Delos Data has raised more than $100 million to build chips, software and systems aimed at a growing constraint inside AI infrastructure: moving data between increasingly diverse computing components. The Palo Alto, California-based company plans to expand its hardware and software engineering teams while accelerating product development and sales. Investors include Matrix, Playground, Socratic Partners, Capricorn’s Technology Impact Fund, Matter Venture Partners and IAG, alongside investors with backgrounds across compute, networking silicon and optical connectivity. The financing places another major pool of venture capital behind the idea that AI performance increasingly depends on what happens between processors, not only inside them.
Delos Targets the Network Behind AI Compute
The architecture of AI infrastructure has started to become more heterogeneous as operators combine GPUs, CPUs, accelerators, memory and storage across increasingly large systems. Delos Data is building its Nonstop AI portfolio around that shift, with products designed to improve how data moves among different endpoints. Its approach includes software, servers, clusters and interface silicon rather than treating networking as a standalone layer. That strategy targets a practical problem for infrastructure buyers: expensive compute delivers less economic value when processors spend time waiting for data.
Delos Data Chief Technology Officer and Co-Founder Dan Daly described the uncertainty around future AI architectures in an interview. “We don’t know what the next infrastructure architecture is going to be for agentic (AI),” Daly said. “But we do know that we can provide the quickest, fastest way to move that data around.” The statement captures the company’s bet that infrastructure diversity will increase the importance of flexible data movement.
AI Infrastructure Moves Beyond Homogeneous Clusters
Early AI infrastructure deployments frequently centered on large GPU clusters connected through tightly integrated networking systems. Inference is now expanding the range of hardware that operators may need to connect, particularly as AI applications become more persistent and distributed across different resources. Delos Data argues that this creates a networking problem that cannot be addressed simply by adding more compute. Its architecture instead seeks to make communication between heterogeneous endpoints a core part of system design.
However, Delos is entering a market where established semiconductor and networking suppliers already invest heavily in high-performance interconnects. The company’s opportunity therefore depends on whether infrastructure operators see enough value in a networking layer designed to accommodate multiple processors, links and topologies. For customers, that question goes beyond raw bandwidth because infrastructure utilization can affect the economics of an entire AI cluster. Networking performance may increasingly influence how much useful output buyers can extract from the accelerators they already own.
New Data Interface Extends Delos Nonstop AI
Alongside the financing, Delos Data introduced its Nonstop AI Data Interface and Nonstop AI Reference Architecture. The company says the interface sits between endpoints and the network, allowing different compute, acceleration, memory and storage resources to operate within a common data domain. Delos offers the interface across three planned form factors covering I/O chiplets, near-packaged optics and cards. The company says these designs target different bandwidth requirements across accelerators, processors, memory and storage endpoints.
Delos lists its I/O chiplet at more than 30 Tbps and its near-packaged optics implementation at more than 10 Tbps, while its card targets more than 400 Gbps. Those specifications are company-provided figures and represent the performance targets of the respective products. Delos also says its Data Interface can detect failures and manage recovery in hardware when problems occur elsewhere in the system. That design reflects the company’s broader focus on keeping AI workloads operating when links, accelerators or software components encounter disruptions.
The company claims its architecture targets tenfold improvements in performance, resiliency and scale compared with what endpoints can achieve today. Those targets remain Delos Data’s own performance claims and will ultimately depend on workloads, deployment conditions and customer configurations. Still, the design highlights an important shift in AI infrastructure development toward system-level optimization rather than processor performance alone. Infrastructure buyers increasingly need to evaluate how compute, memory, storage and networking interact under production workloads.
Funding Backs a Broader AI Infrastructure Portfolio
The financing will support engineering expansion as Delos develops the wider Nonstop AI portfolio. The company has already introduced Nonstop AI Clusters and a Nonstop AI Server, while the newly announced Reference Architecture and Data Interface extend the stack toward networking silicon and system design. Delos says its cluster platform lets infrastructure teams model different network topologies and failure modes before committing to deployment. That capability could become relevant as operators weigh increasingly expensive infrastructure configurations.
Delos says its Nonstop AI Clusters are already in production using existing infrastructure, while its platform is available for customers designing next-generation systems. The Nonstop AI Server is expected to sample to customers by the end of 2026. These milestones give the company several layers through which it can engage infrastructure operators, from software and simulation to systems and silicon. The commercial challenge will be translating that broad architecture into measurable improvements across real-world AI deployments.
Idle Compute Sharpens the Economics of Networking
Pat Gelsinger, former Intel CEO and now a general partner at Playground Global, highlighted the financial implications of processors waiting for data. “I can have all the computers in the world, and if I can’t get these devices communicating effectively with one another, I’ve got lots of hot hardware, burning watts and sitting here twiddling their thumbs,” Gelsinger said. His comment points to a basic infrastructure problem: capital-intensive accelerators cannot generate useful work while stalled by other parts of the system. For AI infrastructure owners, utilization can therefore matter almost as much as the headline performance of individual chips.
Meanwhile, Delos CEO and Co-Founder Ed Doe argues that current networking architectures were not designed around persistent agentic workloads. “The most expensive idle asset in a data center is a GPU, CPU or an accelerator waiting on the network. Inference workloads move data in a way that today’s interconnect was never designed to serve,” Doe said. “This funding positions the company to expand our team and offerings to solve the data transfer problem holistically by optimizing the network as the core of the AI system, thereby maximizing infrastructure ROI and ensuring that AI can scale to keep pace with massive growth in AI Inference.” The company is effectively positioning networking efficiency as a way to extract more value from installed AI compute.
Delos Makes Networking Part of the AI Computer
Delos Data’s strategy reflects a broader infrastructure question emerging as AI systems become more complex. Buying faster accelerators does not automatically guarantee proportional application performance if data cannot reach those accelerators quickly enough. Memory, storage, interconnects, topology and failure recovery can all influence how efficiently a cluster performs. That makes the boundaries between compute and networking increasingly important for infrastructure design.
For AI capacity buyers, the implications could extend into procurement decisions. A cluster with powerful processors but weak data movement can leave expensive hardware underused, while an architecture designed around workload communication may improve the amount of productive compute available from the same installed base. Delos Data is betting that operators will increasingly evaluate infrastructure at that system level. Its more than $100 million financing gives the company additional capital to test that thesis as AI inference expands.
The competitive question now shifts from how many accelerators an operator can deploy to how effectively those accelerators can work together. Delos does not need networking to replace compute as the defining component of AI infrastructure for its strategy to matter. It needs data movement to become important enough that operators allocate more capital and engineering attention to the connections between compute resources. As AI infrastructure becomes more heterogeneous, that connection layer could become an increasingly consequential part of the economics of delivering AI capacity.
