AI silicon startup d-Matrix said on August 3, 2026 that it has acquired Wallaroo.ai, a Pittsburgh-based enterprise AI inference deployment and orchestration platform, folding the software and its engineering team into an end-to-end stack that runs generative and agentic workloads from d-Matrix's Corsair rack-scale silicon straight through to production endpoints.
Why d-Matrix wants an orchestration layer
d-Matrix has spent the last year hardening its Corsair PCIe accelerator — a digital in-memory compute card that pairs custom SRAM tiles with a 256 GB LPDDR5 memory pool to serve 100-billion-parameter models with up to 9,600 TFLOPs of MXINT4 throughput. Corsair entered full production in June and is now shipping to priority customers alongside NVIDIA's Blackwell rack in the hybrid inference reference design the two firms announced last month. What has been missing is a first-party deployment layer that lets enterprise data-science teams push models onto that silicon at scale.
Wallaroo.ai plugs the deployment gap
Wallaroo.ai has spent nearly a decade building exactly that. Its platform, used across financial services, healthcare, manufacturing and defense, gives ML engineers a unified way to package, deploy, monitor and observe models across heterogeneous CPU, GPU and NPU targets. d-Matrix said the deal brings Wallaroo.ai's software platform, intellectual property and expert engineering talent under its own roof.
Consolidating a rack-scale inference roadmap
The Wallaroo.ai integration follows the acquisition of GigaIO's data-center business in April, and further advances d-Matrix as a category leader for rack-scale heterogeneous AI solutions that pair GPUs with specialty XPUs. Combined with the company's Open Secure AI Alliance membership and its Series C investors — Bullhound Capital, Triatomic Capital, Temasek and strategic backer Microsoft — the buy signals that d-Matrix intends to compete for the enterprise inference budget rather than resell into someone else's stack.
Reporting based on coverage from PR Newswire, SiliconANGLE and BigDATAwire/AIwire.