Switch-side CPO lasers still come from two suppliers. The scale-up grid was set so standard lasers can meet it, and challengers now argue manufacturing.
Lumentum said on September 21, 2026, that it will demonstrate an eight-wavelength external laser module designed to meet the requirements of the Optical Compute Interconnect (OCI) multi-source agreement, or MSA, during the European Conference on Optical Communications (ECOC), with initial availability in the first half of 2027. The pluggable uses the company's ultra-high-power laser platform in an eight-channel configuration, in two four-channel bands near 1311 and 1331 nm, demonstrates up to 24 dBm (about 250 mW) per wavelength at the fiber, and draws roughly 12 W. The release names no customer.
The announcement is modest, but it marks a shift in the role of the data center laser. In a conventional pluggable transceiver, the indium phosphide (InP) laser and the high-speed modulator are on a single chip, an electro-absorption modulated laser (EML), and the engineering difficulty is making that chip switch fast. Co-packaged optics (CPO) moves the optical engine onto the package beside the switch or accelerator die and does the modulating in silicon. The laser is left with one task: supply steady, unmodulated light, usually from a replaceable module on the faceplate.
Co-packaged optics needs two kinds of that light, and they are on opposite trajectories. Switch-side CPO, shipping today, wants a handful of very high-power single-wavelength lasers per module, and the public list of merchant suppliers shipping that part has two names on it. The scale-up links that follow- the accelerator-to-accelerator connections inside an AI supercomputer- want many moderate-power wavelengths on one fiber, and they are the second of co-packaged optics' two front doors. That second market was specified to be open: the body writing its interface told SCN that any laser source is allowed and that the wavelength grid was set wide enough for standard lasers to meet it, which puts the incumbents inside scale-up from day one. Wavelength precision, the ground several challengers chose to argue on, is not, on the spec's evidence, what decides the first design-ins. What is left is manufacturing: how many wavelengths one part can carry, what it costs to build millions of them, and whether anyone can show the parts survive in the field. On that last question, none of these companies have published anything.
This map draws on filings, standards text, and company documents; an interview Supercomputing News (SCN) conducted with Lightmatter CEO Nick Harris during the AI Infra Summit; and written questions SCN sent to 16 companies, organizations, and researchers. Six answered: Lightmatter, Scintil Photonics, Xscape Photonics, OpenLight, the OCI MSA and the market research firm LightCounting. Tower Semiconductor declined to comment. Nine did not respond by publication.
In a silicon photonics engine, the modulation happens in a micro-ring or Mach-Zehnder structure on the silicon die, or in newer modulator materials such as electro-optic polymers added to it, and one continuous-wave (CW) laser can be split across several lanes. NVIDIA's technical description of its Quantum-X and Spectrum-X Photonics switches gives the ratio. Each external laser source holds eight lasers and feeds 32 of the Quantum-X switch's 576 transmit lanes, which works out to one laser per four 200G lanes, and NVIDIA says the design cuts the number of lasers in the data center by a factor of four against legacy designs. That trade isn't free, LightCounting CEO and chief analyst Vladimir Kozlov told SCN: four times fewer laser chips means chips four times the power, and supply for the higher-power chips is the more constrained of the two.
Each remaining laser works harder. The Optical Internetworking Forum's co-packaging framework document states the trade: keeping the laser outside the package improves reliability and thermals, but the added insertion loss must be made up for with optical power, whereas a laser integrated on the engine has less loss to overcome. A 2022 paper by Lumentum and Facebook engineers put the requirement for a 51T CPO switch at roughly 22 to 25 dBm per laser, or 158 to 316 mW. Their prototype met a preliminary power-conversion-efficiency target of 10%, which means about nine-tenths of the electrical input leaves as heat.
OIF's pluggable laser-module specification, ELSFP, defines power classes up to 26 dBm (400 mW) per wavelength, caps each fiber at 24.5 dBm for systems that rely on automatic power reduction instead of an interlock to stay within eye-safety limits, and warns that at these levels a contaminated connector burns beyond field repair. A naming trap for anyone comparing data sheets: Lumentum's "UHP" brand describes a 400 mW chip, while OIF's "UHP" class means 23 dBm in the fiber.
OIF's framework document also sets the lifetime bar. Its target is 50 FIT for each laser source feeding a 3.2 Tb/s engine, the whole source and not each laser inside it (one FIT is one failure per billion device-hours), and it says the roughly 300 FIT typical of pluggable modules would fall short if simply scaled to CPO.
