Delos Data raises $100M-plus to keep AI inference running through failures

Why it matters

Delos Data is betting that AI infrastructure value will move toward the links between accelerators. The financing gives two veteran network builders capital to test that thesis across software, servers and interface silicon, but named production customers and independent benchmarks remain absent.

Delos Data co-founders Ed Doe and Dan Daly are building an AI interconnect spanning cluster software, server architecture and interface silicon.

The Palo Alto, California-based Delos Data says it has raised more than $100 million and launched its Nonstop AI Reference Architecture on September 15th. Delos Data says its Nonstop AI Reference Architecture lets customers connect a mixture of GPUs, CPUs, other accelerators, memory and storage within one data domain while keeping inference workloads running through hardware and link failures.

Agentic inference, in Delos Data's framing, consists of persistent workloads that run across multiple devices instead of handling one request and stopping. The funding announcement cites an industry projection of roughly 300x growth in token demand by 2030, with agentic inference accounting for most of that increase. The announcement does not identify the projection's author.

Backers include Matrix, Playground Global, Socratic Partners, Capricorn's Technology Impact Fund, Matter Venture Partners and IAG, along with unnamed investors from the computing and networking industries. Delos Data plans to use the capital to expand its software and hardware engineering teams and accelerate product development and sales.

Delos Data's press release is the primary source for the financing and product claims. The available sources do not identify the exact financing amount, round structure, valuation, lead investor or total funding to date.

Two networking careers converge on inference

Doe and Daly have spent much of their careers working on systems that move data between processors. Their wager at Delos Data is that the same discipline now needs to be applied to AI inference, where an expensive accelerator can sit idle while it waits for data from elsewhere in a cluster.

"The most expensive idle asset in a data center is a GPU, CPU or an accelerator waiting on the network," Doe said in the September 15th announcement.

Doe held senior networking roles at Intel, including leadership of Barefoot Networks, which Intel acquired in 2019. Barefoot's programmable Tofino Ethernet switch technology became part of Intel's networking portfolio.

An Open Networking Foundation biography of Daly says he joined Intel through its 2011 acquisition of Fulcrum Microsystems. Daly joined Fulcrum in 2001 and served as technical director of software and systems, working on software control planes for high-speed, low-latency switch hardware. The biography says he holds degrees from Caltech and USC.

The founders' networking backgrounds inform Delos Data's thesis that inference infrastructure needs a purpose-built data path. The processors, models and storage systems can vary; Delos Data wants to supply the architecture that connects them and keeps work moving when a component fails.

Software first, silicon next

Delos Data's announced product set includes Nonstop AI Clusters, Nonstop AI Server, Nonstop AI Data Interface and Mosaic software. Nonstop AI Clusters is a resilient cluster architecture for agentic inference. Nonstop AI Server is a disaggregated scale-up server architecture intended to let customers mix processors and switches. Nonstop AI Data Interface is planned silicon for linking compute, memory and storage.

Delos Data says its Data Interface delivers 10x lower latency and 10x higher efficiency. The supplied materials include no independent benchmark, test configuration or named customer result supporting those figures.

A May 27th, 2026, Delos Data release announced the cluster architecture and server design, which Delos Data says are intended to scale across thousands of GPUs, CPUs and accelerators. Delos Data listed reduced cost per token, improved tokens per watt and increased tokens per second as intended benefits. Delos Data material described 1,000-GPU scale-up domains as practical and 10,000 GPUs as potentially possible, but those figures have not been independently validated. In the September financing announcement, Matter Venture Partners founding managing partner Wen Hsieh described Delos Data as having built software, followed by an AI server, and said Delos Data was building interface silicon.

At Computex 2026, The Register reported that Delos Data displayed its Nonstop AI network and allowed attendees to pull links at random while the network automatically reacted and corrected itself. The demonstration showed the failure-response feature in a controlled setting rather than providing an independent performance benchmark. The Register described Delos Data's modular system as an alternative to rack-scale approaches from Nvidia and AMD and noted that AWS has a competing reference direction.

In the financing announcement, Delos Data said an AI Infra Summit demonstration at booth 1344 would show heterogeneous endpoints operating in one data domain and surviving failures without stopping the workload.

A May 27th Delos Data announcement said early-access software deployments were available to select customers and broader availability was planned for Q4 2026. Delos Data did not name the customers or say whether any deployment was carrying production traffic.

Delos Data is also recruiting for physical design. A remote, full-time physical design engineer listing calls for an early member of the physical design team to own implementation work from floorplanning through signoff, including timing closure, routing and physical verification.

Delos Data has not disclosed revenue, pricing, customer count, usage totals or headcount in the materials reviewed.

Inference infrastructure draws bigger checks

d-Matrix said it raised $275 million at a $2 billion valuation on November 12th, 2025, for an inference platform combining accelerators, networking hardware and software. Tensormesh said it raised $20 million on May 27th, 2026, for software that reuses cached model state to reduce repeated computation.

The competitive field extends across several layers. Nexthop AI markets Ethernet switches for large AI clusters. Gruve raised $50 million in February 2026 to expand distributed inference capacity, while AMD acquired inference startup MK1 in November 2025.

These rivals address different sources of waste. d-Matrix is building alternative compute, Tensormesh targets duplicated work, and Delos Data is focused on movement between components. Buyers have already committed enormous sums to accelerators, creating demand for infrastructure that helps those processors spend less time waiting.

Delos Data's hardware-neutral pitch could help customers combine processors from several vendors or avoid designing an entire cluster around one proprietary interconnect. It also gives Delos Data a difficult integration burden. Supporting multiple accelerators, switches, physical links and failure modes means proving the architecture under combinations Delos Data does not fully control.

The $100 million-plus financing gives Doe and Daly room to attempt that proof in hardware. Delos Data has yet to disclose named production deployments or reproducible measurements showing how Nonstop AI performs under customer inference workloads. Investors are funding the founders' contention that data movement has become expensive enough to support Delos Data's expansion across software, servers and silicon.