S Korean chip firm Panmnesia secures $10m AI data center R&D project

Panmnesia, a semiconductor startup developing interconnect chips for AI data centers, secured an R&D project worth about US$10 million.

The project focuses on building controller and switch SoCs for accelerator links based on UALink and Ethernet.

The company said the project targets chips that enable high-speed data exchange between AI accelerators. Its switch SoC, designed for accelerator-centric interconnects such as UALink, is expected in the second half of 2027.

UALink is an open standard backed by companies including AMD, AWS, Google, Microsoft, and Meta to connect AI accelerators across vendors.

Panmnesia said it already offers CXL products for memory expansion in data centers and introduced a CXL-over-XLink architecture last year.

🔗 Source: Panmnesia

🧠 Food for thought

Implications, context, and why it matters.

Panmnesia joins a packed race to build AI connectivity

Today’s huge AI models need “scale-up” networks that link hundreds or thousands of accelerators with very low latency, which standard Ethernet was not built to handle 1.
Nvidia leads this layer with its proprietary NVLink interconnect 2.
Companies are building open options, including UALink for shared-memory workloads, plus the Ethernet Scale Up Network (ESUN) initiative that aims to reshape Ethernet for scale-up AI networking 1.
Panmnesia plans chips for UALink and for Ethernet, which could make it a supplier to open-standard groups working on multi-vendor substitutes for Nvidia’s more closed interconnect approach 3.

Open interconnects could move leverage toward parts makers

Open interconnect work such as UALink fits a wider move away from single-vendor, fully integrated AI systems 2.
That shift opens room for focused chip vendors like Panmnesia, where gear from multiple companies can operate together in one setup 3.
Large cloud providers supporting these efforts want less vendor lock-in, a steadier supply chain, plus the ability to mix accelerators with networking equipment 3.
Breaking the AI stack into separate pieces may echo how PCs and servers evolved, which could intensify competition among individual components instead of whole proprietary platforms 1.