AI interconnect and networking startups

Arrays of tiny microLEDs take the place of lasers in Avicena’s LightBundle links. Data travels through multicore fiber over hundreds of parallel channels, connecting processors, memory and switches within packages or across boards and racks. The Sunnyvale company began shipping evaluation kits carrying 1 Tbps, with up to 335 channels at 3 Gbps each, in August 2026. It showed a version with a detachable fiber connector at ECOC in September 2026. These remain evaluation kits; volume production has not been announced.

Ayar Labs puts the optical engine inside the processor package. Its TeraPHY chiplet sits alongside a GPU or switch chip, exchanges data with it through UCIe and sends that data over fiber, with reach up to 2 km. The laser sits elsewhere: a separate pluggable SuperNova module supplies the light. That division gives Ayar a role in both the packaged optical connection and the external source feeding it. The company is based in San Jose, California.

In March 2026, Ayar called its co packaged optics production ready and said it would expand high volume production and test capacity. Its September 2026 update described a transition from technology development to commercialization and production, with customer programs under way. Ayar has not published shipment volumes.

Financing has accompanied that transition. The March 2026 release reported total funding of $870 million and a $3.75 billion valuation. A $500 million Series E led by Neuberger Berman in March 2026 was followed by another $150 million in September, bringing primary capital raised during 2026 to $650 million. The stated work ahead is production and test capacity, alongside those customer programs.

Thin film lithium niobate is the material behind HyperLight’s optical modulator chips, which encode data onto light at low drive voltage. Transceiver makers incorporate its TFLN Chiplet platform into pluggable modules and co packaged optics. The Cambridge, Massachusetts company offers chips supporting up to 448 Gbps per lane; its funding release distinguishes 200G per lane products on offer from 400G per lane chips in sampling. An $80 million Series C led by MediaTek in June 2026, with UMC Capital, Jabil and Foxconn participating, supports manufacturing capacity and customer qualification.

Polymer waveguides connect optical components to fibers in Lessengers’ Direct Optical Wiring process, replacing the lens arrangements used in other modules. Its products include 800G OSFP transceivers and variants for linear and retimed links for data center servers and switches. The current site describes transceiver variants as available. Its company history records sampling of data center transceivers and engines alongside production of earlier modules. Work on optics mounted near processors and switches also appears in the portfolio, but demonstrations of those assemblies do not establish commercial supply. Lessengers is headquartered in Pohang, South Korea, with production in Gwangju.

Passage is Lightmatter’s family of silicon photonics products for carrying data between AI processors and switches. The Passage L20 optical engine carries 6.4 Tbps in each direction and mounts on the circuit board beside a processor or switch. In its March 2026 announcement, Lightmatter said sampling was expected to begin in late that year. The Passage L20 CPX module, announced in September 2026, has evaluation kits expected to ship in the first quarter of 2027.

Customer access is further along for reference platforms. The M1000 reference platform is offered as an evaluation kit, and Lightmatter says reference platforms are already in customers’ hands. That does not move the announced sampling or evaluation dates for the other products forward. The Mountain View, California company lists $850 million raised and a $4.4 billion valuation on its about page, without dating those figures there.

LightSpeed Photonics’ LightKonnect module is soldered near a processor and uses VCSEL lasers. It launched at OFC in March 2026. The module carries 400 Gbps across four 100G channels. The Singapore company’s website references evaluation platforms, but gives no sampling or shipping date.

Lumilens has reported customer delivery for its initial optical transceiver. In August 2026, it said its first scale out product had completed qualification and was shipping into a hyperscaler’s production AI data centers under a supply agreement. That evidence covers the initial transceiver, rather than every type of optical connection in development.

Its LumiCore platform combines silicon photonics, mixed signal circuits and optical packaging across pluggable modules and optics mounted nearer processors. The company addresses networks between AI servers as well as direct accelerator connections. Near packaged and co packaged optics remain a separate development program. Lumilens also develops its own manufacturing processes and test automation.

