Akhetonics Photonic Computing: 1,000x Faster CPUs!

Learn about Akhetonics’ prototype that turns photons into a programmable, terahertz-class processor — and the memory, packaging, and power challenges they still need to overcome.

Akhetonics says their photonic CPU is the beginning of a revolutionary new type of all-optical computer: data would enter as light, be switched as light, circulate through memory as light, and leave as light. Leonardo Del Bino, the company’s co-founder and chief technology officer, reduces its hardest engineering problem to five words: “Light doesn’t like to stop.” Michael Kissner, Akhetonics’ co-founder and chief executive, says the company plans to deploy its first commercial machine with a major customer by the end of 2026. Between those statements lies the entire wager — a published two-bit optical CPU architecture, nearly 30 generations of photonic chips, a compiler intended to accept familiar code, ambitions for terahertz clocking, and unanswered questions about power, packaging, memory, manufacturing, and whether any of it can outperform electronics outside the lab.

A Photonic Computer Made Of Fibers And Pods

The first thing to understand about Akhetonics’ optical computer is that it’s not one thing yet. It’s a collection of devices, test structures, packaged chips, software, architectural ideas, and increasingly integrated prototypes. In footage from the company’s Munich operation, Del Bino apologizes for the disorder as he walks through a room where one of the early systems is being assembled. Inside a conventional rack enclosure sits a computing board intended to hold four photonic chips in pods, together with an orchestrator responsible for routing information. Fiber arrays bring light into the chips and carry the output away. Electrical connections power selected devices. Instruments measure whether the components behave as designed.

The scene can look familiar until the purpose of all those fibers becomes clear. Photonics is already embedded in the modern digital world: fiber-optic networks carry internet traffic across continents and oceans, optical transceivers connect data-center equipment, and silicon photonics is moving steadily closer to processors. But most of those systems convert light into electrical signals when computation begins. Akhetonics wants to eliminate that boundary. “We don’t have any electronic conversion,” Kissner says when describing the intended data path. He means that the information being processed should remain optical — not that the machine contains no electronics at all.

Lasers still need electricity. So do amplifiers, tuners, monitoring circuits, control systems, host machines, and other supporting equipment. The company’s more precise claim is that computational data doesn’t have to cross repeatedly between the optical and electronic domains. In the finished architecture Akhetonics describes, information would arrive through an optical interface, move through optical logic, circulate through optical memory, and pass among specialized optical units. Conventional power would keep the machine alive, but photons would carry the bits that matter.

Several stages of development can become blurred when the company discusses “the prototype.” There’s the published two-bit processor architecture designed to demonstrate general-purpose optical computation. There are individual optical logic gates, decoders, memory structures, routing devices, and multi-chip laboratory assemblies. Then there’s the much larger commercial system that Kissner says will consist of tens or hundreds of photonic chips. The essential question isn’t whether Akhetonics has made light perform computation — it has documented a credible architecture and component work. The question is whether all those pieces can become a reliable, programmable, economically useful computer.

The Security Question That Led To Light

Kissner didn’t arrive at photonic computing through the usual path of semiconductor research. He spent 12 years as an airborne infantry officer in the German military and says he worked around special operations, while maintaining a parallel life in computing, programming, cybersecurity, and white-hat hacking. He later completed doctoral work in computer science on a neural-symbolic framework for mental simulation, with research touching inverse graphics, intuitive physics, meta-learning, episodic memory, and systems that combine symbolic reasoning with neural networks. “I’ve always been a nerd,” he says, acknowledging that the résumé doesn’t resemble the standard preparation for running an optical-chip company.

His original concern wasn’t how to build the fastest possible processor. It was how to build one that could be trusted. Modern high-performance chips contain billions of transistors, intricate supply chains, layers of proprietary intellectual property, and design details that customers can’t realistically inspect gate by gate. For someone thinking about backdoors and secure computation, complexity itself becomes a risk. “How do you build a CPU that’s ultra-secure?” Kissner recalls asking. Any useful answer also had to preserve enough performance to run demanding cryptography rather than turning security into an unbearable slowdown.

Fully homomorphic encryption illustrates the problem. It can allow computation on encrypted data without exposing the underlying information to the machine performing the work, offering a conceptual path toward cloud services that never see their customers’ unencrypted data. But the computational cost can be enormous. The mathematics exists; the hardware frequently makes it impractical. Kissner’s search for architectures that could combine inspectability, security, and speed eventually led him to optical computing — and to a laser scientist who had already spent years trying to make light behave like logic.

