Delos Data raises over $100M for Nonstop AI inference networking

Delos Data said it closed more than $100 million from Matrix, Playground, Socratic Partners, Capricorn’s Technology Impact Fund, Matter Venture Partners, IAG, and industry investors to launch Delos Nonstop AI™. Alongside the raise, it introduced a Nonstop AI Data Interface and Reference Architecture aimed at lower-latency, more resilient data movement for agentic AI inference across mixed hardware.

HARDWARE desk — same-day primary $100M+ raise plus interconnect product push shows capital shifting from “more GPUs” toward networking silicon for agentic inference economics.

Use of proceeds, as the same newsroom has it: grow software and hardware engineering teams and accelerate product development and sales.

Thesis, still company: for agentic inference, the network — data movement between GPUs, XPUs, accelerators, CPUs, memory, and storage — is the next chokepoint, not power, memory, or compute alone. Inference means running a trained model to answer or act. Agentic here means long-running AI that keeps calling tools and moving data, rather than a one-shot request. An XPU is another AI accelerator besides a GPU.

Demand cite, company-attributed — not a desk forecast: Delos Data cites industry projections of roughly a 300x increase in token demand by 2030, with agentic inference as the bulk of that growth. A token is a chunk of text or data a model processes.

New products, still company: a Nonstop AI Data Interface and a Nonstop AI Reference Architecture. They join previously announced Nonstop AI Clusters and Nonstop AI Server. The company says all three sit on that Reference Architecture so customers can compose clusters from their own mix of hardware, models, and topologies in one domain.

Performance targets, company claim — label as a target, not a bench this desk ran: the newsroom says the Data Interface is aimed at 10x lower latency and 10x higher efficiency versus what endpoints achieve today. The same-day PR Newswire also states company targets of 10x faster performance, 10x stronger resiliency, and 10x higher scale.

Data Interface form factors, still company: one interface in three shapes for different endpoint classes — an I/O chiplet at 30-plus Tbps for GPUs, XPUs, and accelerators; near-packaged optics at 10-plus Tbps that pair the interface with optical engines; and a card at 400-plus Gbps for CPUs, flash, and memory joining the same data domain. Tbps means trillions of bits per second. Gbps means billions of bits per second. Near-packaged optics means putting the optical engine close to the chip.

Availability, company: Nonstop AI Clusters are in production on existing infrastructure; the platform is available now for customers to co-design next-generation agentic AI infrastructure; Nonstop AI Server sampling is targeted for the end of 2026.

Show calendar, company: Delos Data is exhibiting at AI Infra Summit booth 1344, including a Nonstop AI Cluster Live demo of mixed endpoints in one data domain.

Named voices on the company pages: Ed Doe, CEO and co-founder; Dan Daly, CTO and co-founder; and investors including Pat Gelsinger, general partner at Playground. Other investor voices on the newsroom — Stan Reiss at Matrix, Christian Gallagher at Socratic Partners, Dipender Saluja at Capricorn’s Technology Impact Fund, and Wen Hsieh at Matter Venture Partners — stay in Sources. Their quotes are color only and stay in Sources. This is not investment advice.

Plain English for the rest of the card: interconnect = the links that move data between chips. inference = running a trained model to answer or act. agentic = long-running AI that keeps calling tools and moving data. Tbps = trillions of bits per second. Gbps = billions of bits per second. XPU = another AI accelerator besides a GPU. near-packaged optics = putting the optical engine close to the chip.