Every accelerator inside an AI data center is rented by the second, which makes waiting the most expensive thing a chip can do. That gap is what Delos Data is selling, and the Intel alumni behind the company collected more than $100M to chase it.
The cheque came from a group that includes Socratic Partners, Matrix Partners, Matter Venture Partners, DYNAMIQ, IAG and Capricorn’s Technology Impact Fund. Gelsinger, a former Intel chief executive, backs the company through Playground Global, where he is now a general partner.
The thesis is that the bottleneck has moved from compute to the wiring between chips. Training clusters were built around Nvidia accelerators and Nvidia networking, but inference work now spreads across GPUs, CPUs, memory, storage and accelerators from AMD, Cerebras and other suppliers. That mix of hardware is exactly where today’s interconnect struggles.
“The most expensive idle asset in a data center is a GPU, CPU, or an accelerator waiting on the network,” chief executive Ed Doe said, arguing the plumbing was never designed for how inference moves information.
Delos plans to spend the money on hardware and software engineers plus product development and sales. It joins optical interconnect players such as Ayar Labs and Celero in betting that mixed-vendor clusters need a new networking layer.
The argument to buyers is blunt: more accelerators deliver little if those chips spend their working hours waiting for information to arrive, and AI infrastructure budgets are already under strain.
