Industries / Technology / Lasers and photonics
Photonic chips
Photonic chips put the parts that guide, switch, split and detect light, and sometimes the laser itself, onto one small chip made in batches the way computer chips are.
7
Milestones achieved, from 1969 to 2023.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1969
Laser circuits on a chip
Optical experiments of the day sat on tables measured in square feet, with separately mounted lenses and mirrors that drifted with temperature, sound and vibration. In the Bell System Technical Journal of September 1969 Stewart E. Miller of Bell Labs proposed "a miniature form of laser beam circuitry": light guided along narrow channels of slightly higher refractive index, patterned on a flat substrate by photolithography, the light-and-mask printing used for electronic chips. He called it integrated optics.
Credit
Stewart E. Miller, Bell Telephone Laboratories
1988
One chip sorts colors
In 1988 Meint Smit of Delft University of Technology proposed a flat component in which light passes through an array of waveguides of different lengths, so that each wavelength comes out at a different place: the arrayed waveguide grating (AWG). H. Takahashi and colleagues reported one in 1990, and C. Dragone a version with two star couplers in 1991. The AWG became one of the main parts for sending many colors of light down one fiber and separating them at the other end, and Smit received the 1997 IEEE/LEOS Engineering Technical Achievement Award for inventing it.
Credit
Meint Smit; H. Takahashi; C. Dragone
2006
Lasers bonded onto silicon
Silicon chips could guide light, and in February 2004 Ansheng Liu, Mario Paniccia and colleagues at Intel reported in Nature a silicon modulator, a switch that writes data onto light, faster than 1 GHz, where the best before was about 20 MHz. A light source made on silicon was still missing. In 2006 A. W. Fang, J. E. Bowers and colleagues, with Paniccia, bonded layers of a compound semiconductor (AlGaInAs) onto silicon waveguides and made an electrically driven laser that ran continuously, with up to 1.8 mW of output. The silicon waveguide alone set the laser cavity, so no fine alignment was needed and hundreds of lasers could be made in one bonding step.
Credit
Ansheng Liu and Mario Paniccia, Intel; A. W. Fang and J. E. Bowers
2006
Shared wafers open the field
On a multi-project wafer (MPW) run, many designers' chips share one wafer, so each pays a fraction of a full run. ePIXfab, set up by the research centers imec in Belgium and CEA-Leti in France, began in 2006 what it calls the world's first open-access MPW service for silicon photonics, completed 50 runs, and handed the service to Europractice by 2015. In the US the AIM Photonics institute began with a cooperative agreement of 9 July 2015 between the Research Foundation for the State University of New York and the US Air Force.
LessonAt Eindhoven, Meint Smit's group found that building every new function with its own process would take years and cost tens of millions of euros each, so it built a fixed library of proven building blocks that one process could combine into many chips, and showed in 2015 that the approach worked.
Credit
imec; CEA-Leti; AIM Photonics
2017
Neural network math in light
Yichen Shen, Nicholas Harris, Dirk Englund, Marin Soljačić and colleagues ran parts of a neural network on a programmable nanophotonic processor, a chip that does the network's multiplications by steering and mixing light. It was published in Nature Photonics in June 2017.
LessonThe paper's gains of at least 100 times in speed and 1,000 times in power efficiency were what the full design "promises"; the chip showed "essential parts" of it, and later claims in the field need the same reading.
Credit
Yichen Shen, Nicholas C. Harris, Dirk Englund, Marin Soljačić and colleagues
2018
Lithium niobate goes thin
Lithium niobate crystals had switched light in telecom networks for decades, but the modulators were bulky, costly and needed high voltages. C. Wang, M. Zhang, M. Lončar and colleagues at Harvard University, with X. Chen, S. Chandrasekhar and P. Winzer of Nokia Bell Labs, made lithium niobate modulators on a chip that run on the low voltages of ordinary computer chips, carry data at up to 210 Gbit/s and lose less than 0.5 dB of light on the chip.
Credit
C. Wang and M. Lončar, Harvard University; X. Chen, S. Chandrasekhar and P. Winzer, Nokia Bell Labs
2023
Intel hands off modules
A pluggable transceiver is the module that plugs into a network switch and turns electrical signals into light. On 30 October 2023 Jabil, a contract manufacturer, said it would take over making and selling Intel's silicon photonics pluggable transceivers and developing future ones. Intel's chief executive, Pat Gelsinger, said the company would focus on "the higher-value component business and optical I/O solutions", and called it the 10th business Intel had left in two and a half years.
LessonMaking photonic chips and assembling them into modules can be separate businesses; Intel kept the first and passed the second to a contract manufacturer.
