Delos Data Secures Funding for AI Server Hardware
A startup focused on AI inference infrastructure has secured over $100 million to build systems that prioritize data movement speed over raw processing power.
Delos Data has raised more than $100 million in new funding to accelerate the development of hardware specifically designed for AI inference. The investment round was led by Matrix, Playground, Socratic Partners, Capricorn, Matter Venture Partners, and IAG, with additional support from angel investors in the data center industry.
The company argues that as AI systems grow, performance is increasingly determined by how quickly data moves between components rather than the power of the processors alone. Delos aims to address this bottleneck with a new server appliance and a dedicated networking chip that claims to significantly reduce latency compared to existing market solutions.
Server Design Focuses on Open Standards
The core product is the Nonstop AI Server, a computing unit optimized for inference workloads. It allows customers to install up to four graphics cards from external suppliers, attaching them via an open-source interconnect called OAM. This approach avoids vendor lock-in, giving data center operators the flexibility to choose accelerators from different manufacturers.
Multiple servers can be linked into a cluster using OSFP, an open-source routing technology. Each graphics card in the system has nine ports with 1.6 terabits per second of bandwidth. According to reporting by GN auto tech/hardware, this network capacity is comparable to recent high-end appliances from Nvidia, suggesting the hardware is competitive in raw data throughput.
New Chip Reduces Network Latency
Alongside the funding, Delos introduced the Nonstop AI Data Interface, a chip designed to connect servers to the cluster network. The company claims this component provides ten times lower latency than rival products. It also monitors the system for issues like graphics card crashes or network outages and automates some recovery tasks.
The technology is available in three forms: a PCIe card for existing servers, a chiplet that can be integrated directly into AI accelerators, and a near-packaged optics version. The chiplet offers the highest speed, handling over 30 terabits per second, while the near-packaged version is simpler to implement but has lower capacity.
Trade-Offs in Integration Options
Customers must balance performance against implementation complexity. The direct chiplet integration provides the highest throughput but requires deeper modification of the accelerator hardware. The near-packaged optics option is easier to deploy but delivers less than half the bandwidth of the chiplet version.
Delos plans to use the new capital to hire engineers and expand its sales team. The company’s strategy relies on the premise that agentic AI workloads will outgrow single-device capabilities, making high-speed data interfaces a critical component for scalable AI infrastructure.
