Enfabrica

Here is a fact about artificial intelligence that gets less attention than it deserves: the expensive part is not always the thinking. It is the moving. A modern AI cluster is a warehouse full of graphics processors, each one a small fortune, and the job of turning them into something useful is largely the job of keeping every last one of them fed with data at the exact moment it needs it. Starve a GPU and you have bought a very expensive space heater. Enfabrica, founded in 2019 in Mountain View, built a company on the premise that the network - the connective tissue between all those chips - had quietly become the bottleneck nobody was pricing correctly.

This is not an obvious thing to build a company around. Networking silicon is unglamorous. It does not generate images of astronauts or write anyone's homework. It sits in the racks and moves bytes, and the highest compliment it can receive is that you never think about it. But the founders had spent their careers in exactly this unglamorous place. Rochan Sankar, the CEO, ran the data-center Ethernet switching business at Broadcom, where he helped ship four generations of the Tomahawk and Trident chips that a large fraction of the internet still runs on. His co-founder, Shrijeet Mukherjee, came out of Google, Cisco and SGI, sits on the Linux NetDev board, and holds 64 patents. These are people who find data movement interesting, which turns out to be a useful trait when data movement becomes the constraint on a trillion-dollar industry.

The product they built has a name that sounds like a boast and mostly isn't: the Accelerated Compute Fabric SuperNIC, or ACF-S. A NIC is a network interface controller, the thing that connects a server to the network. A "SuperNIC" is Enfabrica's claim that it built one from the ground up for AI, rather than adapting an existing design. The specific number is 3.2 terabits per second, which the company says is roughly four times the bandwidth of any other GPU-attached NIC, with a feature it calls multipath resiliency - the ability to spray traffic across up to 32 ports so that when a single link fails, and in a cluster of that size something is always failing, the billion-dollar training run does not simply stop.

There is an argument buried in the marketing here, and it is worth pulling out because it is the whole company. The word "fabric" is doing real work. A fabric is a thing you only notice when it tears. Enfabrica's bet is that as AI clusters grow - to 10,000 GPUs, to 100,000, to numbers that stop meaning anything intuitive - the failure modes stop being about any single chip and start being about the connections. Resiliency becomes the product. If you have spent hundreds of millions of dollars on accelerators, the thing you will pay almost anything for is the guarantee that they keep working together.

The second product is stranger and, in some ways, more interesting. In the summer of 2025 Enfabrica unveiled EMFASYS, which it calls the industry's first Ethernet-based AI memory fabric. The pitch inverts a piece of datacenter orthodoxy. GPUs come with a fixed amount of very fast, very expensive memory called HBM, and when AI teams run out of it - which happens constantly during inference, the process of actually running a trained model - the standard answer is to buy more GPUs. You do not want the compute. You want the memory that is welded to it. Enfabrica's response is essentially: stop doing that. EMFASYS uses the ACF-S chip to connect up to 144 lanes of comparatively cheap CXL-based DDR5 memory to the network, exposing as much as 18 terabytes of shared DRAM per node that any GPU server can reach over standard Ethernet ports. The claim is that you can stop overpaying for accelerators you only wanted for their memory, and make better use of the ones you already own.

Whether that reframing holds up in production is the kind of thing that gets litigated in benchmark wars for years. But the strategic logic is clean, and it is the sort of first-principles move - treat memory as a shared appliance instead of a fixed tax on every server - that tends to make incumbents nervous.

The money agreed. Enfabrica raised a $125 million Series B in 2023 led by Atreides Management, with Nvidia participating, and then a $115 million Series C in November 2024 led by Spark Capital - a round whose investor list reads like a who's who of the people who both compete with and depend on this kind of technology: Arm, Cisco Investments, Samsung Catalyst Fund, Maverick Silicon, VentureTech Alliance, and Nvidia again. That is a telling cap table. When your customers, your partners, and your would-be competitors are all writing the same check, it usually means you are building something structural rather than optional.

Then came the part of the story that is either an ending or a beginning, depending on how you read it. In September 2025, Nvidia agreed to pay more than $900 million, in cash and stock, to license Enfabrica's networking and memory technology and to hire Sankar and much of his engineering team. Crucially, this was not an acquisition. Enfabrica remained a standalone company; Nvidia licensed the intellectual property and hired the people. This structure has a name in Silicon Valley - the "acquihire" - and in 2025 it became something close to a genre. Buy the talent and the technology, leave the corporate shell standing, and in doing so sidestep the antitrust review that a full merger of an AI-networking startup into the world's most valuable chipmaker would surely attract.

It is a savvy piece of deal engineering, and it says something about the moment. The most valuable company on earth looked at a startup of roughly forty people and decided that its roadmap was worth nine figures and its CEO worth relocating. If you want a single data point on how much the industry has come to believe that the network is the new frontier of AI performance, that is the one. The GPUs get the headlines. The fabric between them just got a $900 million vote of confidence.

What happens to Enfabrica as a standalone entity - with its CEO and a chunk of its team now inside Nvidia, its technology licensed out, and a February 2025 R&D hub in India still spinning - is an open and genuinely interesting question. But the company already accomplished the hard part, which was to make a large number of very smart, very well-funded people agree that the boring problem was the important one. In infrastructure, that is the whole game.