Nvidia intends to combine Enfabrica’s tech stack with its own hardware and software to offer more efficient training clusters for more powerful frontier models, analysts say.
Nvidia has hired away the CEO and other staff of chip interconnect maker Enfabrica, and licensed its core technologies in a deal worth over $900 million, according to reports from CNBC and The Information.
Demand for computing capacity to power generative AI for the likes of OpenAI, Anthropic, Mistral, AWS, Microsoft, and Google is posing a challenge to Nvidia: how to build a unified, fault-resistant GPU cluster that can handle such enormous workloads.
[ Related: More Nvidia news and insights ]
Nvidia has been an investor in Enfabrica since September 2023 and, according to analysts, sees the integration of Enfabrica’s technology as critical to increasing the efficiency of its clusters and making them capable of training the next frontier AI model.
“By pulling in Enfabrica’s SuperNIC and pooled-memory fabric, Nvidia can help move data faster through its chip clusters, scale clusters beyond current network limits, and cut dependence on costly high-bandwidth memory (HBM),” said Rachita Rao, senior analyst at Everest Group.
Rao was referring to Enfabrica’s Accelerated Compute Fabric SuperNIC (ACF-S) silicon that the company claims is designed to deliver higher bandwidth, greater resiliency, lower latency and greater programmatic control to data center operators running data-intensive AI and HPC.
Enfabrica contends that ACF-S is more fault tolerant than traditional interconnect systems as it replaces point-to-point GPU connections with a multi-path architecture that reduces congestion, improves distribution of data, and ensures that GPU link failures don’t stall compute jobs.
Another Enfabrica technology that’s of interest to Nvidia, according to Forrester principal analyst Charlie Dai, is Elastic Memory Fabric System (EMFASYS) that became generally available in July.
EMFASYS provides AI servers flexible access to memory bandwidth and capacity through a standalone device that connects over standard network ports.
The combination of ACF-S and EMFASYS, according to Dai, might help Nvidia unlock higher GPU utilization rates and lower total cost of ownership — key metrics for hyperscalers and LLM developers operating at the cutting edge of AI.
Acqui-hires instead of acquisitions
Nvidia’s $900 million deal to absorb Enfabrica’s leadership and core technology can also be seen as a broader trend sweeping Silicon Valley, where traditional acquisitions are being replaced by strategic acqui-hires to prioritize talent and intellectual property.
Meta set the tone earlier this year with a $14.3 billion investment to onboard Scale AI founder Alexandr Wang and key personnel, acquiring a 49% stake in the startup to lead its superintelligence division. Google followed with a $2.4 billion agreement to bring in Windsurf CEO Varun Mohan and several R&D staffers, licensing the startup’s agentic coding tools for its Gemini AI platform.
Microsoft and Amazon’s deals with Inflection AI and Adept are also reminiscent of this pattern. The Inflection AI deal saw Mustafa Suleyman join Microsoft to head its AI division, while Adept co-founder David Luan was hired to head the e-tailer’s AGI efforts.
Analysts see this trend as a measure to sidestep regulatory scrutiny.
“Gutting Enfabrica allows Nvidia to move fast, sidestep messy product overlaps, and reduce regulatory drag. It gets the people and the IP that matter without the burden of integration baggage,” Everest Group’s Rao said.
However, Rao warned that regulators are now attuned to such workarounds.
“The FTC’s scrutiny of Microsoft–Inflection and Amazon–Adept shows that Nvidia’s play could easily attract attention even if it is not a traditional acquisition. But Nvidia can always argue that bringing Enfabrica’s talent in-house will accelerate innovation and solve a major industry bottleneck faster than if the startup remained independent,” Rao said.
Nvidia declined to comment on the reported acquisition. An email inquiry to Enfabrica went unanswered.