Del Bino had worked on lasers for ultracold atoms and quantum simulation before turning toward all-optical logic gates and memory during research at the National Physical Laboratory in Britain and the Max Planck Institute for the Science of Light in Germany. Kissner contacted him on LinkedIn during the COVID era. Del Bino’s first reaction was suspicion: “Who’s this crazy guy?” When restrictions eased, they met in a Bavarian beer garden, made rough calculations, and concluded that no obvious law of physics killed the idea. More important, they recognized that Del Bino’s device-level approach and Kissner’s architecture-level approach would have to be developed together.

When Faster Computers Stopped Getting Faster

For decades, computer users experienced progress as a simple rhythm: new processors arrived, software became more demanding, and the next generation of hardware made everything feel fast again. Shrinking transistors improved density, performance, and economics so consistently that the industry could treat scaling almost as a law of nature. A program that strained one generation of hardware might run effortlessly a few years later. The transistor count rose, clock speed climbed, and programmers could expect tomorrow’s machine to rescue today’s inefficient code.

That pattern broke in the mid-2000s. As voltage and transistor dimensions stopped scaling together as neatly as before, pushing clock frequencies higher became thermally expensive. Processor makers turned increasingly to dual-core, quad-core, and many-core designs. Parallelism preserved overall progress, but it changed the bargain. A single sequence of instructions didn’t automatically become dramatically faster just because a new chip contained more cores. Software had to expose work that could be performed simultaneously, and some problems stubbornly remained serial.

Modern processors are incomparably more capable than their ancestors, yet mainstream CPU clocks are still measured in a few gigahertz. Gains now come through better microarchitecture, speculation, larger caches, vector instructions, additional cores, specialized units, advanced packaging, chiplets, and faster memory interfaces. Much of a processor die is devoted not to arithmetic but to retaining and moving data: caches, interconnects, controllers, and mechanisms intended to prevent fast execution units from waiting on slower memory.

This doesn’t mean Moore’s Law simply ended, or that electronic computing stopped improving. Nor should modern node names such as “3 nanometer” be interpreted as one literal feature on every transistor. Those labels now describe technology generations rather than a single physical dimension. The more consequential point is that leading-edge fabrication has become extraordinarily difficult and expensive, while shrinking devices no longer guarantees the effortless performance and efficiency gains of earlier eras. Akhetonics is betting that a different physical medium can open a different scaling path.

AI’s Hidden Tax Is Moving Data

Artificial intelligence makes the limits of the current architecture unusually visible. GPUs can perform vast numbers of mathematical operations in parallel, particularly the matrix calculations that dominate neural-network training and inference. But the arithmetic is only part of the job. Model weights, activations, intermediate results, prompts, context, and generated outputs must move constantly among accelerator cores, high-bandwidth memory, host processors, networking devices, and storage systems.

A GPU can be ready to calculate while waiting for the next data. A cluster of GPUs can be ready while network switches move information between racks. An AI agent can call a model, wait for an answer, decide which tool to invoke, retrieve more data, and send another job back to an accelerator. Every transition imposes latency. Every electrical trace and cable has physical limits. Every conversion, retimer, switch, and memory access adds energy to a process whose total cost can be dominated by movement rather than multiplication.

That’s why photonics has become central to the plans of companies that aren’t trying to build optical CPUs. Replacing longer electrical links with optical ones can increase bandwidth and reduce energy per transmitted bit. Co-packaged optics moves optical interfaces closer to switching silicon. Optical chiplets can reduce the electrical distance between a processor and the fiber carrying its data. None of those changes requires abandoning electronic CPUs or GPUs. They improve the roads between them.

The scale of recent industrial investment shows how seriously the established semiconductor world takes the issue. In March 2026, Nvidia announced separate $2 billion investments in Coherent and Lumentum to expand optical research, components, manufacturing capacity, and access for future AI systems. Its Spectrum-X photonics program has promoted co-packaged optical switches with 1.6 terabits per second per port and substantial vendor-claimed gains in efficiency and resiliency. In May, Nvidia and Corning announced plans to increase Corning’s U.S. optical-connectivity capacity tenfold and expand domestic fiber capacity by more than 50 percent. The mainstream answer to AI’s bandwidth problem is already becoming more photonic.

