Delos Data builds its inference server around an OAM socket buyers fill themselves

The socket is the part a procurement team can act on. Delos says the Nonstop AI Server takes up to four graphics cards from outside suppliers, attached to the motherboard over OAM, an open-source interconnect built for AI work ^[4]. Servers then link into a cluster over OSFP, an open-source routing technology ^[5]. Both are open standards, and neither came from Delos. A team that wants to keep a second accelerator supplier in the plan can do that without buying a different chassis.

Nine OSFP ports at 1.6 terabits per second on each of four chips works out to 36 ports and 57.6 terabits per second for a single server ^[6][16]. SiliconANGLE reports that this gives the system network capacity comparable to Nvidia's Rubin appliances ^[7]. The article does not name a price, a customer, or the Rubin figure being compared against ^[18].

The new part is the Nonstop AI Data Interface, a chip that connects the server to the network of the cluster it sits in ^[8]. Delos says it provides 10 times lower latency than rival products, and claims it is more efficient and more scalable ^[9]. The second function listed for the chip matters more to the operations team on call: it monitors the host cluster for malfunctions, including graphics card crashes and network link outages, and automates some of the recovery work ^[10].

Dan Daly, the company's co-founder and chief technology officer, said agentic AI is scaling beyond a single device, and that once it does, performance is not settled by the processors alone; it is decided by how the data moves between them. What agentic AI needs, he said, is a data interface that is fast and does not break under load ^[14].

The three ways to buy that chip are not equivalent commitments. A PCIe card drops into servers a data centre operator already owns ^[11]. The chiplet processes more than 30 terabits per second and is integrated directly into the customer's own accelerator ^[12]. There is also a near-packaged optics version, rated at more than 10 terabits per second, which sits on the same substrate as the accelerator and runs slower than the chiplet while being simpler to implement ^[13]. So the chiplet moves roughly three times the traffic of the NPO part ^[17], and taking it means committing at accelerator design time.

Chassis and PCIe card come out with a screwdriver and a maintenance window. The chiplet goes into someone's tape-out and stays there, and Delos is spending this round, raised from Matrix, Playground, Socratic Partners, Capricorn, Matter Venture Partners, IAG and angel investors from the data centre hardware industry ^[2], on hiring more engineers and accelerating sales ^[15]. The 10x latency figure is the company's own, so a buyer's own traffic is what has to confirm it before the fabric choice is locked.