A group of prominent investors including Socratic Partners and IAG is betting that specialized networking ICs are the key to unlocking future AI scaling. This $100 million injection into Delos Data, led by former Intel engineers, marks a fundamental shift in how the industry perceives hardware limits. As agentic inference grows, raw compute is no longer the primary performance driver. The focus has moved to the networking fabric connecting GPUs and memory. By expanding engineering teams and accelerating development, Delos is positioning itself to solve the persistent bottlenecks plaguing data centers. This funding highlights the need for a new class of interconnects that can manage the continuous workloads of autonomous agents without the delays of older, request-based systems. By targeting the architecture of the interface, these innovators are attempting to prevent the massive inefficiencies that arise when high-end processors are forced to wait on slow data delivery.
Data Centers: Redefining the Connectivity Landscape
The Transition from Request to Persistence
The shift toward agentic inference represents a fundamental change in data center functions, moving away from simple request-and-response mechanisms toward persistent workloads. As these AI agents become more sophisticated, they rely on a fragmented landscape of distributed hardware resources that must remain in sync for extended periods. Forecasts indicate that token demand is set to increase 300-fold from 2026 to 2030, a surge driven primarily by complex agentic behaviors rather than traditional search interfaces. Delos Data argues that existing networks, built on legacies like TCP/IP, cannot handle the massive throughput and continuous synchronization required without significant overhead. This mismatch creates a scenario where the network itself becomes a ceiling for performance, preventing advanced models from reaching their full potential because the communication fabric is optimized for a different era. Solving this requires a departure from legacy protocols toward a more fluid data exchange.
Eliminating the Silent Cost of Idle Silicon
One of the most pressing issues in modern data centers is the phenomenon of idle assets, where high-cost GPUs sit stagnant while waiting for data. When communication bottlenecks occur, the total cost of ownership for AI infrastructure skyrockets because the hardware is not utilized to its capacity. Delos is addressing this by focusing on specialized networking ICs that specifically target the data center interconnect. By optimizing how data moves between disparate endpoints like flash storage and processing units, the startup aims to ensure no part of the compute stack is left waiting. This efficiency is critical for maintaining the economic viability of large-scale AI operations. Without a more streamlined way to manage these transfers, organizations risk over-investing in raw compute power they cannot harness, essentially paying for performance trapped behind the limitations of sluggish interface cards. By making data movement more transparent, the startup intends to reclaim lost compute time.
Technical Progress: Innovations for Agentic Workloads
Low Latency through Nonstop AI Design
To bypass the congestion of traditional infrastructure, Delos introduced the “Nonstop AI” architecture, which prioritizes speed at the hardware level. While typical cards often operate with microsecond-level latency, the Delos solution targets a response time of approximately 100 nanoseconds. This improvement is achieved by allowing data to move between different endpoints without high-latency intermediate translation steps that bog down systems. For instance, data can flow directly from a GPU to a dataflow engine or from a CPU to flash storage while maintaining consistent semantics. This direct-bridge capability is essential for agentic tasks requiring rapid-fire updates across multiple memory nodes. By reducing the time it takes for a signal to travel and be understood, the architecture minimizes the processing delays that have historically hindered the development of real-time autonomous agents. This technical leap allows for a much tighter integration between storage and compute than previously possible.
Establishing a Standard for Resilient Intelligence
The transition to specialized networking began with the Morpheus PCIe-based development card, which allowed users to validate their intellectual property against complex network topologies. This move toward a co-design model ensured that infrastructure was built to match the unique requirements of AI workloads, rather than forcing software to adapt to rigid hardware. By integrating end-to-end resiliency directly into the silicon, the design relieved engineers of the burden of managing I/O reliability, which previously stood as a major technical hurdle. This strategic shift helped decouple software innovation from physical hardware limitations, fostering a flexible environment. As a result, architects found they could maintain stable data transfers through routine hardware failures without sacrificing overall throughput. This holistic approach established a standard for data centers where specialized interconnects served as the essential backbone for sustaining the next generation of intelligence.