Optical interconnects: which startup is ahead?

Which optical interconnect startups actually matter right now?

The serious optical interconnect startup field currently has eight independent contenders, with Ayar Labs and Lightmatter clearly separated from the rest.

We define optical interconnect startups as companies developing technology to move data between AI accelerators, switches, memory systems, boards or racks using light. This includes optical I/O chiplets, co-packaged and near-packaged optical engines, integrated laser sources and optical circuit switches built specifically for AI infrastructure.

We exclude large public suppliers such as Broadcom, Marvell, Coherent, Lumentum and Ciena. We also leave out conventional transceiver manufacturers, telecom equipment companies, lidar developers, quantum-photonics businesses and startups mainly building photonic processors rather than connections between processors.

Ayar Labs and Lightmatter form the top group. Ayar has now raised just over $1 billion and Lightmatter about $850 million. The next-best-funded independent company, Salience Labs, has raised more than $155 million.

Ranovus still leads the next group on deployment readiness. Its integrated ODIN engine, manufacturing agreement with Jabil and work with Cerebras put it ahead of several better-funded startups, even though it has announced little new since 2025.

Avicena is pursuing microLEDs rather than conventional laser-and-modulator systems. Xscape Photonics, SCINTIL Photonics and Quintessent focus mainly on the lasers that provide multiple wavelengths of light. Salience Labs operates one layer higher, using silicon photonics to switch optical traffic between computing systems.

Celestial AI and Nubis Communications remain useful benchmarks even though they are no longer independent. Marvell acquired Celestial, while Ciena bought Nubis.

Funding totals are approximate. Private companies disclose grants, extensions and strategic investments unevenly, so small differences should not be treated as exact rankings.

Startup	What it is building	Approximate capital raised	Where it stands
Ayar Labs	Optical I/O chiplets and remote light sources	About $1.02 billion	Live rack demo; volume qualification targeted for end-2027
Lightmatter	Photonic interposers, optical engines and laser systems	$850 million	Validation platforms; low-volume shipments planned for 2027
Avicena	MicroLED optical interconnects	About $120 million	Shipping 1 Tbps evaluation kits
Xscape Photonics	Multi-wavelength comb lasers	About $94 million, plus an ARPA-E award	First commercial laser module launched
SCINTIL Photonics	Integrated DWDM laser sources	About $77 million	Evaluation kits and foundry-qualified production path
Ranovus	Integrated CPO and NPO optical engines	About $70 million disclosed	Moving toward mass production; no shipment data
Salience Labs	Optical circuit switches	More than $155 million	32-port product launched; 64-port switch demonstrated
Quintessent	Quantum-dot lasers for silicon photonics	About $12 million	Early product and supply-chain development
Celestial AI	Photonic Fabric optical chiplets	At least $515 million before acquisition	Acquired by Marvell
Nubis Communications	Compact optical engines	About $50 million before acquisition	Acquired by Ciena

Is Ayar Labs’ reported $5 billion valuation a real lead over Lightmatter?

Yes, but a narrow one. Ayar Labs is ahead because its modular chiplet is easier to adopt and closer to qualification, not because it has better raw bandwidth. Lightmatter still owns the biggest headline number.

The money now leans toward Ayar. It closed a $500 million Series E at a $3.75 billion valuation in March 2026 and added $150 million in September. A separate $225 million secondary sale, which brought no new cash to the company, reportedly valued Ayar above $5 billion. Ayar has not confirmed that figure; the last valuation it disclosed is the $3.75 billion set in March. Lightmatter’s last disclosed valuation is $4.4 billion, set in an earlier round.

Lightmatter wants photonics to become part of the package itself. Its Passage M1000 is a 4,000-square-millimeter photonic interposer offering 114 Tbps of total bandwidth, 256 fiber connections and built-in optical switching.

Ayar sells an 8 Tbps optical I/O chiplet that can sit beside an accelerator or switch. Several TeraPHY engines can surround the same processor, and the package design shown with Alchip and GUC exceeds 100 Tbps per compute package.