The San Jose, California company announced more than $700 million in new financing in August 2026. Named investors included Atreides Management, Bain Capital Ventures, Meritech, Seligman Ventures and Spark Capital.

Built in lasers are central to NewPhotonics’ photonic integrated circuits for transceivers linking AI servers and switches. Foundry partner OpenLight said NewPhotonics placed its first volume production orders for 800G and 1.6T laser integrated chips in March 2026, using its process with Tower Semiconductor. Gemtek subsequently introduced an AiPhoton 1.6T OSFP transceiver built on the NPG10204 chip in May 2026. In the interview posted on that page, NewPhotonics expected validation to finish in the second half of 2026 and mass production in 2027. Foundry orders therefore precede the stated mass production timetable. The company is based in Petah Tikva, Israel.

NLM Photonics has moved from supplying organic electro optic materials to designing active photonic integrated circuits. Its Selerion materials help modulators encode data onto light for transceivers connecting AI servers and switches. The Seattle company’s September 2026 announcement confirms customer sampling of 1.6T and 3.2T PICs. It also reports investment from Pangaea Ventures and Mitsubishi Chemical’s Diamond Edge Ventures, alongside existing backers, without disclosing an amount. Manufacturing qualification and commercialization remain part of the work funded; sampling does not establish volume manufacturing.

Customers can use OpenLight’s process design kit to design their own photonic chips, or commission OpenLight to design and make them. Its technology integrates lasers onto silicon using Tower Semiconductor’s PH18DA process, and more than 25 companies use the kit. In March 2026, OpenLight made alpha samples of its 3.2T DR8 transmitter available, with beta samples expected in the fourth quarter; the 1.6T DR8 chip was in beta sampling with qualification orders. NewPhotonics placed the first volume production orders on the process that March. OpenLight is headquartered in Santa Barbara, California.

Dublin’s Pilot Photonics launched a 16 channel O band laser array in March 2025 to feed external laser modules for co packaged optics and optical input and output chips. No sampling date has been published for those arrays. The same release said its nano-iTLA tunable laser for coherent links was sampling to customers.

A single comb laser supplies several wavelengths in Quintessent’s approach, giving a fiber multiple data channels without a bank of separately controlled lasers. The design integrates quantum dots on gallium arsenide with silicon photonics. Its QCOMB-1310-08-08-200-EVK evaluation kit produces eight wavelengths and began customer sampling in August 2026. Customers can use it in test beds while qualification and manufacturing work continues. Based in Santa Barbara, California, Quintessent announced a $40 million Series A led by Cycle Capital in August 2026, with Ciena among the new investors.

The ODIN optical engine combines lasers, modulators, detectors, drivers and amplifiers on one electronic and photonic chip. Ranovus announced a Jabil manufacturing collaboration in March 2025 intended to enable high volume production. Its news page lists no later sampling or shipping news; its about page describes near packaged optics as current and co packaged optics as the next step.

LEAF Light supplies closely spaced laser wavelengths to co packaged optics in AI scale up networks. Scintil Photonics builds the lasers onto silicon photonics wafers through its SHIP process on Tower Semiconductor production lines. A March 2026 evaluation kit announcement specified up to eight fiber outputs and targeted the second quarter. The subsequent September 2026 release describes customer access through an early access evaluation program; volume shipments have not been announced. Scintil is headquartered in Grenoble, France.

Inside linear optical modules, TeraSignal’s TS8401 and TS8402 drivers drive the devices that encode data onto light. The Irvine company announced customer and partner samples in March 2024, then offered TSLink reference designs with firmware and driver chips to selected customers.

Xscape Photonics supplies the light for optical links between GPUs, switches and memory. Its FalconX external laser uses the pluggable ELSFP format and produces up to eight O band wavelengths with more than 1 W of optical power. Built on the ChromX platform, it serves co packaged and near packaged optics. In a 2026 update, the chief executive said FalconX production was ramping, without publishing shipment figures. Xscape also announced an EagleX evaluation kit for ChromX. It has offices in Fort Lee, New Jersey and Santa Clara, California.