Akhetonics Refuses To Stop At Better Optical Wires

Most of the industry is taking a hybrid route for good reason. Optical links can relieve communication bottlenecks while leaving proven electronic compute and memory largely intact. A data center can adopt photonic networking without rewriting every application or replacing every processor. The approach reduces technical risk, preserves familiar manufacturing and software ecosystems, and lets companies introduce optics where the economic case is clearest.

Akhetonics thinks that solution stops too early. Kissner compares hybrid optical-electronic systems to hybrid cars: useful machines, but not the architecture one would design after committing fully to the new medium. As long as optical signals must be converted to electrons for logic and memory, the company argues, the system retains conversion latency, electrical data movement, and many of the bottlenecks photonics was supposed to avoid.

Its proposed computer is therefore much broader than an optical interconnect. An optical data interface would receive and transmit information. A digital optical processor would control program flow. Optical local and stack memory would hold working data. Read-only optical memory could contain code, while larger global memory could store models and datasets. Reconfigurable function units would perform specialized operations. Multiple wavelengths could carry independent streams of information through shared structures. Electricity would power the supporting apparatus, but the data itself would stay optical.

The company originally called the broad architecture an XPU, or Cross-Domain Processing Unit, because it was designed to combine digital, analog, and potentially quantum photonic functions. Its nearer-term commercial identity is now the RPU, or Reasoning Processing Unit. The change isn’t merely branding. It reflects a strategic decision to focus first on serial, branch-heavy workloads such as automated reasoning, verification, and optimization — problems Akhetonics believes are underserved by the parallel architectures that dominate AI.

Teaching One Beam Of Light To Switch Another

A modern electronic computer is built from transistors: tiny devices in which one electrical signal controls another. Reproduce that basic function in light, Kissner argues, and the foundations of a photonic digital computer begin to emerge. The required device is an all-optical switch — one beam of light determining whether another beam passes, changes, or is suppressed.

Kissner uses photochromic sunglasses as an intuitive analogy. Sunlight changes the lenses, causing them to darken and affect other light traveling through them. The real devices inside Akhetonics’ chips are far more complicated, but the concept is similar: light changes the optical behavior of a material, and that change controls a second optical signal. This is a nonlinear interaction, because the output isn’t merely a proportional copy of the input.

The difficulty is that light generally doesn’t want to interact with other light. That’s one reason fiber-optic communication works so well: multiple signals can share a medium without automatically colliding. Computing requires the opposite behavior at carefully chosen locations. The machine needs strong, controllable nonlinearities for logic gates, signal restoration, branching, memory, analog activation functions, and quantum feed-forward. Akhetonics says it uses materials from the III-V semiconductor family and two-dimensional materials to create those interactions in photonic integrated circuits.

Peter Caruana, an Akhetonics photonic design engineer with a background spanning physics, optics, computer science, machine learning, and graphics, calls low-power optical nonlinearity on an integrated chip a “holy grail” of photonics. Silicon offers mature manufacturing but can introduce unwanted nonlinear effects such as two-photon absorption. Silicon nitride is attractive for comparatively low-loss waveguides. Indium phosphide can provide active behavior, amplification, and useful nonlinear effects. The art lies not in choosing one magical substance, but in combining materials, devices, and foundry processes into a system that works repeatedly.

The Laboratory Where Theory Meets Fabricated Reality

Photonic design begins in software but eventually returns as matter. Engineers send layouts to a foundry, wait for fabrication, package the resulting dies, connect fibers, apply power, and measure what actually happens. The simulation can predict losses and switching behavior, but microscopic deviations in fabrication can change how light propagates through a waveguide or couples into another component.

“You can’t just trust theory,” Caruana says. The company needs its own measurements because optical devices that look correct in a model may behave differently once manufactured. A test station can inject light through a fiber array, vary wavelength and power, control electrical bias, and record the optical output. Automated routines allow the team to characterize many structures rather than tuning each device by hand.

In the earlier company video, Akhetonics team members show blue fiber lines feeding packaged chips and explain how the system collects output for analysis. These experiments aren’t merely pass-fail tests. They produce the data needed to refine models, define process tolerances, adjust future layouts, and understand how devices interact. A switch that works in isolation may fail when inserted into a longer path with loss, noise, timing variation, and multiple fan-outs.

This is one reason Akhetonics describes itself as a full-stack company. A better material can change the architecture. A longer optical path can change the clock. A memory device can change the compiler’s allocation strategy. Packaging loss can determine how many chips can be connected before signals require amplification. In a conventional computing ecosystem, different companies and decades of standards separate these layers. Akhetonics is trying to build them almost simultaneously.

