each step.

That last point is the architectural claim, and the one to keep in mind at the booth. Most photonic accelerators this desk has covered, including Q.ANT’s processors now being piloted on IONOS bare-metal servers, convert between digital and analog domains as they work. LightSolver’s argument is that keeping the iterative loop in light avoids that conversion cost. The counter-argument is that an analog system pays for that avoidance in precision: noise, calibration drift, and limited dynamic range all bound how accurate an answer can be. The only honest way to judge the trade is to see the machine solve problems whose answers are known and check how close it gets.

It is not a general-purpose computer, and not a neural network accelerator: its home territory is physics-based simulation, not LLM inference.

The Boeing presentation: the demo’s serious companion

Alongside the booth demo, LightSolver will present its work with Boeing on PDE-based corrosion simulation at the SC26 Exhibitor Forum. The presentation, according to the release, “will set out the engineering problem and present comparative benchmark results against existing solvers.” The work is part of a paid, multi-year development program with Boeing, announced in April 2026.

The booth demo shows the machine solving problems; the Boeing presentation is where numbers are supposed to appear. Corrosion simulation is a partial-differential-equation problem — the kind of workload the LPU is built for. A comparative benchmark against existing solvers, presented to HPC engineers, is a claim that can be argued with on the spot. What matters is what the presentation contains: which solver, on what problem size, at what precision, and with what error. The release promises “comparative benchmark results,” not a speedup figure, and none is given in advance. A named baseline and a measured result on physical hardware would be new evidence; projections from the emulator would be the same evidence the field already has, in a new room.

LightSolver’s September investor presentation reports projected corrosion-modeling acceleration of up to 100 times, based on emulation and proof-of-concept studies — a projection from the company’s own materials, not a measured result, and not a number from Boeing.

The Birds of a Feather session, and the DOE report behind it

LightSolver will also take part in a Birds of a Feather session on analog computing at SC26, informed by the US Department of Energy report “Analog Computing for Science: From Emerging Technologies to a Shared Research Agenda.” Dr. Talya Vaknin, LightSolver’s Head of Research, will present the company’s case for keeping most of the computation in the optical domain, alongside researchers from HLRS, Pacific Northwest National Laboratory, TU Dresden, and other institutions.

The DOE report comes out of a DOE Advanced Scientific Computing Research workshop held September 11–13, 2024, and was published March 20, 2026 (DOI 10.2172/2476965). It identifies six priority research directions, including mathematical foundations — a rigorous theory of analog computation, with noise-resilient algorithms and formal precision bounds — and analog architectures beyond narrow kernels: the questions a sceptic would ask about the LPU, namely how do you know the answer is right, and does it work on anything but a narrow kernel? A Birds of a Feather session is a discussion, not a peer review — but the company has chosen to argue its case in a room organized around exactly the hard questions.

The numbers that are not results

The HLRS figures are the numbers most likely to be repeated about LightSolver this week — and the numbers most often repeated wrongly.

In September 2026, LightSolver and the High-Performance Computing Center Stuttgart published joint research in the proceedings of the 23rd ACM International Conference on Computing Frontiers: “Accelerating Sparse Linear Solvers with an Optical Laser Processing Unit” (DOI 10.1145/3801488.3808041). Using an emulator of the LPU architecture, it benchmarked projected LPU performance against state-of-the-art iterative algorithms running on a GPU, and projected time-to-solution acceleration of approximately 40 times to more than 80,000 times, depending on the algorithm and problem evaluated.

The study was peer-reviewed, and HLRS is a serious partner — but the numbers are projections from an emulator, a conventional computer simulating how the laser system should behave. No laser produced the 80,000, and the authors note that a physical machine brings noise, calibration, and precision effects an emulator cannot reproduce. The range — 40 to more than 80,000 — says the answer depends enormously on which problem is handed to the machine. Treat the figures as what they are: a peer-reviewed projection that the architecture could be dramatically faster than a GPU on selected sparse linear systems, if the physical machine behaves like the emulator. The SC26 demo is the first public chance to see whether it does. What it can do is show the machine working at all, which the emulator cannot.

The 2028 commercial-availability target is likewise a roadmap statement, not a result.

The honest ledger

First, a demonstration is a vendor-chosen test: the problems, the problem sizes, the precision, and the success criteria are all chosen by the company showing the machine. A demo that runs smoothly proves the machine can solve the problems its makers chose, under conditions its makers control. That is a real achievement for a first public outing, but it is not a benchmark, and it should not be reported as one.