Lightmatter leads the single-platform benchmark, while Ayar reaches nearly the same package-level range through smaller building blocks. Ayar’s UCIe interface also lets accelerator designers add optical links without rebuilding the whole package around one photonic platform.

Lightmatter has broadened its offer with co-packaged engines and laser cards. At the OCP Global Summit in October 2026 it showed a 204.8 Tbps co-packaged switch tray built with Flex and a 102.4 Tbps compute tray. The switch tray is a production-validation system and the compute tray a prototype; neither is a product on sale.

Ayar currently asks customers to change less. Lightmatter could move ahead if a major accelerator company or hyperscaler publicly puts Passage into production.

Company and product	What the figure measures	Disclosed bandwidth	How to read it
Lightmatter Passage M1000	Complete photonic interposer	114 Tbps	Largest platform-level figure; mass-production readiness planned for late 2028
Ayar Labs with Alchip and GUC	Complete accelerator design	More than 100 Tbps	Uses several optical chiplets; shown in a live rack
Celestial AI Photonic Fabric	One optical chiplet	16 Tbps	Now owned by Marvell
Ayar Labs TeraPHY	One optical I/O chiplet	8 Tbps	Modular building block
Ranovus ODIN	Integrated CPO engine	6.4 Tbps	Includes an integrated laser
Avicena LightBundle eKit	Evaluation link	1 Tbps	Raw throughput from 335 microLED channels

Which optical interconnect startup has found real customers?

Ayar Labs has the best evidence of customer pull in optical interconnects, while Ranovus still has the clearest named deployment project.

No independent startup in this comparison discloses optical-interconnect revenue, shipment volume, backlog or a list of paying production customers. None has publicly proved a large commercial business. With no startup disclosing paying production customers, we look at who the buyers are and which problems they would pay to solve in our report on optical interconnects.

Ayar works with Alchip and GUC, which design custom accelerators and switches for cloud companies. Wiwynn, which builds servers and racks for data centers, went from partner to strategic investor in 2026. Ayar’s strategic backers now also include AMD, Intel, MediaTek and Nvidia.

Those names cover nearly the whole path from chip design to rack delivery. They still do not prove volume orders. Ayar itself refers only to unnamed “customer programs worldwide.”

Ranovus has fewer public relationships, though its Cerebras project is unusually concrete. The companies are working on a DARPA-backed system connecting wafer-scale processors with wafer-scale co-packaged optics.

Lightmatter says its Guide lasers and Passage engines are being evaluated at the largest hyperscalers, foundries and chip designers, and that two major players are working on M1000 designs. It has not named them. Its partner list keeps growing, from GlobalFoundries, Amkor, ASE and GUC to Flex, TSMC and Tower.

Celestial AI had the strongest customer claims before Marvell acquired it. Marvell said several hyperscalers were working with the technology, but meaningful revenue was still expected later.

Ayar leads because its route to customers is broader and now partly backed by those partners’ own money. Ranovus has the strongest single named project.

Which optical interconnect startup is closest to production at scale?

Ayar Labs is closest to large-scale production, with Lightmatter now ahead of Ranovus on visible manufacturing work, but behind Ayar on timing.

Ayar’s TeraPHY engine has gone well beyond a lab demonstration. At the OCP Global Summit in October 2026, Wiwynn ran a live co-packaged optics rack built from a GUC accelerator package, Ayar TeraPHY engines and Ayar’s SuperNova external lasers, with detachable fiber connectors for field servicing.

That is the strongest system-level proof any independent startup has shown. It is still a demonstration rack, not a customer deployment.

Ayar’s timeline is explicit. Its CEO said in March 2026 that customers aim for volume ramps around 2028, and that Ayar’s products must be qualified by the second half of 2027. Reuters later reported a target of volume-production qualification by the end of 2027. The company is also opening a design center in Bengaluru, alongside its teams in the US and Hsinchu, Taiwan.

Lightmatter has the broadest manufacturing network. Intel, GlobalFoundries and Tower can fabricate its photonics, Amkor and ASE handle assembly, and Flex is using Lightmatter’s switch tray to validate its own photonics production line. Lightmatter told reporters in October 2026 that low-volume co-packaged shipments start in 2027 and high-volume deployment in 2028, with full-scale production running into 2029.

