Delos Data
Delos Data builds interconnect systems that scale AI inference across thousands of GPUs and accelerators.
Delos Data builds interconnect systems that scale AI inference across thousands of GPUs and accelerators.
Nonstop AI is Delos Data's unified system architecture for accelerating AI inference through cabled scale-up, combining a modular server platform with network orchestration software.
The Delos Server chassis brings scale-up interconnect to the front panel through OSFP ports compatible with existing GPU and accelerator ecosystems, while the Delos Mosaic software provides fine-grained visibility into scale-up domains. Early access software deployments are available to select customers, with broader availability planned for the fourth quarter of 2026.
Delos Data positions its offering against an AI inference interconnect market it estimates will reach roughly one hundred billion dollars by 2030, as accelerator clusters grow into the thousands of devices.
The company targets hyperscalers, AI infrastructure builders, and enterprises scaling inference, emphasizing reduced cost, power consumption, and latency per token as the workload mix shifts from training toward inference.
Delos Data argues that its modular, disaggregated server design offers greater flexibility than integrated rack-scale systems, letting operators size scale-up domains to the workload and use standard cabling and off-the-shelf switches.
By relying on OSFP ports and merchant packet switches, the platform supports flexible topologies and avoids the custom mechanical, thermal, and power engineering that proprietary racks require, which the company says lowers the barrier for AI chip startups reaching rack scale.
The Register observes that Delos Data's reliance on pluggable OSFP optics could make power consumption problematic for larger compute domains, since copper cabling offers a fraction of the power draw of pluggable optics.
The same modular simplicity that gives Delos flexibility also means its platform is not a custom-integrated rack like Nvidia's or AMD's designs, which the report notes would be easier to service in some respects. Delos counters that its modular design trades that integration for greater workload flexibility.