Software controls the paths through iPronics’ Optical Networking Engine, choosing which fiber inputs connect to which outputs on a silicon photonics chip. The switches rearrange optical connections within and between racks of AI servers. Its ONE-32 switch has 32 ports and was launched with availability from May 2025.

The Valencia, Spain company reported customer wins in March 2026 and announced a dedicated packaging, assembly and testing line at Fabrinet. That line is intended to support a production ramp. The announcement describes manufacturing expansion and customer activity, without providing shipment volumes.

A $125 million Series B in September 2026 was led jointly by Maverick Silicon and Light Street Capital, with NVIDIA joining the investors. Commercial deployments, hiring and expansion in Silicon Valley are the stated uses of the financing.

Operating as nEye.ai, nEye Systems combines silicon photonics, MEMS mechanical switches and CMOS electronics in an optical circuit switch on a chip. It aims to rearrange connections between processors and memory as workloads change. The product remains in development, moving from the lab toward manufacturing; no customer samples have been announced in the reviewed sources.

Salience Labs’ optical circuit switches rearrange links between AI accelerators across racks. They change the path light takes through the network without converting traffic to electrical signals at each switching point. The architecture is designed to work with existing transceivers and infrastructure, and Salience is working with Tower Semiconductor and Keysight on manufacturing and testing.

The Oxford, England company announced availability of its 32 port silicon photonic switch in March 2026. A larger 64 port switch was demonstrated in September 2026. Those are different delivery stages: the larger switch’s demonstration does not establish the same availability as the smaller product.

Salience announced a $125 million Series B led by Bullhound Capital in October 2026, with AMD Ventures and Intel Capital participating. It plans to expand manufacturing and its American operations with the funding.

Baya Systems licenses the fabric that moves data between processor cores, memory and input and output blocks, both within chips and across chiplets in a package. WeaveIP supplies that network; WeaverPro models, checks and configures it before a chip design is finished. In July 2026, Kandou AI licensed both for development of its next connectivity chips. Baya also works with Aion Silicon and Openchip on chiplet designs. Based in Santa Clara, California, it raised more than $36 million in a Series B led by Maverick Silicon in January 2025, with Synopsys, Matrix Partners and Intel Capital participating.

Eliyan’s physical layer circuits connect chiplets within packages and chips across circuit boards. NuLink PHY IP supports links between dies on standard organic packages that would otherwise need costlier silicon interposers, and supports standards including UCIe. NuLink-XS extends the approach to separately packaged chips across a board. NuGear chiplets use the same underlying technology, while newer NuLight chiplets drive microVCSEL optical links in modules.

The delivery evidence is broader than a named customer shipment. Eliyan’s July 2026 funding release calls NuLink PHYs and NuGear chiplets silicon proven and says hyperscale, AI accelerator and memory customers are evaluating or deploying its technology. It does not name those customers. That wording should stay attached to the stated products rather than becoming a claim that the whole portfolio is shipping in volume.

The Santa Clara, California company completed a $145 million Series C at a $1 billion valuation in July 2026. Seligman Ventures led, with Cisco Investments and Lumentum joining.

For customers designing chiplets, EXTOLL supplies licensable physical interface circuits. Its 16G UCIe PHY IP for GlobalFoundries 22FDX and 22FDX+ connects chiplets using standard packaging and organic substrates. EXTOLL announced it was available and ready for customer integration in May 2026. Customers incorporate the design into their own chips. Applications include edge AI, physical AI, robotics and communications, alongside automotive and aerospace. Based in Mannheim, Germany, EXTOLL began in high performance computing and also offers serializer and deserializer interfaces and networks on a chip.

Copper carries the data in Kandou AI’s chips and interface IP. Its Chord signaling codes data across several wires at once for links between processors, memory and accelerators. The Regli retimers clean up and resend signals inside servers, extending how far those signals can travel.

