What happens when an inference chipmaker plugs into NVIDIA's rack-scale fabric?

d-Matrix makes chips for AI inference. The company announced on September 10, 2026 that its next processor generation will plug directly into NVIDIA's platform.

The connection uses NVLink Fusion. This is the high-bandwidth, low-latency technology that links custom XPUs and CPUs to the NVIDIA stack. Sid Sheth spoke during a press briefing the day before. He is the cofounder and CEO of d-Matrix. He said demand for inference is soaring. Capital remains finite. Time remains finite. Energy remains finite. With NVLink Fusion and MGX, the company can integrate Raptor XPUs into a broadly deployed, liquid-cooled architecture. This gives customers a faster, lower-risk path to deploy and scale ultralow-latency inference.

The integration path

Building an XPU is only the first step. Deploying it at AI factory scale requires a complete platform. The platform spans networking, rack architecture, power, cooling, software and a proven supply chain. Each step adds time. Each step adds costs. Each step adds risks. Sourcing chips. Integrating high-speed interfaces. Validating a scale-up networking solution. Designing and certifying rack architecture. NVLink Fusion lets silicon innovators connect directly into NVIDIA's proven platform.

The technology extends the platform's openness to XPUs and CPUs. Silicon companies can focus on their processor innovations while using NVIDIA infrastructure to deploy at AI factory scale. The company plans to connect its XPUs in a single high-bandwidth, low-latency scale-up domain using NVLink. Its racks can also work alongside NVIDIA GPU-based systems like NVIDIA Vera Rubin NVL72 for disaggregated inference.

What the rack provides

The MGX ecosystem offers mature, validated rack designs. It includes supply chain. It includes power and cooling infrastructure. By standardizing on a common rack, data centers can be built once and support GPUs, CPUs and XPUs without requiring a separate rack architecture for each processor type. This is the practical change. A facility built for one compute type suddenly accepts another.

The firm also plans to integrate NVIDIA Vera CPUs. It plans to integrate NVIDIA ConnectX-9 SuperNICs. It plans to integrate NVIDIA BlueField-4 DPUs and NVIDIA Spectrum-X Ethernet networking. Together with NVLink and MGX these technologies give the company a proven foundation to deploy specialized inference alongside NVIDIA systems within flexible, unified AI factories.

Where the XPU fits

The NVIDIA full-stack AI factory platform includes NVIDIA Vera Rubin NVL72. It includes Groq 3 LPX. It includes the Vera CPU rack. It includes Vera BlueField-4 STX storage and Spectrum-6 SPX Ethernet networking. It is designed to be completely fungible. The platform runs every AI workload. It runs every model and model architecture with the best performance per watt and lowest cost per token. NVLink Fusion gives customers the flexibility to match the right compute to each workload within a common AI factory platform.

Silicon innovators like d-Matrix can use NVLink Fusion to increase performance. They can accelerate time to market. They can reduce the risk of deploying semi-custom AI factories. The announcement does not specify how many Raptor XPUs will ship or when the first rack-scale deployment goes live. It also does not disclose power draw or latency figures for the integrated system.