The Two-Bit Machine Built To Prove A General-Purpose Idea

The company’s strongest public technical foundation is a paper titled “An All-Optical General-Purpose CPU and Optical Computer Architecture,” written by Kissner, Del Bino, Felix Päsler, Caruana, and George Ghalanos. Published in the Journal of Lightwave Technology in 2024, it asks what comes after optical interconnects. If photons eventually handle much of the communication among electronic chips, the authors argue, computation, memory, and optical-electrical conversion become the next major targets.

For its demonstration architecture, the team chose SUBLEQ, short for “subtract and branch if less than or equal to zero.” SUBLEQ is an example of a one-instruction computer. It’s deliberately primitive, but it can be Turing complete: given enough memory and time, it can express any classical computation. The point isn’t that future customers should replace modern instruction sets with a painfully inefficient one-command machine. It’s that general-purpose computation doesn’t inherently require billions of different switching elements.

The paper argues that a basic SUBLEQ processor can be implemented with fewer than 100 logic gates. Its demonstrator uses two-bit data paths, optical decoding, crossbar-style logic, delay-line registers, and optical memory concepts based on phase-change materials. A two-bit machine isn’t limited to numbers between zero and three any more than an old eight-bit processor was unable to manipulate larger values. Wider numbers can be assembled from multiple words, though the additional operations reduce speed.

That distinction — ability versus efficiency — is central to the paper. A minimal processor may be able to emulate a richer one, but slowly. Adding gates, native instructions, vector operations, larger word sizes, and specialized units improves performance. The published architecture therefore demonstrates a path toward general-purpose optical data processing; it doesn’t prove that Akhetonics already has a terahertz machine outperforming modern CPUs. It establishes a starting point from which the company says it has continued to scale.

From Oscilloscope Blobs To Thirty Chip Generations

The first logic gate appeared roughly six months into the company’s life, according to Del Bino. It wasn’t a sleek product demonstration. It was a device on an experimental table, surrounded by instruments, cables, and the debris of laboratory work. Kissner remembers the team watching “little blobs moving around on the oscilloscope” and realizing that something designed months earlier had finally returned from fabrication and switched correctly.

From there, Del Bino describes a progression from one optical gate to 10, then 100, 1,000, and 10,000 gates on a chip. In the earlier video, Kissner said the company had designed and fabricated 24 chips. In his June 2026 interview, he updated that history to nearly 30 chip generations or tapeouts. Those are company-reported milestones rather than independently audited production figures, but they indicate the amount of iteration occurring beneath the public architecture.

Iteration in hardware has a brutal clock. A design can be submitted with an expected turnaround of six months and return a year later. In some cases, Kissner says, the delay can stretch toward 18 months. If a component needed by the next chip doesn’t work, every dependent design may be postponed. He says foundry delays nearly killed the startup in its early years. Software can be patched the night a bug is found; a fabricated photonic circuit may need another expensive trip through design, fabrication, packaging, and measurement.

Akhetonics’ response was to make more chips rather than wait for one perfect design. Photonic tapeouts on mature processes could be affordable enough to support repeated experiments, and even a failed structure could reveal propagation loss, thermal sensitivity, fabrication variation, or unexpected nonlinear behavior. Thirty generations don’t prove commercial readiness. They do suggest a company trying to replace theoretical certainty with accumulated evidence.

What “1,000x Faster” Actually Means

Kissner describes Akhetonics’ essential ambition as taking a CPU from the gigahertz domain into the terahertz domain. “We’re just 1000x-ing the frequency,” he says. The phrase captures the company’s aspiration but requires careful arithmetic. One terahertz is 1,000 times one gigahertz. It’s about 333 times three gigahertz. Compared with a five-gigahertz processor, it’s 200 times the frequency.

Clock frequency also isn’t application performance. A processor’s useful speed depends on the amount of work completed per cycle, the number of cycles required per instruction, memory latency, path length, branching behavior, signal recovery, I/O, and the degree to which a program can use the architecture. A terahertz switch inside a device doesn’t automatically create a terahertz computer, and a terahertz computer wouldn’t make every application run hundreds of times faster.

Kissner offers more restrained performance ranges elsewhere in the interview, saying the company expects an optical processor to be 10 to 100 times faster than an electronic processor and claiming approximately 100-fold acceleration for certain automated-reasoning tasks. Those are major claims, but the public record doesn’t yet provide the complete benchmark information needed to evaluate them: workload, baseline processor, problem size, precision, compiler settings, physical hardware configuration, wall-clock time, and power.

