Akhetonics says its all-optical CPU reaches a customer in 2026

Share
Akhetonics says its all-optical CPU reaches a customer in 2026

Light-based computing keeps promising more than it delivers — but one Munich startup has just put a date on its bet, and the interview laying it out is doing the rounds on Hacker News this week.

Akhetonics says it will deploy its first commercial machine with a major customer by the end of 2026, with several more planned for 2027. The company, founded by Michael Kissner and Leonardo Del Bino, is building a computer where data enters as light, is switched as light, and circulates through memory as light — with no repeated optical-to-electronic conversions on the computation path. The pitch targets AI's real bottleneck: not arithmetic, which GPUs handle in bulk, but moving weights, activations, and context between chips and memory. So the first product is deliberately not a photonic GPU — it's a "Reasoning Processing Unit" aimed at the serial, branch-heavy workloads where adding cores stops helping: automated reasoning, formal verification, combinatorial optimization, the coordination overhead an agent burns between model calls.

The technical foundation is real enough to take seriously: a peer-reviewed 2024 paper in the Journal of Lightwave Technology describing a general-purpose optical CPU architecture built on a minimal one-instruction design, plus roughly 30 fabricated chip generations since. What's missing is the evidence that would settle the argument. No published benchmark names the workload, the baseline processor, the wall-clock time, or the power draw. Kissner's headline line — "we're just 1000x-ing the frequency," from gigahertz to terahertz — is an aspiration about clock rate, not application speed; his more measured claim of 10 to 100 times faster than electronic processors on reasoning tasks remains unverified. The company has raised 6 million euros to get there. Photonic computing has a long history of impressive laboratories and unsolved memory, packaging, and economics problems — this time the deadline, not the physics, is the story.


A YouTuber built a camera that only tracks police cars — and police came to his door. Anthony Sistilli, a Canadian software engineer, built an automated license-plate reader that exclusively logs police vehicles — "a personal Flock surveillance camera to Flock the Flockers," as he called it — to demonstrate how ordinary the technology has become. Two Peel Regional Police officers visited his Brampton home on September 20, he says in a video posted October 7, and their worry was that citizens could learn where officers live and when they work — the same privacy concern residents raise about Flock Safety's network, which the company says spans more than 120,000 cameras across 49 states and billions of vehicle scans a month. The asymmetry cut the way Sistilli intended: if plate data about officers is dangerous in private hands, the argument applies in both directions — as we noted when A cop searched 19,000 Flock cameras to type "LMAO".

What to watch: whether Akhetonics publishes a real benchmark — named workload, named baseline, power at the wall — before its year-end delivery claim comes due.

If a photonic CPU genuinely delivers 100x on reasoning workloads, which part of the AI stack should worry most — GPU makers, or cloud margins? Write to us via our contact page.

Get SiliconNoon in your inbox

One email a day with the AI stories that matter — what happened, the numbers, and why it matters. No spam, unsubscribe in one click.

Read more

NVIDIA Kumo Relational: 4 AUROC points over XGBoost, no training

NVIDIA Kumo Relational: 4 AUROC points over XGBoost, no training

Most enterprise prediction still runs on gradient-boosted trees fed by weeks of hand-built features. Nvidia's pitch is a foundation model that reads the database itself and predicts in a single pass — and it says that model now beats the tuned pipelines it replaces. Nvidia published Kumo Relational on Wednesday: a pretrained foundation model that makes predictions directly against a relational database, with no task-specific training, no feature engineering, and no flattening of tables into…

GigaChat 3.5 Reasoning lands under MIT — and beats DeepSeek on math

GigaChat 3.5 Reasoning lands under MIT — and beats DeepSeek on math

Sber's GigaChat team has open-weighted a 432-billion-parameter reasoning model under an MIT licence, and the numbers put it within touching distance of DeepSeek V4 Flash. GigaChat 3.5 Reasoning runs 28 billion active parameters, trains natively in FP8, and mixes multi-head latent attention with GatedDeltaNet linear-attention layers — a hybrid aimed squarely at long-context cost — with a 262K-token window and three speculative-decoding heads bolted on.

Nvidia ships Nemotron 3.5 Lightning and an open model router

Nvidia ships Nemotron 3.5 Lightning and an open model router

The enterprise AI story today is about choice: Nvidia shipped a fast, easily customized open model plus a router that picks between models, IBM bet $240 million on open-model inference, and Spotify decided AI-generated artist identities shouldn't get recommendations. Nvidia released Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model aimed at the high-volume, always-on workloads that power AI agents, alongside NeMo Switchyard, an open-source library that routes each step of…

Odyssey-3 opens a free public preview of its world model

Odyssey-3 opens a free public preview of its world model

The world-model race added a third serious entrant you can actually touch this weekend — Odyssey opened its Odyssey-3 to the public, and the benchmark claims came with strings attached. Odyssey-3 is now a public research preview, and the free demo runs today. California-based Odyssey — the 2023 lab founded by Oliver Cameron and Jeff Hawke — first showed off Odyssey-3 in mid-September; what's new is public access, a published technical write-up, and numbers.

Microsoft builds its decision model on Qwen, not OpenAI

Microsoft builds its decision model on Qwen, not OpenAI

Sunday's haul mixed one real model launch with a milestone for China's driverless delivery fleets — and a robotics round that's really about drug-making capacity. Microsoft launched Microsoft-Decision-1, a fast decision-scoring model built by post-training Alibaba's Qwen3.5-9B. Unlike an LLM that generates text, a decision model returns a calibrated probability — the kind of call an agent makes when it routes a ticket, flags an incident, or approves a refund.

Byteification turns Qwen3 and Llama 3 into byte-level models

Byteification turns Qwen3 and Llama 3 into byte-level models

The tokenizer is the last piece of the language-model pipeline that was never learned end to end — and a Nature paper out this week shows it can be swapped out for less than one percent of a pretraining budget. Researchers have retrofitted four open subword language models into true byte-level models using less than 1% of a typical pretraining budget, and the result — published 7 October in Nature — closes the performance gap that has kept byte-level models a niche curiosity for nearly a…