Second, no independent measurement exists yet. HLRS’s work used an emulator, and the Boeing and DDR&D programs’ results are not public — the latter, given the customer, may never be.

Third, the precision question is unanswered. An analog machine’s answers are approximate by nature: a linear solver 1 percent off is useless for some workloads and fine for others. Nothing in the announcement says what precision the demo will run at, or what error the answers will carry. Watch for whether the demo shows error bars or a comparison against a known answer.

Fourth, scale is unproven and the energy claim is still a “potential.” The roadmap talks about 100,000 variables in 2027 and a million in 2028; nothing in the announcement says how many variables the demo machine handles. The release says the LPU offers “the potential for reductions in energy consumption,” but no power figure for the demo system is given. Energy is the argument that would decide whether a datacenter buys this machine; it is also the argument with the least public evidence behind it.

What to watch at SC26

Four things will tell readers whether this week mattered. First, does the demo run, repeatedly, over four days? A machine still solving problems on Thursday, for different visitors, has passed the only test a trade show can administer. One that runs once, for a rehearsed audience, has not.

Second, what does the Boeing presentation actually contain? A named baseline solver, a problem size, a measured result on physical hardware, and an error figure would be the week’s most significant disclosure.

Third, does anyone outside the company get hands on the machine? LightSolver’s IIA grant program includes tools intended to let customers conduct independent R&D on the LPU — but a tool that exists is not a customer using it.

Fourth, what comes after the show? The next dated event is the November 24 shareholder vote. The next technical event is whatever LightSolver publishes about the demo — or does not. A company that publishes the demo’s problems, answers, and errors after the show is inviting scrutiny; one that moves on without doing so treated the demo as a marketing event. The distinction will be visible within days.

Until then, read this week narrowly and literally. A laser computer is stepping onto a show floor in Chicago, in front of the supercomputing community, for the first time. Everything else — speed, energy, precision, scale, and whether any of it matters to a customer — is still to be demonstrated. The demo is the beginning of the evidence, not the evidence itself.

Sources and further reading

Primary sources: CollPlant, “LightSolver to Unveil the First Public Demonstration of its All-Optical Computer at SC26,” GlobeNewswire, November 11, 2026 (booth #5102, exhibition November 16–19, demo problems, Boeing Exhibitor Forum presentation, Birds of a Feather session, Ben Shlomi quotation, forward-looking-statements warnings); CollPlant, “LightSolver Secures Paid Collaboration with the Israel Ministry of Defense,” GlobeNewswire, October 29, 2026 (DDR&D framework, initial 12-month order, no contract value disclosed; defense market estimate of up to $5 billion annually attributed to Hyperion Research 2025); CollPlant, “LightSolver Secures ~$1.8 Million Grant from Israel Innovation Authority,” GlobeNewswire, November 5, 2026; CollPlant, “CollPlant to Begin Trading on Nasdaq Under New Ticker Symbol ‘OPHI,’” GlobeNewswire, November 12, 2026 (ticker effective November 13; name change subject to November 24 shareholder vote and Registrar of Companies consent); CollPlant proxy statement, Form 6-K, November 3, 2026 (extraordinary general meeting November 24, 2026, 10:00 a.m. Israel time). Peer-reviewed research: LightSolver and HLRS, “Accelerating Sparse Linear Solvers with an Optical Laser Processing Unit,” proceedings of the 23rd ACM International Conference on Computing Frontiers, September 2026, DOI 10.1145/3801488.3808041 (emulator-based projections of approximately 40× to more than 80,000× vs. GPU baselines). Policy context: US Department of Energy, “Basic Research Needs for Analog Computing for Science,” workshop September 11–13, 2024, published March 20, 2026, DOI 10.2172/2476965. Company figures flagged as unverified — the 1,000 TOPS / ~100 W description, the projected up-to-100× Boeing acceleration, the 2027–2028 roadmap, and the ~$89 billion HPC market estimate (Hyperion Research 2024) — are from LightSolver’s September investor presentation, not from independent measurement. For the desk’s earlier LightSolver coverage, see the October 30 piece on the Defense Ministry agreement, the November 9 piece on the Innovation Authority grant, and the November 13 piece on the OptiPhi rebrand.