At the recent OCP Global Summit, NVIDIA announced that Samsung Foundry has become a key partner in its NVLink Fusion ecosystem. This collaboration brings Samsung’s advanced design and manufacturing capabilities into the fold, enabling the development of custom CPUs and XPUs that integrate seamlessly with NVIDIA’s high-bandwidth NVLink technology. The partnership positions Samsung to provide comprehensive support, covering everything from silicon design and verification to integration and tape-out for chips designed to connect directly with NVLink-enabled MGX and OCP rack architectures.

Expanding Custom Silicon Options for AI Data Centers

As hyperscalers and AI data centers seek higher performance and tighter CPU-GPU integration, the addition of Samsung Foundry to the NVLink Fusion ecosystem opens new possibilities for custom silicon development. NVLink Fusion serves as both an intellectual property (IP) and chiplet framework, allowing non-NVIDIA processors and accelerators to participate natively in NVIDIA’s data center fabric. This means that custom chips from various vendors can now interoperate within the same MGX and OCP designs, enhancing flexibility and scalability for large-scale AI workloads.

Samsung’s entry brings not only additional fabrication capacity but also robust design services, complementing the work of existing partners and chipset makers. NVIDIA’s engineering efforts on the Vera Rubin OCP rack and MGX integration have been instrumental in enabling these cross-vendor connections. Industry leaders such as Intel, Fujitsu, and Qualcomm are already part of this expanding custom-silicon network, and Samsung’s involvement is expected to further accelerate innovation.

Integrating Samsung Foundry into the NVLink Fusion ecosystem is set to streamline development cycles, reduce reliance on single-source suppliers, and offer organizations more options when designing CPUs and XPUs optimized for large language models and next-generation AI applications. For system integrators, this translates to faster prototyping of workload-specific hardware and more straightforward paths to volume production. For example, partners can now connect custom CPUs directly to NVIDIA GPUs using the high-speed 900 GB/s NVLink-C2C interface.

However, the collaboration comes with important restrictions. Any custom chip developed within this framework must be connected to an NVIDIA product, and NVIDIA maintains strict control over the critical software and hardware layers that manage these connections. Specifically, NVIDIA retains authority over the communication controller and PHY layers in NVLink Fusion, which are essential for initializing and managing links. Additionally, third-party hardware requires a license to use NVIDIA’s NVLink Switch chips, limiting the ability to create fully independent, mix-and-match systems.

Overall, Samsung Foundry’s participation in the NVLink Fusion ecosystem marks a significant step forward for custom silicon in AI and data center environments, offering new opportunities for performance optimization while maintaining NVIDIA’s established standards for integration and control.