Lumentum's ultra-high-power (UHP) laser does not come from its EML line. The company's 2026 reliability white paper says the part is built on its 600 mW Raman pump-laser platform, the lasers that drive amplifiers in long-haul and subsea networks. Initial supply comes from the San Jose fab, with UK shipments starting in late summer 2027, and the former Qorvo fab in Greensboro, North Carolina, now being retrofitted for 6-inch InP, is expected to ramp CW and UHP laser production in mid-2028, with NVIDIA named as a customer.
CEO Michael Hurlston made the scarcity argument on the company's February 3, 2026 earnings call: "the power level that needs to be delivered here, 400 milliwatts, is not something that many people can do." On the August 11 call, he said Lumentum is well behind on high-power laser shipments, and Wupen Yuen, president of global business units, added that the ramp is on track and the shortfall reflects demand. Lumentum did not respond to a request for comment by publication.
Until this week, the only support for that outside the vendors' own statements was the shape of the supplier list. Kozlov put the position more bluntly. Lumentum is qualified at NVIDIA now, and Coherent will be qualified next, he said, and both companies bring three decades of building very high-power lasers for erbium-doped fiber amplifiers, a technology that carries over to CPO. "We do not expect any other suppliers to get qualified in the next 2-3 years because reliability of these lasers is very hard to prove," Kozlov told SCN. That is an analyst's reading of a customer relationship rather than a statement from NVIDIA or from either supplier, and neither company has described its qualification status in those words.
The reliability case is a vendor's, and Lumentum's white paper is open about its basis. The field record it cites (nearly half a million lasers, more than 20 billion device-hours, under 1 FIT) belongs to the pump laser. For the UHP part, the evidence is more than 100 million accelerated device-hours with zero failures, which Lumentum computes as 9 FIT at 60% confidence and 23 FIT at 90%, and the company rates the laser at 20 FIT at 400 mW. That is a per-laser figure, which the paper calls a conservative calculation limited by accumulated test hours, and which Lumentum expects field data to beat. The UHP laser has no field history yet, and the acceleration factors are not disclosed. Lumentum's executives also concede that the laser earns less per port. Hurlston said in February that laser dollars in a CPO port are lower than a transceiver would bring, offset by high share, and that selling the whole laser module could bring two to two and a half times the revenue of bare chips.
Coherent is the second source. It sampled a CW laser delivering more than 400 mW at 55°C in September 2025. CEO Jim Anderson told investors on the company's August 12, 2026 call that its Texas facility has begun ramping UHP CW lasers for CPO, with first CPO revenue expected in the December 2026 quarter, and that its 6-inch InP lines in Texas and Sweden are yielding above its 3-inch lines. Converge Digest's account of Coherent's ECOC media and investor event dates volume shipments of the UHP CW laser to the fourth quarter of 2026 and puts up to eight of them in one replaceable external source. None of that material gives an output power for the shipping part, and Coherent has announced no multi-wavelength source for the scale-up grid. Coherent did not respond to a request for comment by publication.
On September 21, the same day Lumentum announced its OCI module, Coherent launched PhotonLink, a platform that bundles its InP lasers, VCSEL arrays, silicon photonics, isolators, specialty fibers and detectors into external laser sources and optical assemblies tailored to a customer's architecture. The release dates PhotonLink revenue to the fourth quarter of 2026, claims more than ten CPO customer engagements and more than ten in near-packaged optics with anchor customers and long-term agreements on both, and cites a shipped base of more than 300 million InP lasers, which it attributes to the company's established datacenter manufacturing base rather than to the CPO part.
NVIDIA put $4 billion into the two companies in March 2026, $2 billion each. One laser maker sits outside that pair because it is also the other CPO switch vendor. Broadcom's Tomahawk 6-Davisson switch takes field-replaceable ELSFP laser modules, and the company, which did not respond to a request for comment by publication, makes its own lasers and demonstrated a CW laser for silicon photonics in March 2024; it has not said publicly whose laser chips go into its CPO modules.
Behind these sit companies with parts or papers and no announced CPO production ramp. Furukawa Electric researchers have published efficiency results for high-power DFB lasers: 21% power conversion efficiency at 200 mW up to 75°C, holding 21% at 400 mW up to 50°C. In China, Shijia Photons told investors on August 4, 2026, that its commercial-temperature 400 mW CW DFB laser is in small-batch shipment while several customers evaluate samples. In SCN's reading, the wider supplier base one would expect for a simpler part has yet to arrive at 400 mW, where the public production list for merchant suppliers, those that sell to any buyer, is two companies with the same strategic investor.