Akhetonics also proposes another kind of scaling. Different wavelengths can travel through the same waveguide and carry independent information. Caruana illustrates the idea with red and blue light: if each color represents a separate computation, one physical path can do two jobs during the same interval. Kissner says the company has developed nonlinear devices that can switch several colors independently. Frequency, wavelength count, core count, and workload throughput are different metrics, however, and combining them into a single “1,000x” figure risks hiding more than it reveals.

Photonic Efficiency Still Has An Electricity Bill

Another easy misconception is that photons make computation almost free because light traveling through a good waveguide encounters no electrical resistance. Optical propagation can indeed be highly efficient, especially over distances where copper becomes difficult. But an optical computer requires light sources, amplifiers, switches, couplers, tuners, detectors for monitoring, control electronics, and packaging. Loss accumulates at bends, interfaces, splitters, crossings, and chip-to-chip connections.

Kissner is unusually direct about the device-level tradeoff. “Photonics is not really efficient,” he says, if one isolates the optical transistors themselves. He estimates that an optical transistor may require tens of milliwatts, a figure that would be enormous if multiplied by billions. Akhetonics’ argument is that it doesn’t need billions. A processor built from thousands or tens of thousands of comparatively large, fast optical devices could still deliver favorable system-level performance.

The arithmetic invites questions. At 10 milliwatts per continuously active device, 1,000 devices would require 10 watts and 10,000 would require 100 watts before accounting for all supporting hardware. If “tens of milliwatts” means 20 or 30, the total grows accordingly. Yet Kissner describes a system in the lower-watt range. That could reflect low activity factors, shared optical power, different definitions of device consumption, or the fact that not every element is powered simultaneously. Public materials don’t yet give enough detail to reconcile the figures.

The meaningful comparison will be made at the wall outlet. How much power does the complete system draw while solving a named problem? How much cooling does it require? How quickly does it finish? How much energy is consumed per verified result? Akhetonics may gain efficiency by using fewer devices, reducing conversion, moving data optically, and exploiting high speed and wavelength multiplexing. But the advantage has to be demonstrated at the system level rather than inferred from the word “light.”

The Problem Of Making Light Wait

Every computer eventually becomes a memory problem. A processor must retain instructions, addresses, variables, intermediate results, and the state of its program. The faster its logic becomes, the more damaging it is to wait for memory. A terahertz optical processor attached to conventional electronic RAM could spend much of its theoretical advantage stalled at a slower boundary.

“The main problem we have at the moment is memory,” Del Bino says in the company video. “You need to stop light. Light doesn’t like to stop.” Caruana points out that optical memory already exists in a basic form. A DVD uses light to read material states representing ones and zeros. But changing a phase-change material and reading it later is far too slow for the femtosecond- and picosecond-scale operation Akhetonics wants.

The company’s fast-memory answer is not to stop the light but to let it keep moving. Kissner describes signals circulating through delay lines or resonant paths until they’re needed, reviving an idea from early electronic and acoustic computing. A pulse carrying a value travels around a loop, is reshaped or amplified as necessary, and returns after a defined delay. The optical CPU paper uses related delay-line concepts for registers and proposes phase-change materials for information that changes less frequently.

This creates a hierarchy rather than one universal optical RAM. Circulating light may provide fast volatile state. Phase-change structures may serve read-mostly or write-once-read-many storage. Separate optical code memory could preserve instructions. Larger global storage might hold datasets or model information. The unsolved questions concern density, random access, loss, error correction, refresh, write endurance, capacity, and physical size. Akhetonics has a memory strategy; public data hasn’t yet shown a commercial optical memory system matching modern RAM in all the ways customers expect.

Compiling Familiar Code For An Unfamiliar Machine

A revolutionary processor won’t matter much if only its inventors can program it. Akhetonics is therefore trying to make the software interface feel ordinary even when the underlying physics isn’t. Max Aigner, a senior software engineer at the company, describes an earlier form factor as something resembling an oversized graphics card inserted into a conventional host computer. Applications would see an Akhetonics device without needing to understand its waveguides or nonlinear materials.

“It doesn’t actually know that this is photonic,” Aigner says of the host software. A driver and compiler would translate familiar source code into an intermediate representation such as SPIR-V, then into Akhetonics’ custom machine instructions. The company says its stack can support C++, LLVM-compatible languages, and software that compiles to SPIR-V. Kissner also discusses C, Python, Vulkan, OpenCL, OpenGL, and work toward a CUDA-like environment.

