Industries / Technology / Telecommunications / Fiber optics
Transmitters and receivers
This work makes the lasers, light detectors, plug-in modules and signal-processing chips that turn electrical data into light at one end of a fiber and back into electrical data at the other.
7
Milestones achieved, from 1970 to 2026.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1970
Lasers at room temperature
Early semiconductor lasers had to be cooled and could fire only in short pulses. In 1970 two groups made them shine continuously at room temperature, using a double heterostructure, a sandwich of different semiconductor layers that holds charge and light in a thin middle layer and cuts the current a laser needs before it starts. The Kyoto Prize citation says Zhores Alferov in the Soviet Union, and Izuo Hayashi and Morton Panish in the United States, succeeded "almost simultaneously."
Credit
Zhores Alferov; Izuo Hayashi and Morton Panish
1988
Light from the chip's face
The vertical-cavity surface-emitting laser (VCSEL) sends light straight up out of the top of the chip instead of out of its cut edge. Kenichi Iga at Tokyo Institute of Technology conceived it in 1977, his team drove the first one with electric current in 1979, and Fumio Koyama of the team ran one continuously at room temperature in 1988, by the IEEE Milestone's account. VCSELs went into wide use in data links, and the citation notes that many can be made at once on one wafer.
Credit
Kenichi Iga and Fumio Koyama, Tokyo Institute of Technology
2001
The optical bubble bursts
In July 2001 JDS Uniphase, then a leading maker of optical components, reported a $50.6 billion loss for its fiscal year and planned 7,000 more layoffs, which would take its cuts since 1 January 2001 past 15,000.
Lessonmost of the loss was the shrinking book value of companies it had bought, not spending on factories.
Credit
JDS Uniphase (reported by Light Reading)
2008
Coherent detection with digital processing
In March 2008 Nortel announced a 40G/100G system using coherent detection, in which the receiver mixes the incoming light with its own laser so it can read the light's phase as well as its brightness. A digital signal processor (DSP), a chip that undoes the fiber's distortions with arithmetic, removed the need for compensating equipment along the route, so 40 Gbit/s could run over a network designed for 10 Gbit/s. In September 2008 Bell Canada said it was using Nortel's system on its Montreal–New York, Toronto–Chicago and Toronto–Montreal routes.
Lessonmoving the hard part into a chip let carriers raise speeds without rebuilding the fiber.
Credit
Nortel
2016
Silicon photonics in volume
Silicon photonics builds optical parts on silicon wafers with the tools of the computer-chip industry. In August 2016 Intel began volume shipments of 100 Gbit/s PSM4 transceivers made this way to large cloud providers and equipment makers. An earlier product was Luxtera's 40G "Blazar" cable, announced in August 2007, which Luxtera called the world's first CMOS photonics product.
LessonIntel still had to add compound semiconductor material to the silicon to make and detect the light.
Credit
Intel; earlier, Luxtera
2026
Lasers run short
Demand for AI data center optics ran ahead of laser supply. In February 2026 Lumentum said demand for its EMLs, indium phosphide lasers with a built-in light switch, outpaced supply by about 30% even after it raised capacity 20% in one quarter, and that all its added EML capacity was committed under long-term agreements through 2027. On 2 March 2026 NVIDIA agreed to invest $2 billion each in Lumentum and Coherent, with purchase commitments and rights to future capacity. TrendForce expected suppliers to double their combined monthly output of EMLs and continuous-wave lasers to about 50.7 million units in 2026.
Lessonbuyers answered with multi-year capacity deals and equity stakes, not spot orders.
semiconductor-today.comlightreading.comconvergedigest.comsemiconductor-today.com
Credit
Lumentum, NVIDIA and Coherent (their own figures); TrendForce
2026
Optics move onto the switch
In co-packaged optics (CPO), the optical engines sit in the same package as the switch chip instead of in plug-in modules on the front panel. On 16 November 2026 NVIDIA and Lambda, a GPU cloud company, said Lambda runs NVIDIA's Quantum-X CPO switches, each with 144 ports at 800 Gbit/s, in production in a cluster of more than 10,000 GPUs. In October 2025 Meta reported more than one million 400G port-hours without a link dropping in high-temperature lab tests of Broadcom's CPO.
LessonNVIDIA kept the lasers in 18 swappable light sources on the front panel, because light sources tend to be the parts that fail.
Credit
NVIDIA and Lambda; Broadcom and Meta
Watch for
Co-packaged optics for Ethernet
Broadcom says its Tomahawk 6–Davisson switch reaches 102.4 Tbit/s with 16 optical engines of 6.4 Tbit/s each, and uses about 70% less power for optics than plug-in modules. The same report says the switch is both "now shipping" and "sampling to early access customers."
Achieved whena customer or independent lab reports it running in production, or publishes measured power or reliability figures.