Underneath every supplier on that list is InP capacity. Taiwanese trade press reported that a TSMC vice president, speaking at the SEMI Silicon Photonics Global Summit in Taipei on August 31, named lasers among the bottlenecks to scaling CPO; SCN's account also covers Coherent's Sherman, Texas expansion, which has a letter of intent for up to $50 million in proposed CHIPS funding, and China's roughly 70% share of refined indium. Anderson said in August that InP capacity remains Coherent's primary constraint, and on Lumentum's August call, Hurlston said the UHP demand surge sent the company to AXT for more substrates. Lightmatter CEO Nick Harris told SCN that both companies are fully booked and argued that fabs, more than substrates, are the first-order limit. In his written answers, he named laser fabrication and testing as two of the biggest constraints holding back the ramp.
Kozlov's forecast splits the problem the same way. Laser chip shortages should generally ease by the middle of 2027, he told SCN, unless demand keeps running ahead of LightCounting's forecast; his firm's published view earlier this year was less patient: in March it expected relief by mid-2026. CPO lasers are the exception he now carves out. Shortages and higher prices for them may last until 2030, he said, and prices today run about ten times those of the laser chips used in pluggable transceivers, possibly narrowing to five times by 2030 as volumes rise.
Scale-up links run on copper today and hold no laser content, which is why every supplier in this article is aiming at them. The governing specification calls for a different kind of source.
AMD, Broadcom, Meta, Microsoft, NVIDIA and OpenAI formed the OCI MSA in March 2026. Its line interface specification, version 1.0, dated March 11, describes a dense wavelength-division multiplexing (DWDM) grid read by cascaded micro-ring resonators. Each fiber is bidirectional and carries eight wavelengths: two groups of four, one group per direction, near 1311 and 1331 nm. Every wavelength runs at 53.125 GBd with simple NRZ signaling, and four of them make one 212.5 Gbps lane.
On the laser, the text reads as a requirement. "The OCI implementation shall use an ELS in compliance with the OIF ELSFP implementation," Section 2.5 says, with light delivered over polarization-maintaining fiber. The document limits linewidth to 1 MHz, sets noise and side-mode limits, and leaves optical power per wavelength and per fiber to the implementer. Each channel must sit within ±0.2 nm of its grid position. By SCN's arithmetic, the channels are 2.3 to 2.4 nm apart, or about 400 GHz at these wavelengths, and ±0.2 nm is about ±35 GHz.
Drew Alduino, the OCI MSA Chairperson, told SCN that the specification is intended to provide only an optical interface specification, and that any type of laser source, "external or integrated," is permitted. Version 1.0 is not uniform on the point: a footnote to the transmitter table defines a laser-input reflectance limit for implementations that have an ELS, wording that assumes some implementations will not. That March text is still the only version published, and SCN did not ask, nor did the MSA say, whether it will be revised to match the Alduino's reading. Both statements stand.
Alduino also answered the question behind the ±0.2 nm tolerance, but his answer focused on spacing rather than tolerance. He said the spacing was chosen for several reasons, and one main benefit is that standard lasers can meet it. For generation two, which targets 400 Gbps in each direction, the MSA is working toward publication with the goal of keeping the same wavelength spacing. He gave no date, no wavelength count, and no per-lane rate, and the same spacing at twice the bandwidth could be reached by faster lanes or by more of them.
In SCN's reading, that goes most of the way toward settling an argument the light-source vendors have been having with each other. The first-generation grid was set so that ordinary sorted lasers clear it, and Lumentum's September module is the demonstration: eight of the lasers it already builds, at the spec's own ±0.2 nm. If generation two holds the spacing, the grid does not get denser for at least two generations. The MSA did not say whether the ±0.2 nm tolerance itself will change, so wavelength precision is not settled, but on the spec's evidence it is not what decides the first scale-up design-ins. Tighter grids of 200 or 100 GHz appear in several vendors' roadmaps, including Scintil's, and no MSA document plans one.
Harris framed the opening by comparing lasers and fiber to high-bandwidth memory: the component that cannot be bought in volume ends up setting the architecture. By his arithmetic, a scale-up domain with eight optical engines per GPU needs about 100 times as many lasers as scale-out switching. The approaches below are ordered by how far they depart from the incumbent design, and the order implies nothing about merit.
