Chip-to-chip connections: which startup is ahead?

Which chip-to-chip connection startups should we compare?

The chip-to-chip connection race currently has seven serious independent startups, led by Ayar Labs and Lightmatter, although several companies compete at different points inside the system.

We include companies whose main product moves data between chiplets, processors, memory, packages, boards or nearby AI systems. That gives us Ayar Labs, Lightmatter, Kandou AI, Eliyan, Avicena, Xscape Photonics and Hyperlume.

The scope needs some care because "chip-to-chip connections" covers several distances. Eliyan mainly connects dies inside one package. Kandou stretches electrical connections across packages and circuit boards. Ayar and Lightmatter bring optical links close to processors and switches. Avicena and Hyperlume use microLEDs for short optical links, while Xscape supplies multi-wavelength lasers that can power several optical architectures.

These companies can complement each other as well as compete. An accelerator could eventually use Eliyan inside the package, Kandou across the board and Ayar or Lightmatter between trays or racks.

We exclude ordinary networking startups whose main products are switches, Ethernet systems or network software. We also exclude public semiconductor groups such as Nvidia, Broadcom and Marvell, even though they may ultimately ship far more interconnect hardware than the startups covered here.

Celestial AI is the most important exclusion. Marvell completed the acquisition on February 2, 2026 after announcing approximately $3.25 billion in upfront consideration. Marvell's later acquisition accounting recorded $3.5 billion in total purchase consideration. Celestial AI would otherwise belong near the top of this comparison, but it no longer operates as an independent startup.

The seven companies below have raised about $2.59 billion altogether. Ayar Labs and Lightmatter alone account for roughly two-thirds of that capital. Adding Kandou takes the top three to about 86%.

Startup	Main focus	Announced cumulative funding
Ayar Labs	Optical I/O chiplets and external laser systems	$870 million
Lightmatter	Photonic interconnects, optical engines and photonic interposers	$850 million
Kandou AI	High-speed electrical links across chips, boards and memory systems	Approximately $505 million
Eliyan	Electrical die-to-die and chip-to-chip interface technology	$150 million
Avicena	Parallel microLED optical links for processors and memory	$120 million
Xscape Photonics	Multi-wavelength laser sources for optical interconnects	$81 million
Hyperlume	Low-power microLED optical connections	$12.5 million

Is Ayar Labs clearly ahead of Lightmatter today?

Ayar Labs leads Lightmatter in chip-to-chip connections today, but the gap is small enough for one major production contract to reverse the ranking.

The market has a narrow top two, followed by Kandou as the main electrical challenger. Ayar and Lightmatter have raised similar amounts, built broad manufacturing ecosystems and moved beyond laboratory demonstrations.

Ayar's recent developments fit together particularly well. TeraPHY offers up to 8 terabits per second per optical engine. Designs with Alchip and GUC place several engines around accelerators, taking total package connectivity beyond 100 terabits per second. Wiwynn is now working with Ayar on rack-scale systems supporting up to 1,024 accelerators.

Ayar has also kept its product focused. TeraPHY handles optical data movement beside an accelerator or switch. SuperNova provides the light source. UCIe connects the optical engine to the surrounding silicon. Customers can add these pieces without redesigning the whole processor around one proprietary platform.

Lightmatter has the stronger headline specifications. Its Passage portfolio includes 32-terabit and 64-terabit optical engines, 1.6 terabits per second through a single fiber and photonic interposers that can sit beneath large compute packages.

That broader architecture could ultimately support more bandwidth, but it also creates more integration work. Customers must qualify the photonic silicon, electronic interfaces, packaging, fibers, lasers, cooling and repair process together.

Ayar offers the shorter path from optical engine to finished AI system. Lightmatter could overtake it quickly if a named hyperscaler puts Passage into a production cluster.

On an informal leadership index, we would place Ayar at 100 and Lightmatter around 93. Kandou sits closer to 60 because its newest AI product has only recently reached working silicon.

If you want more recent data on this point, please see our latest semiconductor industry report.

Which chip-to-chip startup has the most mature product and strongest customer traction?

