AI: d-Matrix adopts Nvidia’s NVLink Fusion for its Raptor accelerators

According to Reuters, future Nvidia-compatible systems are expected to be available in 2027, while d-Matrix plans to finalize the design of Raptor by the end of 2026. The equipment is particularly aimed at applications where latency is critical, such as programming assistants, voice agents, and conversational services.

NVLink Fusion allows custom processors, or XPUs, to integrate into Nvidia’s hardware ecosystem. The American company presents this technology as a high-bandwidth, low-latency interconnection layer designed for semi-custom AI infrastructures. Nvidia states that NVLink 6 can connect up to 72 XPUs in an “all-to-all” mode, with a bandwidth of up to 3.6 terabytes per second per XPU. This enhancement is part of a broader expansion of the company’s infrastructure, which also aims for up to 2 GW of AI capacity in Australia by 2027.

For d-Matrix, this integration comes at a time when the AI market is gradually shifting from model training to large-scale execution. Inference refers to the phase during which a pre-trained model responds to user requests, generates text, processes voice, or produces code.

Raptor focuses on 3D memory

Raptor represents the next generation of accelerators from d-Matrix. In its technical documentation, the company explains that this architecture is based on its 3DIMC technology, which stacks computation and memory to reduce data transfers. d-Matrix presents this approach as a means to enhance both inference speed and energy efficiency for systems executing large models.

The start-up, based in Santa Clara, California, has been supported by Microsoft since a $110 million funding round completed in 2023. It was subsequently valued at $2 billion following a $450 million funding round in 2025, according to Reuters. Its first AI accelerator began shipping at the end of 2024.

A battle over inference

The alignment with the NVLink ecosystem illustrates the rise of specialized solutions that seek to operate alongside GPUs rather than completely replace them. d-Matrix is also collaborating with Astera Labs on connectivity solutions designed to ensure rapid data transfers within these systems. The financial terms of the agreement with Nvidia have not been disclosed.