The most conservative answer uses the same kind of InP DFB laser and more of them. Lumentum's OCI module is one version. Ayar Labs' SuperNova source is another: a 16-wavelength module built around DFB laser arrays from Sivers Semiconductors' InP platform. In this map, Ayar Labs is a laser buyer, exposed to the same InP capacity as everyone else. Electronics Weekly reported on September 16 that Sivers has a 100 mW O-band CW DFB chip and an eight-channel array, and is using its Glasgow facility for pilot manufacturing, with no customer named and no CPO program attached. Neither Ayar Labs nor Sivers responded to a request for comment by publication.
This row has more outside evidence behind it than the other challenger approaches, chiefly the long field history of discrete InP DFB lasers. The cost is assembly, since the InP parts have to be placed and aligned module by module.
Scintil Photonics, based in Grenoble, France, goes after that assembly cost. Its LEAF Light source is a single photonic integrated circuit carrying two sets of eight DFB lasers, configurable as 8 wavelengths at 200 GHz spacing or 16 at 100 GHz, with an ELSFP-format module as the target package. The company's process, which it calls SHIP, bonds unprocessed III-V dies to the back side of processed silicon photonics wafers and defines the laser gratings in silicon with deep-UV lithography. In SCN's reading, that ties channel spacing to the foundry's lithography and removes the step of sorting individual laser chips. The device concept appeared in an OFC 2023 conference paper, which reported about 1 mW of fiber-coupled power per wavelength on a four-wavelength prototype.
The foundry is Tower Semiconductor, on its PH18M platform. A joint release on February 17, 2026, said the process is validated on Tower's production lines and offers a path to millions of units a month. Scintil's evaluation kit, announced March 11, targets qualified customers for the second quarter of 2026. The same release targets a 50% power reduction against single-wavelength CPO, a system-level figure that depends on the customer's engine. NVIDIA is an investor; Scintil has named no customer.
Crowley answered SCN's questions, and his reading of Lumentum's announcement is close to the MSA's own. Generation one was written with incumbent laser manufacturers in mind, he said, and Scintil expected established suppliers to address it with the lasers they already make; a module from a supplier of Lumentum's standing validates the architecture and shortens the conversation about why anyone needs several wavelengths on one fiber. He expects to be one of several sources on generation one, and called a multi-sourced market easier for a new company to enter than a single-sourced one. Crowley described an OCI MSA-compliant ELS and a four-wavelength bidirectional part, with 8- and 16-wavelength versions already developed on the same platform. No Scintil release announces those, so they stand as company statements rather than launches.
Asked whether integration eases the indium phosphide shortage or leans on it, Crowley answered that it does both. Scintil buys InP gain blocks from the same epitaxy supply as everyone else, he said, but uses only a small coupon of the material, and the flow skips steps a conventional DFB needs: multiple depositions, e-beam gratings, cleaved facets, anti-reflective coatings. Those comparisons characterize a rival device, offered by a company selling the alternative. Scintil claims a roadmap to a tenfold reduction in the total InP wafer area needed per laser versus a traditional DFB, through smaller coupons and recycling. He also argued that wafer-level integration lets Scintil test and burn in before dicing, so only known-good die go into a module.
"Making a few samples of good multi-wavelength laser sources has never been the problem. Making millions of them, each one holding its channel plan over years of field deployment, is the real challenge."
The same reply concedes the gap. "Making a few samples of good multi-wavelength laser sources has never been the problem. Making millions of them, each one holding its channel plan over years of field deployment, is the real challenge," Crowley told SCN, and he said buyers who raise the thin field history of integrated lasers are right to. His answer is a reliability program built ahead of orders: parts made at Scintil's expense to get a test population large enough to mean something, in-house labs customers can visit, and the observation that it takes 5,000 hours to run a 5,000-hour test. By SCN's arithmetic, a million device-hours means two hundred parts held on test for all 5,000 of those hours, which is seven months of calendar time before any of the data exists. On field behavior, he points to precedent rather than Scintil data: Intel reported in June 2024 shipping more than 8 million photonic integrated circuits carrying over 32 million on-chip lasers, with a laser failure rate below 0.1 FIT, a different design on a different platform, as Crowley acknowledged. LEAF Light has been sampled, he said; ELSFP module builds are under way, and the first few hundred thousand chips are in process at the fab. Full product qualification, customer qualification, and a production module with Scintil's own control electronics stand between that and volume.
