Delos Data raises over $100M, unveils AI inference networking chip

The funding was provided by Matrix, Playground, Socratic Partners, Capricorn, Matter Venture Partners and IAG, Delos said in its announcement. Angel investors from the data center hardware industry also participated.

The raise comes three months after Delos debuted its first product, the Delos Nonstop AI Server. The appliance is optimized for inference and can be equipped with up to four graphics cards from external suppliers. The accelerators attach to the system's motherboard through an open-source interconnect called OAM that is specifically optimized for AI use cases.

Customers can link multiple Nonstop AI Servers into a cluster. The machines exchange data through OSFP, an open-source routing technology. Each of the server's four chips has nine OSFP ports with 1.6 terabits per second of bandwidth apiece. According to SiliconANGLE, that gives the system network capacity comparable to Nvidia Corp.'s Rubin appliances.

Alongside the funding announcement, Delos introduced the Delos Nonstop AI Data Interface, a chip designed to work with the Nonstop AI Server. It allows the server to connect to the network of the inference cluster in which it is deployed.

The company said the new networking chip provides 10 times lower latency than rival products and is more efficient and scalable. In addition to powering network connections, the chip monitors the host cluster for malfunctions, detecting issues such as graphics card crashes and network link outages. It also automates some of the work involved in recovering from such failures.

Delos sells the technology in three forms. The first is a PCIe card that data center operators can attach to their servers. The other two implementations let customers integrate the technology into their systems at the component level. Delos offers a Nonstop AI Data Interface chiplet that can process more than 30 terabits of traffic per second, which customers can integrate directly into their AI accelerators. There is also a near-packaged optics, or NPO, version with capacity of more than 10 terabits per second, designed to be placed on the same substrate as the accelerator that uses it. That arrangement provides lower speeds than a chiplet integrated directly into the accelerator but can be simpler to implement.

“Agentic AI is scaling beyond a single device, and when that happens performance isn’t solely determined by the processors, it’s decided by how the data moves between them,” said Delos co-founder and Chief Technology Officer Dan Daly. “Agentic AI needs a data interface that’s fast and doesn’t break under load.”

Delos said it will use the newly closed funding round to hire more engineers and accelerate its sales efforts.

Editor's Summary

Delos Data raised more than $100 million from Matrix, Playground, Socratic Partners, Capricorn, Matter Venture Partners and IAG and launched the Nonstop AI Data Interface networking chip. The company says the chip offers lower latency than rivals and supports recovery from cluster failures. The funding will be used to add engineers and expand sales.