Ayar Labs has the most mature new optical product and the strongest visible customer pipeline, while Kandou AI has the deepest history of shipping connectivity silicon.

Ayar describes TeraPHY as production-ready and has tested it under rapid temperature changes while maintaining its target error performance. The engine uses established CMOS and silicon-photonics processes, connects through UCIe and fits packaging methods already used for advanced processors and switches.

Its partnerships cover most of the route from chiplet to rack. Alchip and GUC create accelerator and custom ASIC designs. Wiwynn contributes board, server and rack engineering. Jabil has worked on external laser infrastructure. Nvidia, AMD, Intel, MediaTek and Alchip have invested in Ayar or worked closely with it.

The sequence is more useful than any single announcement. Ayar demonstrated optical links, aligned with UCIe, passed reliability testing, created accelerator reference designs and moved into rack-level integration. It is now expanding manufacturing capacity.

Lightmatter comes next. Passage products are sampling, evaluation systems are available, and GUC plans to bring co-packaged optical products based on the platform to hyperscale customers. GlobalFoundries, ASE, Amkor, Synopsys and Cadence support its manufacturing and design ecosystem.

Lightmatter has not named a cloud company operating a production cluster built around Passage. Evaluation systems and development agreements show serious interest, but not paid adoption at scale.

Kandou says more than 20 million chips using its connectivity technology have shipped. Most came from earlier USB, PCIe, retimer and consumer-connectivity products rather than Tigerwing, so the figure proves manufacturing experience rather than current AI traction.

Eliyan has working advanced-node silicon. Avicena offers a complete evaluation kit. Xscape has launched a commercial laser module. No independent startup has disclosed enough shipment, revenue or field-reliability data to claim mass adoption.

Which chip-to-chip startup has the best performance?

Lightmatter leads optical bandwidth, Kandou AI has the strongest long-distance copper result, and Avicena has reported the lowest transmitter-energy figure.

Lightmatter has sampled a Passage chiplet carrying 1.6 terabits per second through one fiber by placing 16 wavelengths on that fiber. Its L200 family reaches 32 or 64 terabits per second per engine, while larger configurations can provide more than 200 terabits per second around one processor package.

Ayar's individual TeraPHY engine reaches 8 terabits per second. Ayar scales by placing several engines beside one accelerator, creating package designs above 100 terabits per second.

Kandou recently demonstrated a 260-gigabit-per-second electrical connection over more than 200 millimeters of ordinary printed circuit board. Reported energy stayed below one picojoule per bit, with a bit error rate below 10^-19.

Eliyan has delivered 64 gigabits per second per bump on 3-nanometer silicon. Its technology targets shorter connections inside packages, where hundreds of closely spaced bumps create several terabits per second of total bandwidth.

Avicena uses hundreds of relatively slow microLED channels rather than a few extremely fast serial lanes. The company has demonstrated four gigabits per second per lane and transmitter energy around 80 femtojoules per bit under specified test conditions.

The figures measure different parts of the connection, so direct comparison has limits. A transmitter-only energy number will naturally look better than a figure covering drivers, receivers, lasers, fibers and control circuitry.

Startup	Strongest disclosed result	Where it looks strongest
Lightmatter	1.6 Tbps per fiber and up to 64 Tbps per optical engine	Optical bandwidth and fiber density
Ayar Labs	8 Tbps per engine and more than 100 Tbps per accelerator design	Modular package-level scaling
Kandou AI	260 Gbps over more than 200 mm at below 1 pJ/bit	High-speed copper across circuit boards
Eliyan	64 Gbps per bump on 3 nm silicon	Dense electrical links inside packages
Avicena	4 Gbps per lane and about 80 fJ/bit at the transmitter	Low-power parallel optical links
Xscape Photonics	Eight wavelengths and more than 1 watt of optical output	Shared laser infrastructure

Can Lightmatter turn its chip-to-chip technology lead into a business lead?

Lightmatter can take the chip-to-chip lead, but it needs a major production customer more than another bandwidth record.