LEAF Light holds ±10 GHz across the operating temperature range, including aging, he said, a guarantee over life rather than a measurement under stated conditions, against about ±35 GHz tolerated by first-generation OCI. Scintil quotes 19 dBm off chip, which, after 2 dB of chip-to-fiber coupling loss, leaves 17 dBm in fiber, and says it has demonstrated 50 mW lasers in a result presented at ECOC 2026, with 100 mW designs in fab before year-end, 200 mW in 2027, and 400 mW on the roadmap. On reliability, the company describes testing on every lot, HTOL capacity of 12,000 units at a time, and labs in Grenoble and Santa Clara, reaching tens of millions of device-hours cumulatively; that capacity figure is what the ovens hold, not the population Scintil has run. It claims no FIT figure, saying the testing runs across multiple lots over several quarters and is not finished. The OCI MSA part comes out of fab in the first half of 2027, the control ASIC returns in the first quarter, field trials run in 2027, and deployments in 2028.
Lightmatter's Guide takes integration further. It puts many lasers, multiplexers, monitors, and spare emitters on one silicon photonics chip run by custom control ASICs, an approach the company calls very large scale photonics, or VLSP. Two products are public. Guide 1 is a 16-wavelength DWDM source on a 200 GHz grid, listed at 13 mW per wavelength and at least 100 mW per fiber, and it is sampling. Guide DR is a single-wavelength source on a liquid-cooled card inside the chassis, listed at 64 lasers and at least 200 mW per fiber, offered in a 1311 nm or a 1331 nm version; the May 21 release puts sampling in the fourth quarter of 2026, and the company's September technical brief says the part comes to market in 2027. No release names a Guide customer.
Harris's argument to SCN focused as much on manufacturing as on performance. He compared today's discrete InP laser assemblies, with their gold plating and wire bonds, to circuits wired from individual transistors before the integrated circuit. "They look like 1960s technology," he said.
Asked why he had cited about 128 lasers per chip when the Guide DR page lists 64, Harris separated the two parts: Guide 1 carries 128 lasers feeding 16 fibers at eight wavelengths each, while Guide DR is the single-wavelength part, 64 lasers, one per fiber, for parallel-fiber DR optics like the Passage L20. Spare lasers sit on the chip on top of both counts, with an optical switch to route around a failure. He did not give the sparing ratio.
He also answered the question Lightmatter's documents left open: where the III-V material comes from. Guide is built at silicon photonics foundries on 300 mm wafers, Harris said, with the InP gain material grown on an InP wafer, bonded to the silicon wafer and processed with it, the technique the industry calls heterogeneous integration. The foundries are unnamed. Read against the supply picture above, that means Guide still consumes InP epitaxy; what moves to the 300 mm silicon photonics lines is the laser processing and the testing, which he says those fabs have far more capacity to absorb than the InP industry does.
The business case follows from the shortage. "We just make money because there's no capacity, and we have new capacity," Harris said. Earlier in the interview, he said Guide could lift world laser supply by something like 30%. Asked afterward for the basis, he said the figure is an estimate built from the wafer starts Lightmatter's foundry partners can support and its die-per-wafer count, and that neither number can be shared publicly, which leaves the denominator undefined. On customers, he went no further than categories, and would not discuss engagements or launch plans before announcement. His interview claim that three large switch companies will ship platforms with Guide and the Passage L20 next year stands, with no name attached.
Wavelength accuracy is the one place where Lightmatter's public specification and its private measurement differ. Harris gave SCN the 4 GHz figure in the interview, and it is not the number Lightmatter publishes. Asked in writing which one describes the part, he explained the gap as a definition rather than a correction. The ±20 GHz on the January launch release and the Guide 1 page describes the absolute worst-case drift Lightmatter guarantees from the frequencies Guide 1 was designed to generate, he said, while the measured accuracy of the laser lines is ±4 GHz, a number he said the company could publish and has not. Full spec sheets go to customers under NDA, so the conditions behind the measured number - population, temperature, and elapsed time - are not public. For scale, first-generation OCI tolerates about ±35 GHz.
"When there's 10 million of anything running, everything that can go wrong will go wrong all the time."
Harris's case for tight control is reliability at fleet scale. A large AI data center will run something like 100 million optical links, he said, and lasers that wander in wavelength or power make links flap. "When there's 10 million of anything running, everything that can go wrong will go wrong all the time." Against that standard, his written answers give Lightmatter's account: he says FIT rates are extremely low, with 2 million device-hours on Guide 1 with zero hardware failures, and a validation data center running hundreds of the platforms continuously. The company does not publish FIT data or third-party test results, and he did not say whether a "device" in that count is a laser, a chip, or a whole source, or whether the hours were accelerated. No third-party measurement of a Guide part is public.