That’s compatibility at the source or framework level, not native compatibility with every existing binary. An executable compiled for x86 or Arm can’t simply be copied onto an unrelated optical instruction set and expected to run. If the source is available, it may be recompiled. A virtual machine could emulate x86 or Arm instructions, Kissner says, but emulation would add complexity and overhead that the initial target applications may not need.

Then there’s Doom, the traditional test of whether a strange piece of hardware can be treated like a computer. Akhetonics’ team has developed a C++ Doom project around its architecture and software stack. The company uses it to emphasize generality rather than gaming performance. Public descriptions remain ambiguous about whether the game has run end to end on fully physical all-optical hardware, on an emulator, on a software model, or through a partially integrated system. Doom is a compelling demonstration motif; it isn’t yet a substitute for a reproducible hardware benchmark.

Old European Fabs Could Find A New Purpose

Photonic circuits don’t need to shrink according to the same logic as electronic transistors. Akhetonics says its chips can use mature 90-, 130-, and 250-nanometer foundry processes. Kissner describes early devices fabricated on a 130-nanometer platform using silicon-nitride waveguides. The optical wavelengths involved range roughly from 600 to 1,550 nanometers, so the structures guiding them are naturally much larger than leading-edge electronic features.

That size sounds like a disadvantage in a semiconductor culture trained to equate smaller with better. Akhetonics argues that it’s also an opportunity. If performance comes from faster switching, low-latency optical movement, and multiple wavelengths rather than maximum transistor density, a foundry that can’t manufacture leading-edge CPUs may still be able to produce valuable photonic compute components. Shrinking below roughly 45 nanometers offers little benefit if the waveguide still has to accommodate much larger wavelengths.

Kissner is careful not to call the process automatically CMOS compatible. A conventional semiconductor fab can perform much of the work, but photonics introduces unfamiliar materials, waveguides, nonlinear layers, optical couplers, and process steps. “The foundries adapt to us,” he says. Existing metallization and portions of the manufacturing flow can be reused, while specialized stages are added or modified.

That creates a sovereignty argument as much as a manufacturing one. Europe contains many mature-node fabs that aren’t positioned to compete directly with the most advanced Taiwanese logic processes. Akhetonics believes photonics could give those facilities a new route into high-performance computing while reducing dependence on distant supply chains. The same logic might apply in other regions with established but non-leading-edge semiconductor industries. Whether the economics work will depend on yield, materials availability, packaging, volumes, and the cost of adapting each process.

Packaging May Be Harder Than Fabricating The Chips

If mature fabrication is an advantage, packaging may become the counterweight. A conventional optical transceiver might connect a chip to one or two fibers. Kissner says an Akhetonics device could require 64 or 128 fiber connections per chip, and a commercial processor could contain tens or hundreds of photonic chiplets. The result isn’t just a large chip. It’s an optical network assembled inside a computing system.

Every connection must align precisely enough to couple light without intolerable loss. Temperature changes can shift optical behavior. Fibers must be routed, secured, tested, and repaired. Chiplets must communicate at high speed without allowing signal degradation to overwhelm later stages. Amplification can restore power but also adds noise and complexity. Packaging yield may determine system cost even if the individual dies are inexpensive.

Akhetonics is building design tools to address the circuit side of this scaling problem. Its internal AtetDesigner system is described as an RTL-to-GDSII flow for optical digital logic: an engineer defines processor behavior in a hardware-description language such as Verilog, and the tool generates a photonic layout. The company has also released Connect-A-PIC, an open-source visual design and simulation environment intended to lower the barrier to creating photonic integrated circuits.

Automating layouts doesn’t automatically solve physical assembly. Electronic VLSI succeeded not only because engineers could draw more transistors, but because manufacturing, packaging, testing, timing models, libraries, and design rules matured together. Photonic computing will need its own equivalents. Akhetonics’ first large systems may resemble early computers built from many modules — not because the company wants a rack-sized CPU, but because integration hasn’t yet compressed the architecture into one package.

Why Akhetonics Is Building An RPU Instead Of A Photonic GPU

Many photonic-computing companies have focused on analog matrix multiplication, a natural fit for neural networks. Light waves can interfere and combine in ways that perform certain linear operations directly through physics. Akhetonics has experimented with analog and cross-domain ideas, but its immediate commercial focus is different. Kissner says the company doesn’t plan to compete first on the matrix-vector operations already served by GPUs, TPUs, and analog photonic accelerators.