Ranovus has demonstrated a 6.4 Tbps ODIN engine with an integrated laser and engaged Jabil to support mass production. It also announced an investment of more than $100 million to expand its Ottawa operations. It has not disclosed shipments since.

Ayar stays ahead because its chiplet, packaging partner, rack builder and production funding now meet in one working system. Lightmatter has more manufacturing partners but a later date for its flagship interposer.

Which optical interconnect startup uses the least power?

Avicena has the most exciting low-power result in optical interconnects, while Ranovus has published the cleaner full-engine measurement.

Avicena reported transmitter consumption of 80 femtojoules per bit for its microLED arrays. Ranovus reported roughly 4 picojoules per bit for its 6.4 Tbps ODIN implementation, including the laser.

Avicena’s figure is 50 times lower, but the measurements cover different parts of the link. Its number applies to the transmitter, while a working connection also needs photodetectors, receiver electronics, control systems and fiber coupling. Transmitter numbers are not whole links, so we separate what works today from what is still on paper in our report on semiconductors.

Ranovus measures far more of the finished optical engine. Its result is less dramatic but easier to connect to real system power.

Avicena sends data directly from arrays of tiny LEDs into a multicore fiber bundle. Conventional silicon photonics generally begins with a laser and uses modulators to encode the data.

Removing separate lasers, modulators and wavelength-control components could cut power and simplify short links. In August 2026 Avicena began shipping 1 Tbps LightBundle evaluation kits, using 335 microLED channels at up to 3 Gbps each. In September it showed a connectorized version at ECOC. It did not publish a power figure for the new kit.

Ayar and Lightmatter publish broader efficiency claims, though their system boundaries and workload assumptions differ too much for a fair comparison.

Avicena leads on technical potential. Ranovus has the strongest disclosed result for a more complete optical engine.

Which optical interconnect startup has the strongest moat?

Lightmatter has the hardest optical interconnect architecture to copy, while Ayar Labs has built the stronger moat around customer adoption.

Lightmatter’s Passage platform combines photonics, packaging, power delivery, thermal design, fiber attachment and optical switching inside one large structure. A competitor would need expertise across all those areas and support from several foundry and packaging partners.

Its multi-reticle interposer is especially difficult to reproduce. Lightmatter is also trying to set the rules around it: in August 2026 it launched an open silicon photonics initiative under the Open Compute Project, starting with 19 companies, including Dell, Qualcomm and Foxconn Interconnect Technology.

Ayar’s individual components may be easier to understand, but the surrounding system is becoming harder to match. TeraPHY combines UCIe, dense wavelength multiplexing, microring modulators and a remote light source.

Ayar then adds reference designs with ASIC companies, packaging suppliers and rack manufacturers. A competing modulator alone would not recreate that integration work.

Avicena has the most distinctive underlying technology. Its microLED approach changes the emitter, receiver, fiber and packaging together, so a conventional silicon-photonics company cannot copy it through a minor process update.

Xscape, SCINTIL and Quintessent have narrower laser moats. They could become important suppliers across several competing optical platforms without controlling the whole interconnect stack.

Lightmatter wins on technical difficulty and standard-setting. Ayar’s advantage is easier for customers to use today.

Is Salience Labs now the fastest-rising optical interconnect startup?

Among the smaller players, yes. Salience Labs is gaining ground fastest, while Ayar Labs still has the strongest deployment momentum overall.

Salience raised an oversubscribed $125 million Series B in October 2026, led by Bullhound Capital, with AMD Ventures and Intel Capital among the new investors. A month earlier, it demonstrated a 64-port all-optical circuit switch at ECOC, double the port count of the 32-port switch it launched in March.

The money will go to volume manufacturing with Tower Semiconductor, a Silicon Valley office and 128- and 256-port products. Salience has not named a customer. Its switch also sits in a different part of the network from the optical I/O beside an accelerator.

Ayar’s progress has been more concentrated. It went from an architecture shown in March to a live rack at OCP in October, and it raised fresh capital twice in 2026.