A comb generates many evenly spaced lines from one source. The device's physics fixes the spacing, so the lines cannot drift relative to one another. Xscape Photonics, a Santa Clara company, makes its comb optically. Its joint release with Tower on August 25, 2025 describes a multi-wavelength source built monolithically on Tower's PH18 silicon photonics platform and pumped by a single external CW laser, with no hybrid III-V integration on the chip. Xscape's first product, FalconX, announced March 11, 2026, is an ELSFP emitting up to eight wavelengths with more than 1 W of total optical power, which the company describes as OCI MSA-compliant; no compliance program exists to test that against. NVIDIA is among its investors, and in July 2026 the US Department of Energy's ARPA-E selected Xscape for its SCALEUP Ready program.
"The main remaining hurdle is the lack of long-term volume data as a new technology."
Co-founder and CEO Vivek Raghunathan told SCN that Lumentum arriving six months after FalconX shows how much momentum the MSA is gaining, and that hyperscale buyers want a multi-vendor ecosystem. His case for the comb is that it controls wavelength spacing and power flatness at the same time, uses what he calls the smallest InP footprint, needs fewer module-level components, and scales to 128 colors with a silicon re-spin. On the gap in his own case, he was direct: "The main remaining hurdle is the lack of long-term volume data as a new technology."
The comb's structural question is the pump, because every wavelength depends on it. Xscape's answer is pump redundancy: one spare pump backs all eight channels, whereas, in Raghunathan's characterization, a discrete approach needs eight more lasers to cover eight lines. Lightmatter uses on-chip spares with an optical switch; Lumentum uses a hot-swappable module; none of the three has published a sparing ratio. The pump is still a III-V laser, so the architecture reduces the InP a link needs without removing it, and Raghunathan said Xscape buys from multiple III-V strategic partners he did not name. FalconX has been validated in customer links, he said, and enters a qualification phase that collects six months of reliability data across multiple modules from a contract manufacturer, with qualification targeted for 2027 and production for 2028. He gave one power figure, a target of up to 28 dBm on a fiber carrying four wavelengths. OIF's ELSFP specification caps a fiber at 24.5 dBm, where automatic power reduction stands in for an interlock; Xscape's answer did not address eye-safety class, and Lumentum has not said how its eight lines map to fibers, so the two modules cannot be compared on that basis. Per-line power, flatness, pump-to-comb efficiency, FIT data, and customers remain unpublished.
Quintessent makes its comb electrically. The first product, sampling as an evaluation kit since August 2026, is a single-chip quantum-dot laser that emits eight wavelengths from one device under one bias control, using gallium arsenide gain material heterogeneously integrated on silicon photonics. The company pitches the material as independent of the InP supply chain and says the design needs no high-power pump laser, a line aimed at optically pumped combs.
Its gain material is the only one in this article outside the InP bottleneck, and Quintessent has the least public evidence behind it. Ciena is among its investors, and the company claims up to 40% lower data-movement power against fast single-lane links. Quintessent did not respond to a request for comment by publication. It has published no product-level independent data, no customers, and no production date.
The last approach deletes the external source. Tower's PH18DA platform integrates InP components on the silicon photonics wafer so that lasers, optical amplifiers, modulators, and detectors share one chip. On September 17, 2026, Tower and Israel's NewPhotonics said laser-integrated engine chips for 800G and 1.6T have begun high-volume shipment, with 6.4T chipsets for socketed near-packaged modules to follow in the first half of 2027. OpenLight's design kit for the same platform became available in Cadence tools in August. This approach moves the laser's reliability exposure into the package, which is the risk the external-source camp organized to avoid. TSMC's longer-term roadmap, in slides reported by Taiwanese trade press and covered by SCN on September 4, also places lasers on the engine, with no date.
OpenLight CEO Adam Carter told SCN that more than 35 customers are engaged with the platform across data center, automotive, and consumer electronics, including hyperscalers and Tier 1 companies. That figure appears in no public document, and one of those customers is public: NewPhotonics, which OpenLight named in a March 5, 2026 release as its first volume production customer on PH18DA. Carter said he cannot name others until they are ready to announce. Asked whether the OCI requirement for an external source keeps on-engine lasers out of scale-up, he declined the framing, saying the question is less about a particular MSA than about which architecture delivers performance and manufacturability. The MSA's answer, which arrived after that question was sent, makes the point moot: integrated sources are permitted.