The RPU is aimed at automated reasoning, formal verification, operations research, combinatorial optimization, code verification, and other workloads dominated by branching, serial dependencies, and low-latency memory interactions. Kissner argues that adding thousands of GPU cores doesn’t necessarily accelerate a task whose next step depends on the previous answer. Such workloads may benefit more from fast single-thread execution than from another expansion in parallel arithmetic.

This focus also fits the rise of agentic AI. A large model may run on GPUs, but an agent must coordinate tools, evaluate intermediate results, follow conditional plans, manage context, check constraints, and decide what happens next. Established chip companies are also emphasizing CPUs, networking, data processors, and memory systems around their accelerators because a GPU doesn’t operate alone. Akhetonics wants its optical processor to become the fast control and reasoning layer that keeps parallel units productive.

Cryptocurrency is less obviously suited to the initial product. Mining workloads are highly parallel, and major proof-of-work currencies are frequently dominated by specialized ASICs rather than general-purpose GPUs. Wavelength multiplexing could eventually let Akhetonics run multiple operations in parallel, but its first competitive claim isn’t that it can beat mature mining hardware. It’s that it can accelerate problems for which parallelism offers diminishing returns.

A Standalone Computer That Will Initially Need A Rack

Kissner says Akhetonics ultimately intends to make the whole computer optical, including processing and memory. That doesn’t mean the earliest commercial machine will operate in isolation like a familiar desktop CPU. It will still need lasers, power supplies, optical packaging, external networking, system management, and likely electronic control and host infrastructure. Earlier product descriptions presented the device as an OpenCL-style accelerator installed alongside ordinary x86 systems.

The distinction is similar to that between a processor architecture and the complete product containing it. An RPU might execute its target workload in the optical domain while relying on a conventional server to load jobs, manage storage, expose APIs, and handle tasks that don’t benefit from the new architecture. Over time, more support functions could move into photonics or specialized optical modules. In the beginning, practicality will matter more than architectural purity at the system boundary.

Size is one of the technology’s clearest limitations. Kissner says individual optical transistors or gates can be visible to the naked eye, unlike nanoscale electronic devices. The first commercial CPU may require tens to hundreds of chips. That rules out optical smartphones and makes laptops unlikely in the near future. The initial machines belong in server racks, research facilities, industrial sites, data centers, or businesses that can justify dedicated equipment.

Del Bino imagines the longer-term result as a return to local computing. Schools, offices, and companies could operate high-performance AI servers on-site rather than sending sensitive information to distant cloud providers. Local systems could improve latency, privacy, and resilience while reducing dependence on a handful of hyperscale facilities. It’s an appealing vision, although it depends on the optical machines becoming not only fast but affordable, reliable, serviceable, and compact enough for ordinary organizations.

A Digital Control Layer For Analog And Quantum Machines

The Reasoning Processing Unit is only part of Akhetonics’ original ambition. The broader Cross-Domain Processing Unit would combine digital optical logic with analog mathematical units and photonic quantum components. From a fabrication perspective, these domains can share waveguides, couplers, nonlinear materials, and similar layouts even though they manipulate information differently.

Kissner describes the digital processor as a kind of “janitor” for the other domains. An analog photonic accelerator may perform a matrix operation efficiently, but something must decide when to run it, route its inputs, interpret the result, and select the next operation. A photonic quantum circuit may require rapid feed-forward based on a measurement. If the control path converts repeatedly to electronics, it can impose latency and limit scalability. Akhetonics wants digital optical logic to provide that control while keeping the surrounding information optical.

The company is involved in several collaborative projects aligned with that idea. GATEPOST focuses on graphene and other two-dimensional materials integrated with silicon-nitride CMOS processes for ultrafast nonlinear optical interactions and secure connected systems. HAETAE is a 36-month European–South Korean project with an overall budget near €2.9 million, targeting multi-material photonic chiplets, transfer-learning accelerators, operation up to 50 Gbaud, and a stated goal of 15 times greater energy efficiency than its comparison point.

The SPOC project with Fraunhofer IZM explores packaging for room-temperature continuous-variable photonic quantum computing, including optical interconnects and integrated gates intended eventually to connect with Akhetonics’ XPU. The company has also publicized an all-optical analog-to-digital converter designed to translate optical analog information into optical digital outputs without an electronic conversion stage. These remain research programs and development targets, but they show that the cross-domain concept hasn’t disappeared behind the RPU branding.