Passage attacks a real physical constraint. Processors traditionally send data through connections placed around their edges. As chips become larger and need more bandwidth, that limited perimeter becomes crowded. Lightmatter can place a photonic layer beneath the processor or stack an electronic interface directly on photonic silicon.

The company has lately focused more on the practical problems around that architecture. It introduced detachable fiber arrays for easier assembly and replacement, a liquid-cooled laser system for dense racks and smaller Passage engines for near-package and on-board uses.

Its partnerships also look increasingly industrial. GUC supports custom silicon, GlobalFoundries provides silicon-photonics manufacturing, and ASE and Amkor provide advanced packaging. Cadence and Synopsys help chip designers integrate the required interfaces.

The missing piece is a named hyperscaler running a large production cluster with Passage. That would count for more than another improvement in peak capacity.

If you want more recent data on this point, please see our latest semiconductor industry report.

Can Ayar Labs manufacture chip-to-chip optical links at scale?

Ayar Labs has the strongest scaling plan among optical chip-to-chip startups, although it has not published the yields or shipment volumes needed to prove mass manufacturing.

TeraPHY uses standard semiconductor manufacturing and packaging flows. The optical engine sits beside a processor or switch and communicates through an electrical UCIe interface. SuperNova keeps the light source outside the expensive compute package.

Separating the laser helps with heat, replacement and product upgrades. Ayar can improve the light source without forcing customers to redesign every processor package.

The supply chain now covers more than fabrication. Alchip and GUC handle accelerator integration. Jabil contributes optical and laser expertise. Detachable fiber connections simplify assembly and servicing. Wiwynn brings the system engineering needed to move from chip-level designs to rack equipment.

Ayar's $500 million financing was directed toward production and test capacity. That is consistent with a company moving from engineering validation toward manufacturing.

The hardest numbers remain private. We do not know Ayar's monthly output, packaging yield, unit price or field-failure rate. For now, Ayar looks production-ready rather than proven at hyperscale.

Could Kandou AI keep copper competitive for chip-to-chip connections?

Kandou AI could delay the move to optical chip-to-chip connections because many customers will keep choosing copper whenever it remains fast enough.

Copper is familiar, inexpensive and easy to manufacture. It uses ordinary circuit boards and avoids lasers, fibers and optical alignment. Its weakness is that electrical losses rise quickly as speed and distance increase.

Kandou uses coordinated signaling across several wires to push that limit further. Tigerwing recently reached working silicon on a 4-nanometer process and transmitted 260 gigabits per second across more than 200 millimeters of circuit board at below one picojoule per bit.

A 200-millimeter connection can link separate processor, memory or switch packages on the same board. Customers could keep components physically separate and easier to manufacture while gaining some of the benefits of a large multi-chip package.

Kandou also has more shipping experience than most optical startups. More than 20 million chips using its earlier connectivity technology have shipped, and parts of its signaling work have influenced industry standards.

Tigerwing still needs a qualified customer deployment. Copper will eventually lose at very long distances and extreme bandwidth, but Kandou can win the large middle zone across boards and nearby packages.

Is Eliyan leading electrical chiplet connections?

Eliyan leads the independent startups building electrical die-to-die connections inside advanced chip packages.

NuLink connects chiplets across silicon bridges, interposers or ordinary organic packaging. Its main appeal is the ability to support different bump sizes and substrates without forcing every customer into the most expensive packaging.

Eliyan says NuLink can deliver high bandwidth and low latency through organic substrates in suitable designs. That could preserve more of the cost advantage that chiplets are supposed to create.

The company has delivered working 3-nanometer silicon at 64 gigabits per second per bump. It has also developed versions for Samsung's 4-nanometer process and longer chip-to-chip distances.

Eliyan has raised $150 million, yet its investors include AMD, Arm, Meta, Coherent, Samsung, Intel Capital and SK hynix. That group covers processors, cloud infrastructure, optics and memory.

The missing evidence is a production processor shipping with NuLink. Until that appears, Eliyan remains the electrical subcategory leader rather than an overall top-three company.

If you want more recent data on this point, please see our latest semiconductor industry report.

Could Avicena's microLED links beat conventional silicon photonics?

