Delos Data, co-founded by Intel veterans, secured over $100 million from investors including Matrix and Playground, to advance its Nonstop AI interconnect technology that targets data movement bottlenecks in heterogeneous agentic inference clusters.
Delos Data, a Palo Alto-based chip and systems startup founded in 2025 by Intel veterans Ed Doe (CEO and co-founder) and Dan Daly (CTO and co-founder), announced that it has raised over $100 million. The capital comes from Matrix Partners, Playground Global (with participation from former Intel CEO Pat Gelsinger as a partner), Socratic Partners, Capricorn’s Technology Impact Fund, Matter Venture Partners, IAG, DYNAMIQ, and industry luminaries spanning compute, hyperscale systems, networking silicon, infrastructure software, and optical connectivity. Proceeds will expand software and hardware engineering teams and accelerate product development and sales.
The company positions itself at the emerging bottleneck in agentic AI inference: data movement. As AI shifts from training centric GPU clusters (often tied to proprietary interconnects) toward always-on agentic workloads that span heterogeneous mixes of GPUs, XPUs, accelerators, CPUs, memory, and storage, traditional interconnects borrowed from HPC, cloud, or training leave expensive silicon idle while waiting for data. Industry projections cited by the company point to roughly a 300x increase in token demand by 2030, driven largely by agentic inference. Delos argues that performance and economics are now determined less by raw compute and more by low latency, high bandwidth, resilient data paths that keep infrastructure productive through routine failures.
What is Delos Data’s technology?
Delos Nonstop AI is a unified system architecture built around a “workload first” redesign of the data movement domain. It aims for a flatter, faster, stronger interconnect that unifies compute, memory, and storage into one resilient data domain, supporting “Mixture of X Infrastructure” (MoXI), heterogeneous hardware, models, switches, links, and topologies serving a single persistent workload.
Key elements include:
- Nonstop AI Reference Architecture: A system level blueprint emphasizing quicker data movement, scale to hundreds of thousands of devices, and higher resiliency against endpoint, network, software, and link failures. It introduces a data interface between every endpoint so heterogeneous resources can be composed to the workload to deliver continuous “Nonstop Tokens.”
- Nonstop AI Data Interface: Delivered in three form factors, I/O chiplet (30+ Tbps, multi protocol, for GPUs/XPUs/accelerators), near packaged optics (10+ Tbps, multi protocol, integrating with optical engines), and card (400+ Gbps, multi protocol, for CPUs, flash, and memory). It targets 10x lower latency, 10x higher efficiency/resiliency, and 10x higher scale versus current endpoint capabilities. Failure detection and recovery occur in hardware at the interface. Latency is described in the ~100-nanosecond range versus microseconds for typical NICs.
- Nonstop AI Clusters: Resilient cluster architecture for agentic inference at scale; already in production on existing infrastructure and includes a platform for designing/simulating topologies (PCIe validation card, pre silicon simulation/emulation, and cluster level performance modeling). Supports co-design of infrastructure with workloads.
- Nonstop AI Server: Disaggregated scale-up server architecture (Asterion, shown at Computex); sampling to customers expected at the end of 2026.
- Supporting software (Mosaic, launched earlier and in production) and development tools (Morpheus PCIe card for IP validation against real workloads across topologies, available now).
The approach supports multiple protocols and topologies (including NVLink, UALink, Ethernet, InfiniBand) for scale-up, scale-out, and scale-across, with copper, pluggable optics, and co-packaged optics. It aims to abstract integration complexity so customers can mix best of breed components without multi year co-engineering between vendors.
Recommended: Locus Robotics Raises $41.6 Million In Series G Funding
Who are Delos Data’s leaders?
Doe and Daly bring deep data center silicon and systems experience (including Intel, Broadcom, Fulcrum Microsystems, Barefoot Networks, and involvement with efforts around Ultra Ethernet, UALink, ESUN, and Open MRC). Investors highlight the team’s two decade track record of shipping data center silicon and systems at scale. The company describes itself as a mix of software, systems, and silicon experts focused on responsive intelligence at scale. Headquarters is in Palo Alto, California.
The funding coincides with the formal launch of the expanded Nonstop AI portfolio at the AI Infra Summit. It addresses a clear industry shift: early AI data centers were often homogeneous Nvidia GPU clusters with proprietary networking; the agentic era features greater diversity (AMD, Cerebras, multiple Nvidia chip types, custom accelerators) and persistent, failure prone workloads where idle silicon represents major wasted capital and energy. Gelsinger and other investors emphasize that effective communication between devices is now the binding constraint, “the network is the AI.” By optimizing the data path as the core of the system rather than bolting interconnect onto existing designs, Delos seeks to raise GPU/accelerator utilization, lower cost per token, improve tokens per watt, and increase tokens (and revenue) per second.
The capital gives the company runway to scale engineering, move from software and reference architectures already shipping into broader silicon and server sampling, and pursue customers across frontier labs, hyperscalers, neoclouds, and sovereign AI programs. The business model emphasizes flexibility and co-design rather than lock-in to a single accelerator vendor.
Please email us your feedback and news tips at hello(at)techcompanynews.com