Lightmatter has released the most products. It launched the Guide DR liquid-cooled laser card in May 2026, set up the open photonics initiative in August, added a 6.4 Tbps bidirectional Passage L20 engine for the Open CPX standard in September and showed two optical trays in October. That shows technical depth, but it also means several product families to industrialize at once.

Xscape is turning FalconX into a manufacturing story. In July 2026, ARPA-E selected it for a program to scale FalconX production in the US. In May it helped launch a multi-source agreement group whose members include AMD, Cisco, Meta and Oracle.

Avicena doubled its evaluation-kit bandwidth to 1 Tbps. SCINTIL has not announced new milestones since launching its LEAF Light evaluation kit.

Ayar has the best deployment momentum. Salience made the biggest jump among the challengers, though it was in funding and port count rather than shipments.

Did Celestial AI already win the optical interconnect race by selling to Marvell?

Celestial AI won the optical interconnect exit race, but Marvell’s revenue timetable shows that the technology had not reached volume sales.

Marvell agreed in December 2025 to pay about $3.25 billion upfront for Celestial AI, in cash and shares, and completed the deal in February 2026. Additional payments of up to about $2.25 billion in shares depend on Celestial reaching cumulative revenue targets by the end of Marvell’s 2029 fiscal year. Deals like this show where investors put money, and we look at what they back and avoid in our semiconductor market report.

Celestial had raised at least $515 million and developed a Photonic Fabric chiplet offering 16 Tbps. Marvell said several hyperscalers and other companies were already working with the technology.

Marvell expects meaningful revenue from Celestial only in the second half of its 2028 fiscal year. It projected a $500 million annualized run rate near the end of that year and $1 billion near the end of the following one. Those are forecasts, not results.

The performance-based payments reflect the remaining commercial risk. Customer designs still need to become products and then reach volume.

Nubis provides a smaller comparison. Ciena acquired the company after it developed compact 1.6 Tbps optical engines and has since introduced a 6.4 Tbps product based on Nubis technology.

Celestial’s sale proves that optical scale-up technology can command a multibillion-dollar price before large revenue appears. Ayar’s reported valuation above $5 billion now suggests investors see an independent path as at least as valuable.

Can optical interconnect startups survive Nvidia, Broadcom and Marvell?

Optical interconnect startups can survive Nvidia, Broadcom and Marvell, but most will do so as suppliers or acquisition targets rather than full-stack giants.

Large semiconductor companies already control many of the chips, lasers, foundries and customer relationships needed to bring optical interconnects into AI systems.

Nvidia’s Quantum-X Photonics co-packaged switches are shipping in 2026, with Spectrum-X Photonics Ethernet to follow. Nvidia has also opened NVLink Fusion to outside chip designers, and both Ayar and Lightmatter have joined that program.

Broadcom offers Tomahawk 6 Davisson, a 102.4 Tbps co-packaged optics switch. Coherent launched its PhotonLink platform in September 2026. Marvell now combines custom AI chips, networking products and Celestial AI’s Photonic Fabric, and showed 2-nanometer optical technology at ECOC in September 2026.

Ciena followed the same path by acquiring Nubis and using its technology inside compact optical products for AI data centers.

Startups still have room where new technology moves faster than the incumbents’ internal roadmaps. The incumbents are also funding them: AMD, Intel and Nvidia back Ayar, and AMD and Intel joined Salience’s Series B.

Ayar can remain independent by supplying chiplets across several processor architectures. Ranovus can operate as a specialized optical-engine supplier. Xscape, SCINTIL and Quintessent can sell lasers across competing platforms.

Startups relying on one impressive demonstration are more exposed. Those already connected to foundries, packaging companies and AI-chip programs have a realistic place in the market.

Who leads each optical interconnect segment today?

Optical interconnects currently have several specialist leaders, but Ayar Labs is the only independent startup near the front across most of the stack.

Ayar leads modular optical I/O for accelerators and switches. Lightmatter leads in large photonic interposers. Ranovus has the strongest position in tightly integrated optical engines containing electronics, photonics and laser technology.

