Enfabrica bet AI clusters starve on memory, not compute; Nvidia paid $900M to license it

A Mountain View startup building ACF SuperNIC chips and software that tie massive GPU clusters into effectively one machine, treating memory and network as one composable fabric.

That AI clusters are bottlenecked by memory and I/O bandwidth rather than GPU compute — and that the fabric tying 100,000+ GPUs together is worth a company.

$125M Series B (Sept 2023, led by Atreides Management, with Nvidia investing) and a $115M Series C (Nov 2024, led by Spark Capital with Arm, Samsung Catalyst Fund, Cisco Investments, Maverick Silicon and VentureTech Alliance) — about $600M post-money per PitchBook.

Enfabrica was founded in 2019 in Mountain View by Rochan Sankar, who had led Broadcom's cloud and data-center switch silicon business and previously worked at Cypress Semiconductor and Cswitch. His thesis: AI scaling would hit I/O and memory walls long before compute walls. The company emerged from stealth in 2023 and raised a $125M Series B that included Nvidia itself.

In November 2024 Enfabrica raised a $115M Series C led by Spark Capital with Arm, Samsung Catalyst Fund, Cisco Investments and Maverick Silicon — a round Crunchbase News noted as it announced the ACF SuperNIC, pitched as the world's fastest GPU network interface controller and able to connect clusters of more than 100,000 GPUs. The chips treat memory and networking as one fabric, so accelerators can be fed and address memory across racks instead of starving on the PCIe links inside a single server.

Nvidia licensed Enfabrica's technology and hired Sankar in a deal of over $900 million in cash and stock that closed the week before CNBC reported it on September 18, 2025. Nvidia structured it as a license-plus-hire rather than an acquisition — the same shape as Meta's Scale AI, Google's Windsurf and Microsoft's Inflection deals — a notable move for a company whose only big acquisition was Mellanox and whose attempt to buy Arm collapsed under regulatory pressure in 2022.

Sankar came from Broadcom's switch-silicon business, so the team knew from the inside that interconnect and memory bandwidth — not FLOPs — was where clusters actually choke.

Owning the fabric layer made Enfabrica strategically valuable to Nvidia regardless of whether it ever won merchant market share on its own.

Nvidia had both the motive (racks like GB200 live or die on interconnect) and the constraint (antitrust scrutiny after Arm) that made a $900M license-plus-hire rational.

Raising from Arm, Samsung and Cisco alongside Nvidia hedged the company's future among several potential acquirers and partners.

Selling a license to the incumbent can beat displacing it: build the one capability the leader still needs, price it high, and skip the antitrust fight that kills startup acquisitions.

As of late September 2025, the deal had closed and Sankar had joined Nvidia; SDxCentral reported the license covers interconnect solutions capable of connecting hundreds of thousands of GPUs. It landed in the same week Nvidia announced a $5B stake in Intel and close to $700M for UK cloud startup Nscale — an unusually acquisitive stretch for a company that had previously bought almost no one. Enfabrica announced no wind-down, and its ACF SuperNIC remained on the availability list announced in November 2024.