Avicena has the most credible alternative to laser-based optical chip-to-chip connections, but its microLED platform remains several steps behind Ayar and Lightmatter commercially.

Conventional silicon-photonics systems concentrate data into a small number of extremely fast lanes. Avicena instead uses hundreds of microLEDs and photodetectors connected through multicore fiber.

Each lane runs at a more modest speed, but the large number of lanes creates high total bandwidth. The current evaluation kit provides a 512-gigabit-per-second link, and Avicena has reported more than two terabits per second per millimeter of bandwidth density.

Slower parallel lanes may need less demanding signal processing than a few ultra-fast serial channels. MicroLED manufacturing could also benefit from equipment and techniques developed for displays.

Avicena has acquired Nanosys assets and is working with TSMC on photodetectors. Its investors include SK hynix, Micron Ventures, Lam Research and Hitachi Ventures.

Customers are only beginning to test the complete system. Avicena still needs to prove long-term LED reliability, fiber alignment, packaging yield and total link power outside the laboratory.

Which chip-to-chip startup uses its funding best?

Eliyan has produced the most validated chip-to-chip technology per dollar raised, while Ayar Labs has built the strongest case for spending at a much larger scale.

Eliyan has raised $150 million and delivered working 3-nanometer silicon, several packaging options and versions covering both die-to-die and longer chip-to-chip connections. That is less than one-fifth of the capital raised by either Ayar or Lightmatter.

Avicena has also used its $120 million efficiently. It developed a new microLED architecture, acquired manufacturing assets, worked with TSMC and released a complete evaluation kit.

Kandou has raised roughly $505 million. That produced several generations of connectivity silicon, more than 20 million shipped units and the new Tigerwing architecture.

Ayar's $870 million supports photonic dies, electronic interfaces, laser systems, packaging methods, test equipment and manufacturing capacity. Lightmatter has spent $850 million on an even broader product range covering several distances and laser infrastructure.

We cannot calculate true capital efficiency without revenue, gross margin and operating-expense data. Based on public engineering output, Eliyan has achieved the most with the least money.

Startup	Funding	What the capital has produced
Eliyan	$150 million	Working advanced-node silicon and several packaging options
Avicena	$120 million	A new microLED architecture and a complete evaluation kit
Kandou AI	Approximately $505 million	More than 20 million historical shipments and working Tigerwing silicon
Ayar Labs	$870 million	Production-ready optical chiplets and a broad system ecosystem
Lightmatter	$850 million	The widest photonic product portfolio and advanced evaluation platforms
Xscape Photonics	$81 million	A commercial eight-wavelength laser module
Hyperlume	$12.5 million	Early microLED interconnect development

Which chip-to-chip startup is moving fastest now?

Ayar Labs has the strongest commercial momentum, while Lightmatter is releasing products faster and Kandou AI is making the sharpest comeback.

Ayar has recently combined fresh production capital, expansion in Taiwan, Nvidia ecosystem access and rack-level development with Wiwynn. The pieces fit: Taiwan puts Ayar closer to advanced suppliers, the financing expands manufacturing, and Wiwynn provides a route into complete systems.

Lightmatter has been more active on the product side. It sampled its high-density Passage chiplet, introduced a smaller optical engine, launched detachable fiber technology and added a liquid-cooled laser system.

Lightmatter has also broadened its strategy. Customers can now enter through near-package and co-packaged products instead of starting with the largest photonic interposer.

Kandou has changed direction more dramatically. New leadership repositioned the company around AI infrastructure, the business raised $225 million, and Tigerwing reached working silicon soon afterward.

Ayar has the most complete momentum because its recent activity stretches from funding to manufacturing and systems. Lightmatter is advancing at least as quickly technically.

If you want more recent data on this point, please see our latest semiconductor industry report.

Could Nvidia, Broadcom or Marvell crush the chip-to-chip startups?

Large semiconductor companies could erase the lead of weaker chip-to-chip startups, but Ayar Labs and Lightmatter have become strategically useful enough to survive or attract a major buyer.

