The queue of startups chasing AI’s interconnect bottlenecks just gained another well-funded entrant. Delos Data raised more than $100M to build networking chips and software for data centers, with most of its engineering bench drawn from Intel.
Matrix and Playground led the group. Socratic Partners joined, along with Capricorn’s Technology Impact Fund, Matter Venture Partners and IAG. The money is earmarked for engineering hires and product work.
The company bets that the network, not compute, is what breaks next. Its Nonstop AI architecture is pitched at roughly 100 nanoseconds of latency, where a conventional network interface card runs in microseconds, and is built to move data between unlike accelerators without a translation step.
Idle accelerators, in chief executive Ed Doe’s telling, are the most wasteful line item in any cluster. His description of the goal involves shuttling data among GPUs, dataflow engines and flash over sockets without stacking on latency at every hop.
A development board called Morpheus is shipping now as a PCIe card, letting customers test their own logic against real workload topologies. Delos expects token demand to rise roughly 300x by 2030, most of it from agentic inference.
The bet has plenty of company. Established vendors and a crowd of startups are all trying to loosen Nvidia’s grip on rack-scale interconnect, which is where AI clusters are increasingly won or lost.
