AI chip startups face massive complexity beyond chip design to build scalable rack architectures that rival Nvidia and AMD. Delos Data introduces a modular platform focused on simplified networking and flexible deployment to fast-track rack-scale systems at hyperscale density.
- Modular design supports thousands of accelerators via flexible switching fabrics
- OSFP ports enable use of standard cables and pluggable optics for network scaling
- Integrated software platform manages resilient, dynamic interconnect topologies
Infrastructure signal
Delos Data is addressing the critical infrastructure challenge of interconnect at rack scale, where traditional GPU server designs integrate custom backplanes and cabling to handle network throughput, power delivery, and cooling. The startup’s chassis looks more like a switch, with 36 OSFP ports that support high-bandwidth accelerator modules through open accelerator module (OAM) sockets, achieving up to 3.6 TB/s per chip interconnect bandwidth. This modular approach allows AI startups to build next-generation rack-scale systems without engineering custom mechanical or thermal solutions from scratch.
This platform supports multi-terabit fabric connectivity using standard packet switches from vendors like Broadcom and Marvell, and can scale from hundreds to thousands of accelerators. By leveraging potential architectures like 2D/3D toruses or optical circuit switches, it opens pathways to extend compute domains beyond single-layer fabrics while maintaining compatibility with off-the-shelf components. However, power demands remain a consideration, especially for systems using pluggable optics at scale, an area Delos is exploring via near-package or co-packaged optics innovations.
Developer impact
For AI system developers, Delos Data’s solution simplifies the traditionally challenging deployment and integration processes tied to rack-scale AI platforms. The modular nature of the hardware means developers can scale their domains up or down with ease, tailoring interconnect fabric sizes based on performance and deployment requirements without redesigning hardware. The use of OSFP ports and compatibility with various communication protocols like PCIe or Ethernet further enhances integration flexibility in diverse developer workflows.
Complementing the hardware, Delos provides a software orchestration layer named the Nonstop AI network, designed to automate configuration and monitor fabric health dynamically. This platform enables real-time traffic rerouting in the event of link failures, improving reliability and reducing downtime risks that can interrupt AI training workloads. Thus, developers benefit not only from hardware scalability but also from enhanced fabric resilience, observability, and simplified network management.
What teams should watch
Infrastructure and platform engineering teams should closely evaluate how modular rack designs leveraging OSFP ports and OAM sockets might reduce time-to-market for new AI accelerator deployments while controlling total cost of ownership. Given the platform’s compatibility with standard networking components and its potential to scale from single nodes to thousands of accelerators, teams focused on hyperscale compute will find new flexibility in design and deployment strategy.
Additionally, teams responsible for observability and platform reliability should monitor advances in Delos’ orchestration software that supports automated topology management and dynamic fault tolerance. As AI workloads continue to demand larger fabrics with complex topologies, integrated software solutions that reduce manual intervention and improve uptime will become increasingly critical. Finally, power management and optics choices remain a key consideration, as adopting novel co-packaged optics or MPO connectors could influence future deployments for power-sensitive environments.