Photonic Computing Challenger Akhetonics: Can an All-Optical CPU Really Be 1000x Faster?

Photonic Computing Back in the Spotlight

A discussion about startup Akhetonics sparked attention on Hacker News, centered on a bold claim: that a fully photonic computing architecture could deliver processors up to a thousand times faster than conventional electronic CPUs. Claims like this tend to generate polarized reactions in tech circles — equal parts excitement about pushing physical boundaries and skepticism about real-world engineering feasibility.

The basic idea behind photonic computing is to use light (photons) rather than electrons to carry and process information. Photons travel through media at close to the speed of light and are not significantly constrained by resistive heat generation, theoretically enabling order-of-magnitude improvements in bandwidth, latency, and energy efficiency. This is why optical computing resurfaces every few years, each time in a different guise.

Why Light, Why Now

Modern electronic chips are approaching a dual ceiling — both physical and economic. Slowing transistor scaling, a persistent power wall, and the fact that data movement (rather than computation itself) has become the dominant energy cost have all prompted the industry to revisit non-von Neumann and non-electronic alternatives.

The appeal of photonic approaches centers on a few key points: extremely high signal transmission bandwidth, inherent parallelism (different wavelengths of light can simultaneously carry distinct data streams), and significantly lower latency and power consumption in interconnect scenarios. In fact, optical interconnects are already deployed at scale in data centers and some chip-to-chip communications, providing an industrial foundation for deeper "optical computing" exploration.

However, there is an enormous engineering gap between optical interconnects and truly all-optical logic operations. Akhetonics's "all-optical" approach means performing logic gates, storage, and control entirely within the optical domain — not merely using light to transmit data.

Optical interconnects in data centers are worth examining separately, as their operating principles and limitations are often misunderstood. The prevailing approach today places electro-optical converters at the edge of chips or modules, converting electrical signals into laser pulses, transmitting them over fiber, and then converting them back to electrical signals. This "electrical-optical-electrical" path offers clear power and bandwidth advantages over pure copper interconnects at distances over a hundred meters, and is widely used in 400G/800G data center switches and inter-node communications in GPU clusters. But fundamentally, this is still "transmit with light, compute with electricity" — the computation units themselves remain transistors. What companies like Akhetonics are attempting to cross is precisely this boundary: migrating the computation itself into the optical domain and eliminating the electrical stage entirely. Understanding this distinction is a prerequisite for evaluating whether claims about all-optical computing are credible.

How Should We Interpret "1000x Faster"?

When faced with headline claims like "1000x faster," careful interpretation is essential. Such comparisons are highly dependent on assumptions: does this refer to theoretical peak performance on a specific workload, or end-to-end real-world system performance? Is the metric clock frequency, throughput, or energy efficiency? The multiplier can vary dramatically depending on which dimension is being measured.

All-optical computing has long faced several stubborn challenges:

• Cascading optical logic gates: Electronic transistors can be chained together into complex circuits with ease, but optical signals degrade and distort across multiple stages of computation, making it difficult to maintain acceptable signal-to-noise ratios.
• Lack of efficient optical memory: Computation requires state retention, but "storing a beam of light" is far harder than storing electrical charge.
• Electro-optical conversion overhead: If the system still requires frequent conversion between optical and electrical domains, the supposed speed advantage is largely consumed by conversion latency and power draw.
• Manufacturing and integration maturity: Compared to the highly mature CMOS process, photonic integrated circuits still lag significantly in yield, cost, and design toolchain readiness.

A truly programmable, general-purpose CPU that operates entirely within the optical domain therefore remains an industry-recognized hard problem. Claims of order-of-magnitude performance leaps need to be backed by reproducible benchmarks and clearly defined use cases — not sweeping peak-performance analogies.

Optical matrix multiplication is currently the sub-field of photonic computing closest to practical deployment, and serves as a useful reference for understanding what all-optical computing can actually do. The core idea is to encode weight matrices as amplitude or phase parameters of optical modulators, with input vectors encoded as light intensity passing through a Mach-Zehnder Interferometer (MZI) network that performs the matrix multiplication — a process that completes at the speed of light, requires no clock cycles, and theoretically yields results in a single propagation pass. Companies such as Luminous Computing and Lightelligence have pursued this approach for AI inference chips. But these architectures are highly specialized, weight updates are difficult, and analog-to-digital conversion is still required to read out results. They demonstrate genuine advantages in "specific linear algebra operations" — not general-purpose programmable computation. This boundary is precisely the baseline that must be clarified when evaluating performance claims from any photonic processor.

Evaluating These Technical Claims Rationally

A reasonable expectation for photonic computing may not be "replacing the CPU," but rather serving as an accelerator in specific domains — such as matrix-heavy AI inference, signal processing, or enhanced optical interconnects. Several photonic AI chip companies in recent years (including startups working on optical matrix multiplication) are following exactly this more pragmatic path.

For an early-stage company like Akhetonics, the key things to watch for are: whether they have publicly disclosed technical details and benchmark data that peers can verify, whether the demonstration is a lab prototype or a deployable system, and whether the claimed performance improvements clearly specify the comparison baseline and workload type. Until that information is fully disclosed, "1000x faster" should be treated as an aspirational target, not a validated product specification.

Photonic computing represents a direction worth tracking over the long term — it carries genuine hope for breaking through the bottlenecks of electronics. But the history of technology repeatedly reminds us: the road from a compelling claim to a scalable, manufacturable product is long and brutally demanding. Staying open-minded while remaining skeptical is the healthiest posture when engaging with frontier topics like this.