Delos Data is expanding its Nonstop AI architecture with a new Data Interface and Reference Architecture designed to connect heterogeneous compute, accelerator, memory and storage resources into large-scale AI infrastructure. The technical announcement follows the company's disclosure that it has raised more than $100 million to develop its approach to AI cluster infrastructure. Delos is targeting agentic AI workloads that run continuously and require sustained data movement across increasingly heterogeneous systems rather than infrastructure built around a single accelerator, interconnect or topology.
At the center of the architecture is the Delos Nonstop AI Data Interface, positioned between endpoints and the network. Delos plans three implementations: a 30+ Tbps multi-protocol I/O chiplet for GPUs, XPUs and accelerators; a 10+ Tbps near-packaged-optics implementation combining the interface with optical engines; and a 400+ Gbps card for CPUs, flash and memory endpoints. The company says the interface is designed to provide high-bandwidth, low-latency connectivity and to detect failures on the network-facing side and manage recovery in hardware. Dan Daly, CTO and co-founder, said: "These engines spend most of their time waiting to be fed. When you're stitching together endpoints from different vendors at this scale, you can't engineer failure away, you must build for it from the beginning."
The accompanying Nonstop AI Reference Architecture extends this approach to clusters containing potentially hundreds of thousands of devices. Delos describes modern AI infrastructure as a "Mixture of X Infrastructure," or MoXI, combining different hardware, models, switches, links and network topologies around a common workload. The company is also adding design and simulation capabilities, including a PCIe validation card, pre-silicon interface logic and cluster-level performance modeling for evaluating topology, scale, transport behavior and failure modes before infrastructure is deployed.
What remains unknown is whether Delos can meet its stated performance targets in customer deployments. The 30+ Tbps and 10+ Tbps figures are company targets rather than independently demonstrated comparative results, and the company's stated goals of 10x performance, resiliency and scale should be treated as design targets until supported by benchmark data. Delos says its current Nonstop AI Clusters are already operating with existing infrastructure, while the Nonstop AI Server is scheduled to begin customer sampling by the end of 2026.