NVIDIA Adds Samsung Foundry for Custom Silicon Manufacturing

Posted on Edited on
NVIDIA expands NVLink Fusion ecosystem by adding Samsung Foundry for end-to-end design and manufacturing of custom CPUs and XPUs aimed at AI data centers.

Quick Report

TechPowerUp reports NVIDIA has added Samsung Foundry to its NVLink Fusion ecosystem, enabling Samsung to provide end-to-end design and manufacturing for custom CPUs and XPUs that interoperate with NVIDIA's NVLink-enabled MGX and OCP rack architectures. The collaboration promises faster prototyping, expanded fabrication capacity, and another route for hyperscalers seeking tighter CPU–GPU coupling.

The piece explains NVLink Fusion acts as an IP and chiplet framework: partners can design chips that speak NVIDIA's high-bandwidth fabric language but must integrate with NVIDIA's NVLink controllers, PHYs and switch chips. Samsung's entry brings both fabrication muscle and design services, shortening development cycles and reducing single-source risk — but partners remain bound by NVIDIA\‘s licensing and interface restrictions.

Key points:

  • Samsung Foundry joins NVLink Fusion to offer design-through-manufacturing services for custom silicon targeting MGX and OCP racks.
  • NVLink Fusion enables non-NVIDIA processors and accelerators to become first-class participants in NVIDIA's fabric, but requires adherence to NVIDIA\‘s proprietary controller/PHY and licensing.
  • The collaboration should help hyperscalers and AI data centers prototype workload-specific hardware faster and access additional production capacity.

Bringing Samsung Foundry into NVLink Fusion is strategically significant: it adds a major fabs partner and design services capability, widening the ecosystem for custom silicon that tightly integrates with NVIDIA GPUs. For hyperscalers this reduces vendor lock-in risk and shortens time-to-volume for specialized XPUs and CPUs optimized for NVLink-C2C interconnects.

However, the proprietary nature of NVLink Fusion means partners still operate within NVIDIA\‘s ecosystem boundaries — they can build custom compute elements, but interoperability and critical link-management remain controlled by NVIDIA. That trade-off will matter for organizations balancing performance, control, and supplier diversity.

Written using GitHub Copilot GPT-5 mini in agentic mode instructed to follow current codebase style and conventions for writing articles.

Source(s)

  • TPU
  • TrendForce