Ayar Labs has raised another $150 million to prepare its optical chip technology for large-scale manufacturing, bringing the new capital it has secured this year to $650 million. The investment addresses a growing problem in artificial intelligence hardware: moving data between processors without spending too much power doing it.
The San Jose company announced the financing on September 10. It said the money would fund product development, validation, manufacturing preparations and a new engineering center in Bengaluru, India.
“Copper interconnect is becoming the limiting factor for AI scale-up,” chief executive and co-founder Mark Wade said in the company’s announcement. Interconnects are the links that carry information between chips.
Ayar separately disclosed that data-center equipment maker Wiwynn had invested earlier this year. Its strategic backers also include AMD, Intel, MediaTek, Nvidia and Alchip. The announcement did not identify the size of Wiwynn’s investment.
Using light to connect processors
Ayar develops co-packaged optics, which places optical communications technology close to processing chips within a shared package. The idea is to carry data using light through fiber, reducing the constraints associated with electrical connections as more processors work together.
Its product system combines a TeraPHY optical engine with an external laser light source called SuperNova. Keeping the light source separate is intended to make cooling, reliability and servicing easier. The company says its optical engines can fit into existing chip-package designs.
The commercial goal is to let expensive AI processors exchange information efficiently across multiple equipment racks. A faster processor offers less benefit if the surrounding system cannot feed it data quickly enough.
Ayar advertises bandwidth and power-efficiency advantages over conventional connections. Those are supplier claims about its technology, not independently verified savings for every customer’s complete data center.
The demand reaches beyond the companies designing processors. We recently reported that IQE’s revenue rose as AI demand lifted semiconductor materials. The businesses occupy different parts of the supply chain, but both illustrate spending on the equipment and materials needed around computing chips.
From a chip design to an operating rack
Ayar and Wiwynn announced a rack-system partnership in March. Wiwynn contributes system design and manufacturing, while Ayar supplies the optical technology.
Their proposed architecture is designed to connect 1,024 AI accelerators and beyond. An accelerator is a processor built for particular computing tasks, such as training or running AI models. That design target describes the system the partners are developing, rather than a confirmed installation of that size.
The companies identified fiber routing, cooling, chip integration and manufacturability as deployment issues. Equipment also needs to be serviceable when a part fails. These requirements help explain why a working optical component does not immediately become a product that a cloud operator can install across its facilities.
The partnership includes a liquid-cooled design. Our earlier coverage of Singapore’s benchmark for liquid-cooled AI data centers explains another part of the same operational problem: removing heat as computing power becomes more concentrated.
Production plans extend into 2027
The new financing extends the $500 million Series E round announced in March. That earlier round valued Ayar at $3.75 billion and included money for production and testing capacity.
Reuters reported a separate $225 million purchase of shares from early employees and investors, valuing the company above $5 billion. A purchase of existing shares pays the sellers; it should not be added to the $650 million of new capital raised by Ayar this year.
Wade told Reuters that Ayar needs its products qualified for volume production by the end of 2027 to fit customers’ plans to increase production in 2028 and 2029. Investors are financing that preparation period, with commercial deployment still dependent on customers completing their own systems.
