Ayar Labs

Walk into any conversation about artificial intelligence and the talk turns quickly to compute - how many GPUs, how many FLOPs, how large the model. Ayar Labs is a company built on a quieter question: once you have all those processors, how do they actually talk to each other? Inside today's AI systems, that conversation still happens over copper wires, and copper is running out of room. It can only carry so much data, over so short a distance, before the power and heat costs become punishing.

Ayar Labs' answer is to swap the copper for light. The company, headquartered in San Jose, California, designs "optical I/O" - input/output technology that sends data between chips as pulses of laser light traveling through optical fiber rather than as electrical signals down metal traces. The pitch is simple to state and hard to build: more bandwidth, less power, longer reach, so that AI clusters can grow from a single server rack into unified systems spanning thousands of accelerators.

"Optical connectivity will be important to scale accelerated computing clusters to meet the demands of AI and HPC workloads." - Bill Dally, Chief Scientist, NVIDIA

The idea is not new - researchers have chased photonic interconnects for decades - but Ayar Labs' particular bet is that optics belongs not in a separate box at the edge of the network, but co-packaged right next to the processor itself. That's the difference between shouting across a room and whispering in someone's ear. The closer the optics sit to the compute, the less energy each bit costs to move.

Modern AI training splits a single model across thousands of processors that must constantly exchange data. When those processors sit centimeters apart on a board, copper handles it. When they sit in different parts of a rack - or in different racks entirely - copper's limits become the ceiling on how large a cluster can grow. Signals degrade, power climbs, and the physical distance a high-speed electrical link can span shrinks as data rates rise.

Ayar Labs frames its mission around a phrase that appears on its own front door: "AI scale-up beyond the rack." By replacing the electrical link with an optical one, the company aims to let system builders connect processors across much greater distances without the bandwidth or energy penalty - stitching many racks into what behaves like one giant machine.

Ayar Labs sells to the companies building the physical layer of the AI boom: chipmakers designing next-generation processors, hyperscale cloud operators, high-performance-computing labs, and defense and aerospace programs. Its relationships read like a roll-call of the semiconductor industry - NVIDIA, AMD, Intel, Hewlett Packard Enterprise, GlobalFoundries, MediaTek, Wiwynn, Jabil and Lockheed Martin among them. The company reports roughly $91.6 million in revenue and employs about 150 people.

Ayar Labs is not alone in chasing photonic interconnects. Lightmatter builds both a photonic interconnect, Passage, and a full AI accelerator; Celestial AI markets a "Photonic Fabric"; and giants like Broadcom and NVIDIA have their own co-packaged optics efforts. What distinguishes Ayar Labs is a combination of approach and posture.

First, standards. In 2025 the company unveiled what it described as the world's first optical chiplet compliant with UCIe - the Universal Chiplet Interconnect Express standard - meaning its optical engine is designed to interoperate with chiplets from other vendors rather than lock customers into a proprietary stack. In an industry moving toward mix-and-match chiplet designs, betting on compatibility rather than a walled garden is a deliberate strategy.

Ayar Labs chose the plumbing, not the shiny part. Optical I/O is invisible when it works - which is exactly why it became a moat.

Second, provenance and manufacturability. The company traces its technology to the DARPA-funded POEM project (Photonically Optimized Embedded Microprocessors), a collaboration across MIT, UC Berkeley and CU Boulder. Its founders spent years proving you could embed optical devices inside standard silicon chip processes - the kind of physics result that is worthless until someone makes it manufacturable at volume. That patient translation from research bench to foundry is the company's core expertise, and it works with partners like GlobalFoundries to actually produce the parts.

Ayar Labs is a business-to-business deep-tech hardware company. It designs and sells optical I/O chiplets and light sources to processor makers, hyperscalers, system integrators and defense customers, and it reaches volume through a foundry-and-packaging ecosystem rather than owning a fab itself. Revenue comes from product sales, development engagements and strategic partnerships - not a software subscription.

Its place in the market is best understood as infrastructure beneath the AI infrastructure. Ayar Labs does not make the GPU, the model or the cloud. It makes the connective tissue that lets those pieces scale - the part that becomes visible only when it fails to keep up. As AI clusters strain against the physical limits of copper, that connective tissue has become one of the most valuable unsolved problems in computing, which is why rivals as fierce as NVIDIA and AMD both wrote checks into the same funding round.

Lightmatter, Celestial AI, Broadcom, Marvell and NVIDIA's in-house optics efforts all circle the same opportunity from different angles - some building full accelerators, some selling fabrics, some keeping optics proprietary. Ayar Labs' wager is that a standards-based, co-packaged, foundry-manufacturable optical chiplet is the version the broadest set of customers can actually adopt.