Photonics Isn’t A Waiting Room For Quantum Computing

Optical computing is frequently confused with quantum computing because both can use lasers, interferometers, waveguides, and specialized optical measurements. Akhetonics’ digital processor is a classical computer. Its bits may be represented by light, but it isn’t relying on quantum superposition or entanglement to perform its basic logic. The programming and computational model are intended to remain recognizable to conventional developers.

Kissner sees quantum computing as a complementary technology rather than the inevitable replacement for all classical processors. Quantum machines may eventually provide dramatic gains for selected algorithms, while ordinary software, control flow, operating systems, databases, networking, and most business logic remain classical. A photonic digital processor could be valuable whether or not fault-tolerant quantum computing arrives soon.

It could also become part of a quantum system. Photonic quantum architectures often require fast classical decisions based on measurements. Optical digital feed-forward could reduce the need to convert the result into electronics, process it, and send a new optical command back into the circuit. In that sense, photonics isn’t simply a bridge between electronics and quantum computers. It could be a classical computing platform in its own right and a control layer for specialized quantum modules.

The same is true of analog optical computing. A system doesn’t have to choose one permanent paradigm. Digital logic can handle branching and precision, analog units can accelerate suitable mathematics, and quantum components can be invoked for narrow problems. Akhetonics’ long-term thesis is that a shared photonic platform might let these domains cooperate more directly than systems assembled around repeated electronic conversion.

The Company Has Money, A Market Thesis, And A Deadline

Akhetonics raised €6 million — reported at approximately $6.3 million at the time — to deliver first-generation hardware to pilot customers. The company has identified aviation, defense, and networking as potential early markets, arguing that deterministic execution, low latency, supply-chain control, and resistance to electromagnetic interference can matter alongside raw performance.

The RPU strategy adds automated reasoning and formal verification to that list. Formal tools are used to prove properties of hardware and software, check whether systems meet specifications, and explore enormous spaces of possible states. Operations research covers scheduling, routing, resource allocation, logistics, and optimization problems faced by airlines, manufacturers, fleets, and infrastructure operators. These tasks can be commercially valuable even if the first optical computer never runs a consumer operating system.

Kissner says one of the company’s first computers will be deployed with a major customer by the end of 2026, with several more planned for 2027. He hasn’t named that customer publicly in the interview. Akhetonics is also deciding whether systems should be sold outright, installed through colocation arrangements, or exposed remotely so outside users can submit workloads. Cloud-style access would let customers test the processor without installing an unusual optical rack in their own facilities.

The deadline transforms the story. Akhetonics no longer presents commercial hardware solely as a distant endpoint. It says a machine is approaching a real environment where a customer expects useful results. That raises the standard of evidence. A laboratory prototype can succeed by demonstrating a principle. A commercial system must operate repeatedly, accept jobs, survive environmental variation, deliver answers, and create enough value to justify its cost.

The Measurements That Will Decide Whether The Bet Worked

The most important next document from Akhetonics may not be another architecture paper. It may be a benchmark report. Which workload ran? What electronic system served as the baseline? What portion of the task executed optically? What was the measured clock or switching rate? How much memory was available, and at what latency? How much wall-plug power did the full system consume? How often did it fail?

Performance claims need enough context to be reproduced. A 100-fold speedup on formal verification could be transformative, but only if the baseline and workload are appropriate. A terahertz device could be scientifically significant without producing a terahertz application. A low-power chip could be overwhelmed by its lasers and packaging. Conversely, an ungainly rack could still be commercially valuable if it solves an optimization problem in minutes that previously took days.

Independent validation will matter especially because photonic computing has a long history of ambitious promises. Kissner acknowledges that skepticism is justified. Optical systems have repeatedly impressed in laboratories and then struggled with memory, control, manufacturing, integration, or economics. Some companies that began with optical-compute ambitions have concentrated instead on interconnects, where the road to revenue is shorter. Akhetonics has chosen the harder claim, and with it the harder burden of proof.

Yet that burden is also what makes the company compelling. It isn’t merely trying to make a better optical cable. It’s asking whether the medium that carries the world’s information can become the medium that reasons over it. The machine in Munich remains part artifact and part argument: chips, fibers, software, measurements, and a conviction that the next computer doesn’t have to be a smaller version of the last one. The future won’t be proven merely when light moves fast. It’ll be proven when the machine can make light wait, remember, and earn its keep.