On the field-history problem, Carter's answer runs on physics rather than data. NewPhotonics did not respond to a request for comment by publication, and Tower Semiconductor declined to comment, saying the executive SCN asked for was traveling. OpenLight's lasers use proven III-V material and have been through extensive accelerated life testing, he said, and they are relatively low-power devices, "not high-power lasers operating near fundamental limits." From a physics perspective, he added, nothing inherent about heterogeneous integration would make them less reliable, and eliminating optical interfaces and facets removes failure points a discrete part has. He offered no qualification data, FIT figure, or field result to back that up, and the low-power argument also shows how far this device class sits from the 400 mW switch laser at the other end of the article.
"Vendor-claimed" means the seller is the only source, including what it told SCN this week. "Outside corroboration" means a source other than the seller: a standards document, a paper, an analyst, a customer or partner statement, or field data. The last column uses only what the send log supports.
Approach | Who | Published power/wavelengths | Status, September 2026 | Vendor-claimed | Outside corroboration | Reply |
|---|---|---|---|---|---|---|
Single-line UHP InP DFB, external module (switch CPO) | Lumentum; Coherent; Broadcom (own switches, sourcing undisclosed); Shijia small batch; Furukawa papers | 1311 nm; Lumentum page: up to 350 mW at 50°C, up to 24 dBm in fiber; Coherent >400 mW at 55°C (2025 sample) | Shipping into scale-out CPO; Coherent UHP volume from Q4 2026; Lumentum says it is well behind on high-power shipments; Coherent calls InP capacity its primary constraint | 20 FIT per laser from accelerated test; capacity | OIF power classes and 50 FIT per-source target; 2022 Lumentum/Facebook paper; LightCounting: Lumentum qualified at NVIDIA, Coherent next, no others for two to three years | Lumentum, Coherent, Broadcom: did not respond |
UHP DFBs on the OCI grid, ELSFP | Lumentum | 8λ, two groups of four near 1311/1331 nm, ±0.2 nm, up to 24 dBm per λ, ~12 W | ECOC demo; initial availability targeted for 1H 2027 | All specifications | Grid matches OCI v1.0; MSA says the spacing was chosen so standard lasers can meet it | Did not respond |
InP DFB arrays, external module | Ayar Labs SuperNova with Sivers arrays | 16λ (SuperNova); Sivers 100 mW/channel 8-channel array; Glasgow in pilot manufacturing | Ayar Labs: volume date not public | System-level energy and bandwidth claims | Long field history of discrete InP DFBs | Ayar Labs, Sivers: did not respond |
DFB array bonded on silicon photonics | Scintil LEAF Light, Tower PH18M | 8λ at 200 GHz or 16λ at 100 GHz; FIT unpublished | Sampled; evaluation kit since March; ELSFP builds under way; first few hundred thousand units in process at the fab; buyers pointed at 2028 systems | OCI MSA-compliant ELS and 4λ BiDi part; 10x InP-area roadmap; wafer-level test and burn-in | OFC 2023 paper (about 1 mW per λ, four-λ prototype); Tower validation release; NVIDIA an investor; Intel's <0.1 FIT for its own integrated lasers, a different design | Replied (Matt Crowley, CEO) |
Integrated silicon photonics laser chip with control ASICs (VLSP) | Lightmatter Guide 1, Guide DR | Guide 1: 128 lasers, 16 fibers × 8λ on a 200 GHz grid, 13 mW/λ, ±20 GHz guaranteed. Guide DR: 64 lasers, one per fiber, 1311 or 1331 nm, ≥200 mW/fiber. Spares plus an optical switch on top, ratio undisclosed | Guide 1 sampling; Guide DR samples Q4 2026, to market in 2027 | ±4 GHz measured; 2M device-hours with zero hardware failures; FIT rates Harris calls extremely low; about 30% of world laser supply, inputs withheld; unnamed 300 mm foundries; InP bonded | Ecosystem releases with GUC, Qualcomm and NVIDIA, none about laser purchases. No paper, no public third-party data (Harris says results exist under NDA) | Replied (Nick Harris, CEO) |
Optically pumped comb on silicon photonics | Xscape FalconX, Tower PH18 | 8λ, more than 1 W total; target of up to 28 dBm per fiber with 4λ per fiber | Validated in customer links; six-month reliability run next; qualification 2027, production 2028 | Single redundant pump for all eight lines; smallest InP footprint; 128 colors with a silicon re-spin | Tower joint release; NVIDIA an investor; ARPA-E selection | Replied (Vivek Raghunathan, CEO) |
Electrically pumped quantum-dot comb on GaAs | Quintessent | 8λ, O-band | Evaluation kits sampling since August 2026 | 40% power saving; no InP in the gain material | Ciena among investors; quantum-dot laser literature | Did not respond |
Lasers integrated on the optical engine | Tower PH18DA: OpenLight PDK, NewPhotonics | Per-lane lasers at 200G per lane | 800G and 1.6T engines in high-volume shipment; 6.4T chipsets 1H 2027 | OpenLight: more than 35 customers engaged; accelerated life testing; says nothing inherent in heterogeneous integration makes the lasers less reliable | Tower/NewPhotonics shipment release; OpenLight's March 5 release names NewPhotonics; MSA says integrated sources are permitted | OpenLight replied (Adam Carter, CEO); NewPhotonics did not respond; Tower declined |
Only two rows of that table describe parts in announced volume shipment: single-line UHP lasers from the incumbents, and on-engine lasers from Tower's customers. Every multi-wavelength external source is at the demonstration, sample, or evaluation-kit stage, Lumentum's included, and the challengers' half of the table rests on manufacturing claims, with the supporting numbers withheld.