Nvidia is developing photonic switches and complete networking systems. Broadcom already sells high-capacity switching silicon and has several generations of co-packaged optics. Marvell now owns Celestial AI and can combine its Photonic Fabric with switches, custom chips and optical interfaces.

These companies already sell processors or networking chips in huge volumes. Their customers trust their manufacturing, software and support, and they can bundle connectivity with products buyers already need.

Marvell's purchase of Celestial AI shows one likely outcome. A strong interconnect startup may create more value inside a large semiconductor platform than as an independent company.

Ayar and Lightmatter still have room to build independent businesses because hyperscalers designing custom accelerators often want components that work across several processor and switch vendors. Open interfaces such as UCIe also make neutral suppliers easier to adopt.

Reliable optical chiplets require mixed electronic-photonic design, advanced packaging, fiber attachment, testing and years of manufacturing work. Large incumbents can copy the concept faster than they can reproduce the complete production knowledge.

What would prove that a chip-to-chip startup has really won?

A named, high-volume production deployment would settle the chip-to-chip startup race more convincingly than any funding round, partnership or benchmark.

The first missing number is shipment volume. We need to know how many optical engines, electrical interfaces or microLED links leave factories each quarter and how many operate in real AI systems.

The second is paid contract value. A multi-year supply agreement worth several hundred million dollars would outweigh a long list of development partnerships.

Manufacturing economics would also change the ranking. Customers need the full cost of photonic dies, electrical interfaces, lasers, fibers, connectors, packaging, testing, cooling and repair.

Reliability data will become more important as products reach the field. Buyers need predictable laser lifetimes, low failure rates and a practical way to replace damaged components.

Real workload results would be especially persuasive. A production cluster running large models would show whether projected gains in training speed and throughput survive software overhead, congestion and system failures.

Ayar has the strongest indirect evidence. Lightmatter has the best technical case. Kandou has the longest shipping history. The first company to connect those strengths to a named, paid and repeatable deployment will pull clearly ahead.

Which chip-to-chip connection startups are actually ahead?

Ayar Labs is the strongest chip-to-chip connection startup overall, followed closely by Lightmatter and then Kandou AI.

Ayar wins because it combines a mature optical engine, a standard UCIe interface, credible chip-design partners, manufacturing investment and a visible route into rack-scale AI infrastructure.

Lightmatter ranks second with only a small gap. Passage offers higher bandwidth density and a broader range of optical architectures. A large named customer could move Lightmatter into first place.

Kandou ranks third because copper remains valuable wherever it can still meet the required speed and distance. The company has shipped millions of earlier connectivity chips and recently proved Tigerwing in silicon.

Eliyan holds fourth place and leads electrical links inside advanced packages. Avicena ranks fifth because microLEDs could eventually provide a simpler optical architecture. Xscape ranks sixth as an important laser supplier without control of the complete link. Hyperlume remains seventh because it is earlier and less publicly validated.

We place Ayar at 100 and Lightmatter around 93. Kandou sits much further back near 60 because its latest AI technology has not yet shown the same system-level adoption path.

A hyperscaler production deployment would probably put Lightmatter first. A large Tigerwing contract could pull Kandou closer to the top two. A successful processor or memory integration could move Avicena above Eliyan.

For now, Ayar Labs is ahead. Lightmatter remains the most dangerous challenger, while Kandou offers the strongest alternative route by extending copper rather than replacing it immediately.

Rank	Startup	Why it ranks here
1	Ayar Labs	Strongest mix of product maturity, accelerator integration, manufacturing readiness and rack-scale access
2	Lightmatter	Best optical performance and broadest platform, with less public evidence of production adoption
3	Kandou AI	Proven shipping experience and a compelling copper interface that recently reached working silicon
4	Eliyan	Leading independent electrical die-to-die specialist with advanced silicon and exceptional strategic backing
5	Avicena	Most promising alternative optical architecture, but still early in customer qualification
6	Xscape Photonics	Valuable multi-wavelength laser supplier without control of the complete interconnect platform
7	Hyperlume	Relevant microLED technology at a much earlier commercial stage

If you want more recent data on this point, please see our latest semiconductor industry report.