Nvidia Swallowed Enfabrica for $900 Million

There is no official press release even a blog post on that move but it is reported by sources close to the deal.Nvidia has silently acquired Enfabrica for $900 million in cash and stocks options. This operation is not a classic acquisition as Nvidia recruits Enfabrica CEO and several key people and licenses the technology, in other words this is a talent and IP acquisition. By some aspects, it reminds us a bit what they did with SwiftStack with just a blog post in 2020, same for Excelero in 2022 and Nebulon in 2024 without any mention on any website, not even a blogpost. Now these websites redirect to Nvidia.com obviously.

Such move allows entities to collaborate without the need to wait legal or regulatory validation like a real acquisition. We saw this kind of operations with Google and Windsurf or META with ScaleAI a few months ago.

Founded in 2019 by Rochan Sankar, co-founder, president and CEO, and Shrijeet Mukherjee, co-founder and CDO, the company has raised a total of $260 million. Nvidia has invested in the $125 million Series B in September 2023 with 8 other investors. These two founders have spent time at Broadcom, Cypress Semiconductor, Google, Cumulus Networks, Cisco and SGI among others, illustrating their strong background in networking and semiconductor domains.

With the AI pressure on infrastructure, new designs and modern solutions are required especially on the interconnect between processors, memory and storage. The motivation is to reduce hops and latency, improve bandwidth and load distribution in an a any-to-any manner to glue together 100,000s GPUs.

This is why Enfabrica has been founded: design a radical new approach coupling hardware and software to build gigantic clusters with tons of GPUs. By leveraging memory disaggregation, Enfabrica wishes to address HBM costs by considering DRAM, HBM, memory pool via CXL and storage. The second goal is to avoid point to point GPU connection and promotes a multipath architecture thanks to its ACF-S – Accelerated Compute Fabric SuperNIC – chip displayed above.

The architecture supports PCIe Gen 5 and 6, CXL, 800GbE, RDMA…instantiated with its Elastic AI Memory Fabric System aka EMFASYS-powered by its ACF-S chip. It allows farms of GPUs to be connected to racks of memory via CXL. The interconnect battle will continue for sure…

Click to enlarge