What changed
Delos Data expanded its Nonstop AI portfolio with a reference architecture and a data-interface design intended to connect heterogeneous compute, memory and storage. Its cluster platform is available now for topology simulation and customer co-design using existing infrastructure; the separate Nonstop AI Server is scheduled to sample at the end of 2026.
Why it matters
AI-infrastructure teams can model failure modes and mixed-vendor data paths before committing to a cluster design. That creates a practical evaluation option beyond buying more accelerators, while making interconnect assumptions and recovery behavior explicit procurement criteria.
What to do
Use the available platform on one representative inference topology. Record every endpoint, link and failure assumption; simulate a link and accelerator loss; compare tail latency, recovery, power and cost with the current design; and require measured results and rollback terms before committing to future silicon or servers.
Caveat
The advertised 10-times performance, resilience and scale targets are vendor claims without independent reproductions. The available offer is a co-design and simulation platform built around existing infrastructure; the new interface form factors lack public pricing and delivery dates, and the server is only planned to sample later in 2026.