No supplier has published a FIT figure for a multi-laser module. Lumentum's white paper separates OIF's module-level target from its own per-laser rating and does not give the module number. By SCN's arithmetic, eight lasers at the rated 20 FIT, with no sparing and any single failure counted, come to about 160 FIT, against OIF's 50 FIT for a source. The same sum applies to the eight-laser modules feeding NVIDIA's switches today. The 20 FIT rating is an upper bound from a test with zero failures, so the sum describes what has been demonstrated so far and says nothing of how the parts will behave in the field. SCN asked Lumentum how a module-level figure is derived for an eight-laser ELSFP, whether through sparing, derating, or a better per-laser number, and had no answer by publication. For Scintil, Xscape, and Quintessent, there is no number to do arithmetic on yet, and Lightmatter's claimed 2 million device-hours cannot be converted into one without knowing what it counts as a device.
Scintil's arrays, Xscape's comb, and the OpenLight and NewPhotonics engines all run on Tower silicon photonics platforms, so three challenger approaches share one foundry, and none of the companies has said how much capacity it has committed to each. Tower declined to comment. Lightmatter claims its chips run at other 300 mm silicon photonics foundries it will not name.
On test, Harris told SCN that the infrastructure for CPO at GPU scale does not exist yet and that about three companies make the testers, none of which he named. As SCN reported in August, the Open Compute Project's map of the silicon photonics stack, a 294-page paper from 19 member companies, names photonic die test as the one layer no organization yet owns.
Near-term timing favors the incumbents. Lumentum said in February that it had a multi-hundred-million-dollar purchase order for UHP lasers for scale-out, shipping in 2027, and told investors in August that it expects the UHP ramp for scale-up in the second half of 2027, ahead of customer deployments in 2028, with customers other than its lead CPO customer prioritizing near-packaged optics first. The challengers' dates cluster in the same window; evaluation kits for the socketed Passage L20 module are anticipated in the first quarter of 2027. Harris agrees on the order of events and expects CPO scale-up by the end of 2028 or in 2029. Neither NVIDIA nor Broadcom publishes unit volumes.
The second-generation OCI text will not sort this out. Its stated goal is to keep the wavelength spacing where it is, and the MSA gave no publication date, so the precision argument remains unsettled, and the spec's evidence does not decide the first generation. What buyers say they need instead is field reliability data, and no supplier of a multi-wavelength source, incumbent or challenger, has published any.
Disclosure: Scintil Photonics is a client of OmniScale Media, a marketing agency co-founded by Matt Walters, who is also the publisher of Supercomputing News. OmniScale Media and Supercomputing News are separate companies. Walters conducted the interviews and requested comments. No other company named in this article is an OmniScale Media client.
Matt Walters is the founder and publisher of Supercomputing News. He also runs OmniScale Media, the marketing agency he co-founded in 2017 to serve AI, HPC, quantum, and deep tech companies.
He's spent 15+ years in this world. Seven of them at Tabor Communications as VP of Digital Strategy, where he grew audience and sponsorships for HPCwire and helped launch Datanami (now BigDATAwire) and EnterpriseTech (now AIwire). Along the way he's built hundreds of campaigns for tech companies of all sizes - from early-stage startups to NVIDIA, Intel, IBM, HPE, and others... including NVIDIA's early push to sell GPUs for AI, back when that was still a bet.
A builder at heart, he spent 15 years in the construction trades before any of it.