Claimed by
Broadcom, October 2025
1.6T coherent pluggables
Marvell calls its COLORZ 1600 "the industry's first 1.6T ZR/ZR+ pluggable module," rated for links of 20 km, 120 km and 1,000 km, and said sampling would begin in the second half of 2026. OIF, the industry body that writes interface agreements, published its 1600ZR agreement on 10 September 2026 for amplified single-span links up to 120 km.
Achieved whenmodules from two or more makers work together to the 1600ZR agreement in an OIF or independent test, and an operator reports carrying traffic on them.
Claimed by
Marvell, March 2026
400 gigabits per lane
At OFC 2026, Tower Semiconductor and Coherent showed a silicon modulator, made in a production-ready process, sending 420 Gbit/s on one lane with PAM4, which sends four brightness levels per symbol so each symbol carries two bits. They aim it at 3.2 Tbit/s transceivers.
Achieved whena customer or independent lab publishes bit-error results for a shipping 3.2T module built on 400G lanes.
Claimed by
Tower Semiconductor and Coherent, March 2026
Open
01
An open transceiver health dataset
A public dataset of the readings that plug-in optical modules report about themselves, called digital optical monitoring (module temperature, supply voltage, laser bias current, transmit and receive power, alarm flags), with labeled events of known kinds (added attenuation, a dirty or unplugged connector, temperature steps within the rated range) and a baseline detector.
The briefCredit · unclaimed02
An open real-time coherent receiver
An open-source FPGA design that runs the whole coherent receiver chain in real time for dual-polarization QPSK on an AMD RFSoC board: undoing chromatic dispersion (pulse spreading), a 2×2 polarization equalizer, frequency-offset estimation and carrier phase recovery.
The briefCredit · unclaimed
Open 01 · the brief
An open transceiver health dataset
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public dataset of the readings that plug-in optical modules report about themselves, called digital optical monitoring (module temperature, supply voltage, laser bias current, transmit and receive power, alarm flags), with labeled events of known kinds (added attenuation, a dirty or unplugged connector, temperature steps within the rated range) and a baseline detector.
Why it is still open
Published failure-prediction results come from operators' private fleets. An August 2026 paper reports 97.7% F1, a combined score of precision and recall, "across millions of optical links in production" AI data centers over ten weeks, and releases no data. GitHub searches on 19 November 2026 for "optical module failure prediction," "DDM DOM dataset transceiver" and "transceiver telemetry dataset failure" returned no repositories.
Built with what exists
Used Class 1 (eye-safe) SFP+ and QSFP28 modules and patch cords on a bench, never a live carrier network; Linux hosts with matching network cards; the open-source vexxhost/transceiver_exporter, which reads module diagnostics through Linux ethtool into Prometheus; fixed or variable optical attenuators; Zenodo for a citable release.
First step
Loop 8 modules back through patch cords on one host, run transceiver_exporter, save a week of one-minute samples, and post the schema for comment.
Done looks like
At least 50 modules from at least 5 makers logged once a minute for at least 90 days, at least 200 timestamped and labeled events, a schema and data card, and a baseline detector whose precision and recall on held-out events anyone can reproduce from the published files.
Open 02 · the brief
An open real-time coherent receiver
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open-source FPGA design that runs the whole coherent receiver chain in real time for dual-polarization QPSK on an AMD RFSoC board: undoing chromatic dispersion (pulse spreading), a 2×2 polarization equalizer, frequency-offset estimation and carrier phase recovery.
Why it is still open
Commercial coherent DSPs are closed chips. A May 2026 study put only the polarization equalizer on a ZCU208 FPGA, and only in fixed-point simulation, processing the rest offline in MATLAB. Searches found offline MATLAB and Python chains and a single-polarization radio QPSK demo, but no real-time optical receiver.
Built with what exists
An AMD Zynq RFSoC board (ZCU208 or RFSoC4x2, both supported by the open strath-sdr/rfsoc_qpsk demo); Vitis HLS; OptiCommPy, an open Python library with dispersion compensation, equalizers and phase recovery, as the reference model; test signals generated in software and fed electrically, so no laser is needed. Coding agents can draft the HLS blocks, each checked against the OptiCommPy reference (see the Note "Code no person read").
First step
Generate a one-million-symbol DP-QPSK test set in OptiCommPy with dispersion, laser phase noise, frequency offset and polarization rotation, record the reference error curve, and port the equalizer stage to Vitis HLS with a C-simulation test against the Python output.
Done looks like
A public repository that processes a continuous 1 GBd DP-QPSK stream for one hour without dropping samples, with simulated dispersion of 1,600 ps/nm (about 100 km of standard fiber) added, and needs no more than 0.5 dB more signal-to-noise ratio than OptiCommPy's floating-point chain to reach a bit error rate of 1e-3 on the same samples.