d-Matrix buys Wallaroo to orchestrate chip inference

Silicon startup d-Matrix announced on August 3, 2026, its acquisition of Wallaroo.ai, a developer of AI inference orchestration software. This acquisition, the chipmaker's second in four months, integrates Wallaroo's platform and engineering staff into d-Matrix. The move addresses a software challenge for d-Matrix's Corsair accelerators, which entered production on June 9, 2026. Manufactured with Alchip on TSMC's N6 node, Corsair uses SRAM-based in-memory compute chiplets and LP-DDR5 memory to run air-cooled. Instead of replacing GPUs, Corsair runs beside them, handling the memory-bound decode phase of language-model requests while GPUs manage the compute-heavy prefill. Wallaroo's control plane, supporting x86, Arm, GPU, vLLM, and SGLang, will orchestrate this split.

This acquisition reflects an industry trend where hardware vendors buy software layers; examples include Qualcomm's July 2026 purchase of Modular, Nscale's acquisition of Anyscale, and Nebius's $643 million deal for Eigen AI. To fund its expansion, d-Matrix relies on a $275 million Series C closed on November 12, 2025, at a $2 billion valuation, co-led by BullhoundCapital, Triatomic Capital, and Temasek. The company previously acquired GigaIO's data center business in April 2026, absorbing its SuperNODE system and FabreX PCIe-based memory fabric, and has partnered with Infineon.

For practitioners, managing mixed-silicon clusters is complex. Wallaroo founder Vid Jain noted on March 16, 2026, that agentic traffic is highly variable: 84% of requests are short (around 2,000 tokens), 15% are medium (8,000 to 64,000 tokens), and under 1% exceed 128,000 tokens. Their cache-aware routing cut worst-case time-to-first-token by 75% in tests. Furthermore, a March 11, 2026, study by Gimlet Labs—co-authored by d-Matrix CTO Sudeep Bhoja—demonstrated the power of this split. Running gpt-oss-120b with a 1.6-billion-parameter draft model, moving speculative decoding to Corsair's 2GB of on-chip SRAM and 150 TB/s bandwidth while keeping prefill on GPUs yielded 2x to 10x faster end-to-end requests. Already, Parasail is deploying Corsair alongside Nvidia Hopper and Blackwell fleets across 40 data centers in 15 countries. By owning Wallaroo's orchestration layer, d-Matrix CEO Sid Sheth can now deliver these performance gains without forcing developers to build custom routing software themselves.