Avicena effectively owns the funded microLED-interconnect category. Xscape and SCINTIL lead different approaches to producing several laser wavelengths, while Quintessent remains an earlier challenger using quantum-dot materials.

Salience leads optical switching among the startups in this comparison, and now has the money to stay there. Its product routes light between systems rather than providing the connection leaving the processor package.

These companies will not always compete directly. A future AI rack could combine an Ayar optical engine, a Xscape or SCINTIL light source and a Salience optical switch.

Optical interconnect segment	Current leader	Closest challenger	Why the leader is ahead
Modular optical I/O chiplets	Ayar Labs	Celestial technology inside Marvell	Standard interfaces, live rack demo and the most production funding
Photonic interposers	Lightmatter	No close independent equivalent	Largest and most advanced 3D photonic package
Integrated CPO and NPO engines	Ranovus	Nubis technology inside Ciena	Integrated laser and electronics with a mass-production partner
MicroLED optical links	Avicena	No direct funded equivalent	Distinct architecture, 1 Tbps kits shipping and very low transmitter power
Comb-laser modules	Xscape Photonics	Quintessent	First eight-wavelength commercial module and US manufacturing support
Integrated DWDM lasers	SCINTIL Photonics	Quintessent	Foundry-qualified process and customer evaluation kit
Optical circuit switches	Salience Labs	Large incumbent systems	64-port switch demonstrated and $125 million Series B
Full Photonic Fabric platform	Celestial technology inside Marvell	Ayar Labs and Lightmatter	Strong hyperscaler interest and multibillion-dollar strategic validation

Which optical interconnect startup is actually ahead?

Ayar Labs is the optical interconnect startup ahead overall, with Lightmatter second and Ranovus a weakening third.

Ayar takes first place because it combines a mature optical engine, a standard electrical interface, complete accelerator designs, a live rack built with Wiwynn and GUC, and the largest production budget of any independent startup.

It still has not disclosed shipment volumes or a named production customer. Neither has anyone else. Its volume ramp depends on qualification by the end of 2027, so the lead is real but not yet locked in.

Lightmatter comes second. Its package-level performance is stronger and its architecture may eventually prove more important. But its own dates put volume co-packaged shipments in 2028 and Passage M1000 production readiness in late 2028.

Ranovus takes third on its 6.4 Tbps integrated engine, its Jabil agreement and the Cerebras project. It has shared almost nothing new, and a challenger with fresh shipments could pass it within a year.

Avicena ranks fourth. Its 1 Tbps kits are now with customers, which narrows the gap, but bandwidth and system maturity remain behind the first three companies.

Salience moves up to fifth. Its $125 million round and 64-port switch make it the best-funded challenger, though it works in an adjacent part of the network and has no named customer.

Xscape takes sixth as the strongest laser specialist. FalconX could sell across several optical architectures, and ARPA-E support helps it scale production.

SCINTIL follows with integrated DWDM technology, a Tower manufacturing path and a customer evaluation program.

Quintessent ranks eighth. Its quantum-dot laser technology could become valuable, but the company remains less funded and less commercially advanced than Xscape or SCINTIL.

The ranking would change quickly if Lightmatter named a major production customer, if Ranovus showed substantial ODIN shipments, or if Avicena demonstrated a complete microLED link at multi-terabit bandwidth.

The gap between Ayar and Lightmatter is real and widened during 2026. Ayar leads by one stage of commercialization, not by a generation of technology.

Rank	Startup	Why it ranks here
1	Ayar Labs	Live rack demo, most production funding and a dated qualification target
2	Lightmatter	Strongest architecture and performance, with volume production planned for 2028
3	Ranovus	Integrated engine, manufacturing partner and named project, but little new evidence
4	Avicena	Most disruptive low-power architecture, with 1 Tbps kits now shipping
5	Salience Labs	Best-funded challenger and a 64-port optical switch in an adjacent layer
6	Xscape Photonics	Multi-wavelength laser specialist with a launched product and US manufacturing support
7	SCINTIL Photonics	Foundry-qualified DWDM laser technology and a credible evaluation path
8	Quintessent	Promising quantum-dot technology at an earlier commercial stage