d-Matrix Launches Squadron Clusters for Megawatt-Scale AI Inference
Event summary
- d-Matrix introduced Squadron, multi-rack systems scaling up to 1,024 Corsair XPUs for ultra-low-latency AI inference.
- Squadron clusters are air-cooled, industry-standard, and deployable in existing data centers without liquid cooling.
- d-Matrix demonstrated Attention-FFN Disaggregation (AFD) with Squadron, achieving 20,415 tokens per second in a heterogeneous setup.
- Squadron is now shipping to priority customers, with Corsair servers and SquadRacks generally available.
The big picture
d-Matrix is transitioning from a chip company to a full-stack data center infrastructure provider, competing directly with NVIDIA’s dominance in AI inference. The launch of Squadron underscores the industry’s shift toward disaggregated, memory-centric architectures for latency-sensitive applications. The success of AFD could redefine how AI workloads are distributed across heterogeneous systems.
What we're watching
- Adoption Pace
- How quickly enterprises and AI labs will integrate Squadron into their data centers.
- Competitive Response
- Whether NVIDIA or other AI hardware providers will counter with similar disaggregated compute solutions.
- AFD Viability
- The extent to which Attention-FFN Disaggregation becomes a mainstream AI inference approach.
Related topics