Availability differs within that family. The Regli page lists the 16 lane KB9003 in production, the 8 lane KB9002 sampling and the 4 lane KB9001 coming soon, supporting PCIe 5.0 and CXL 2.0. Separately, the Tigerwing chip to chip interface ran at 260 Gbps over 200 mm of circuit board in a May 2026 silicon demonstration. Tigerwing is not yet a product; that demonstration should be read separately from Regli’s production and sampling status.

Headquartered in Saint Sulpice, Switzerland, Kandou AI closed a $225 million Series A led by Maverick Silicon in March 2026, with SoftBank, Synopsys, Cadence and Alchip participating.

Cornelis Networks connects servers in AI and supercomputing clusters through Omni-Path network adapters and switches. Its CN6000 SuperNIC is an 800G adapter that also runs RoCEv2 and Ultra Ethernet over PCIe 6.0. This gives the portfolio both its existing fabric technology and support for those Ethernet protocols.

In September 2026, Cornelis said CN5000 adapters and switches were shipping, while CN6000 was sampling with customers, with wider availability expected in the fourth quarter. The expected availability remains a timetable, rather than confirmation that the newer adapter has reached general supply.

The same announcement introduced Active Compute Fabric, an architecture extending Cornelis into scale up networking within racks through UALink and ESUN. It also disclosed $205 million in funding, without naming a round or lead investor. Cornelis is based in Wayne, Pennsylvania.

MangoBoost implements its BoostX data processing units on AMD FPGAs, allowing networking functions to change as protocols and workloads evolve. The RoCE AI variant supports 400 Gb Ethernet; other variants address storage or custom processing. Its timeline says BoostX supply to global cloud service providers is accelerating in 2026. The company also offers inference software and server systems, while its own programmable DPU data paths provide the networking technology here. MangoBoost is headquartered in Bellevue, Washington, with operations in South Korea.

Xsight Labs supplies both Ethernet switch chips and DPUs for cloud, edge and AI data center networks. Its X2 switch handles 12.8 Tbps through up to 16 ports of 800G. An open instruction set, XISA, lets its data plane be reprogrammed after deployment; Xsight says the chip is ready for Ultra Ethernet. SpaceX has chosen X2 for next generation Starlink satellites, giving the switch a named application beyond data centers.

The E1 DPU has 64 Arm cores and processes traffic at 800 Gbps, handling networking, security and storage work that would otherwise occupy server processors. In its July 2026 release, Xsight said it was shipping both DPUs and switches and described that DPU as the first to ship at the stated speed.

The Tel Aviv, Israel company announced more than $300 million raised at a $2.8 billion valuation, led by Fidelity.

Panmnesia uses CXL switches and controllers to connect processors and accelerators with pooled memory. Its PCIe 6.4-CXL 3.2 Fusion Switch has 256 lanes, joining GPUs, CPUs, memory expanders and accelerators; switches can be chained across racks. The Daejeon, South Korea company said in April 2026 that it supplied pre release silicon and pilot systems to early access partners. An October 2026 update describes work with system manufacturers on appliances built around the switch. That appliance work does not change the stated early access stage of the silicon.

Primemas builds CXL memory controller chips and the cards and systems around them, using its Hublet chiplet based hub. Its PMA14 and PMA16 cards hold 14 or 16 memory modules and up to 4 TB of DRAM per card. In August 2026, it scheduled the cards to ship to Micron in September for testing and qualification in Abaco, a rack system for Pacific Northwest National Laboratory with more than 100 TB of pooled memory. The source states a schedule, rather than confirming delivery. Primemas also introduced SLiM, a shared memory pool without a separate CXL switch. It is headquartered in Santa Clara, California.

UnifabriX’s MAX system pools, shares and maps memory across a rack. The Haifa company announced expanded early access after public demonstrations, with agreements for validation. Its current site offers systems by quotation, but the reviewed sources do not confirm customer delivery.