Credit
Intel; Jabil
Watch for
Photonic processors at a supercomputer
Q.ANT says its second-generation photonic processors, built on thin-film lithium niobate chips and plugged into ordinary servers, now run at the Leibniz Supercomputing Centre (LRZ) in Germany. In benchmarks there it reports more than 50 times the matrix-multiplication throughput, 25 times faster ResNet-18 image inference and 6 times lower energy than its own first generation, and says photonic designs can use up to 90 times less power per workload.
Achieved whenLRZ or another independent party publishes measured energy per task against a current GPU on the same workload.
Claimed by
Q.ANT, March 2026
Optical servers run Llama
Lumai, a 2021 spin-out of the University of Oxford, says its Iris Nova server, which pairs digital processing with an optical tensor engine that computes with light in three dimensions, is the first optical computing system to run billion-parameter language models (Llama 8B and 70B) in real time, with up to 90% lower energy than GPU systems.
Achieved whenan outside party measures tokens per second and energy per token on a public benchmark such as MLPerf Inference.
Claimed by
Lumai, April 2026
Open
01
AI-designed chips, made and measured
A public test in which an AI agent designs 20 silicon photonic parts from written specifications, such as bends, splitters, ring filters, grating couplers and a small wavelength splitter; the layouts are made on a shared-wafer run; and each part's measured performance is published beside its target and beside a version drawn by a human designer.
The briefCredit · unclaimed02
Photonic computing claims, scored
A public scorecard of photonic AI computing claims from 2017 to now, by company and laboratory, recording for each figure whether it was measured or projected, what it was compared against, and whether the lasers, the converters between electrical and optical signals, and memory access were counted; with energy per operation recomputed on one basis where the data allow.
The briefCredit · unclaimed
Open 01 · the brief
AI-designed chips, made and measured
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test in which an AI agent designs 20 silicon photonic parts from written specifications, such as bends, splitters, ring filters, grating couplers and a small wavelength splitter; the layouts are made on a shared-wafer run; and each part's measured performance is published beside its target and beside a version drawn by a human designer.
Why it is still open
Published design agents stop at simulation. PhIDO, by Ankita Sharma, Joyce Poon, Dirk Englund and colleagues (August 2025), turned 102 written requests into mask files and reached about 57% end-to-end success on circuits of up to 15 components, and lists "extended verification" among its next steps; PICasso (2026) reports only simulated loss. The one agent-designed device found that was built and measured, from VortexChat (2026), is a terahertz beam multiplexer, not a photonic chip. An arXiv search on 20 November 2026 for language-model photonic design found no fabricated chip results.
Built with what exists
gdsfactory, an open-source Python library for chip layouts; KLayout for design-rule checks; an open simulator such as Meep; a current AI model such as Claude driving the design loop (see the Note "Three working devices", in which Anthropic says Claude Fable 5.2 designed three prototype devices that worked, with no outside test reported); Applied Nanotools' NanoSOI service, whose multi-project wafer runs come every six to eight weeks with turnaround as short as three weeks, and which also lists measurement among its services, so the chips can be tested by the foundry or a university photonics lab rather than by the project itself.
First step
Write five single-part specifications in the style of the PhIDO testbench, have an agent draw them in gdsfactory, and run the layouts through KLayout design-rule checks for the NanoSOI process.
Done looks like
20 parts fabricated and measured; insertion loss, center wavelength and bandwidth for each reported against its specification; a published tally of specifications met by the agent's designs and by the human designs; all prompts, layouts and raw spectra released.
Open 02 · the brief
Photonic computing claims, scored
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public scorecard of photonic AI computing claims from 2017 to now, by company and laboratory, recording for each figure whether it was measured or projected, what it was compared against, and whether the lasers, the converters between electrical and optical signals, and memory access were counted; with energy per operation recomputed on one basis where the data allow.
Why it is still open
The claims use different baselines. Q.ANT's March 2026 gains compare its second generation with its first; Lumai's April 2026 release gives "up to 90% lower energy" without test conditions; Shen and colleagues' 2017 paper called its gains promises. Tanner Andrulis, Joel Emer, Vivienne Sze and colleagues released an open model in 2024 showing that converting data between domains and fetching it from memory "can consume a significant portion of photonic system energy". An arXiv search on 20 November 2026 found modeling studies of research designs but no scorecard of published claims.
Built with what exists
The open-source model released by Andrulis and colleagues; company releases and papers; Python; a Claude-based document reviewer to pull each figure and its citation for a person to check (see the Note "Seven more flagged").
First step
Score the Q.ANT, Lumai and Shen 2017 claims on a one-page rubric, then test the rubric on two more systems.
Done looks like
At least 15 systems scored; every figure cited and labeled measured, projected or unstated; two people score each system separately and agree on at least 90% of labels; energy per operation recomputed for at least five systems and sent to each claimant for comment before release.