Rochan Sankar bet the whole company on the plumbing nobody was watching
While the AI world fought over GPUs, he spent six years building the silicon that feeds them. Then Nvidia paid more than $900 million to bring the work in-house.
In 2019, the smart money in Silicon Valley was pouring into one thing: raw compute. Faster GPUs, more of them, packed into ever-denser racks. Rochan Sankar looked at the same picture and saw something the crowd had skipped over. The chips were getting fed by a network card that was, in his words, a "tiny element that was designed originally for pairing with a CPU." The processors were hungry. The plumbing hadn't kept up.
That gap became a company. Sankar left a comfortable senior role at Broadcom, teamed up with veteran network engineer Shrijeet Mukherjee, and started Enfabrica to solve a problem most of the industry had not yet bothered to name: in the age of generative AI, the bottleneck is not the compute. It is moving data to the compute fast enough to keep it busy.
Six years later, in September 2025, Nvidia agreed to pay more than $900 million to license Enfabrica's technology and hire Sankar along with much of his team. The bet on the plumbing had paid off.
The setupTwenty years learning one thing
Sankar did not arrive at this idea by accident. He is, by his own accounting, a 26-year veteran of the semiconductor business, and the career reads like a slow accumulation of exactly the expertise Enfabrica would need. He studied electrical engineering at the University of Toronto, then added an MBA from the Wharton School - a combination of deep silicon knowledge and business fluency that is rarer than it sounds.
He started at Cypress Semiconductor in 1999 as an applications engineer and chip architect, moved through a stint at the startup Cswitch, and in 2010 landed at Broadcom, where he eventually ran the data center Ethernet switching business. That is not a small posting. Broadcom's switching silicon is the connective tissue of the modern cloud, and Sankar was responsible for bringing multiple generations of the company's Tomahawk and Trident chips to market. He holds six issued patents.
We were the first to draw up the concept of a high-bandwidth network interface controller chip optimized for accelerated computing clusters. Rochan Sankar
So when he argued that data movement was about to become the defining constraint in computing, it was not a hunch from an outsider. It was a diagnosis from someone who had spent two decades building the very components that move bits through data centers. He knew the choke points personally.
The insightThe memory wall, in plain sight
The technical case behind Enfabrica is easier to grasp than it sounds. Picture a rack of GPUs chewing through an AI model. They can process staggering amounts of math per second. But all that math is useless if the data - the model weights, the training batches - cannot reach the chips quickly enough. Engineers call this the "memory wall": the widening gap between how fast a processor can compute and how fast you can feed it.
Why AI hits the wall · growth since ~2020 (illustrative)
GPUs make the wall worse. They have far more cores than a CPU and an "insatiable appetite for data," as Sankar has put it. The network interface card strapped to them - often a 100 or 200 gigabit part - was designed for a gentler era of general-purpose servers. Enfabrica's answer was the Accelerated Compute Fabric, or ACF: a single chip-plus-software system that consolidates networking components, connects GPUs, CPUs and memory pools directly, and moves data at multi-terabit speeds.
AI is pumping so much data in and out of the server nodes through a 100 or 200 Gig NIC - a tiny element that was designed originally for pairing with a CPU. Rochan Sankar
The flagship product, the ACF SuperNIC unveiled in its "Millennium" generation, runs at 3.2 terabits per second and is designed to knit together as many as 100,000 GPUs in a single fabric. When Enfabrica announced it, the company called it the fastest GPU networking chip in the world.
The moneyBuilding the picks and the pipes
Investors came around to the thesis in stages. Enfabrica raised roughly $240 million in total across three rounds, and the cap table itself tells a story about how central AI infrastructure had become.
The $115 million Series C, announced at the SC24 supercomputing conference in November 2024, was led by Spark Capital and drew new backers including Maverick Silicon and VentureTech Alliance. An earlier $125 million Series B was led by Atreides - and notably included Nvidia as a strategic investor, a detail that would look prophetic within two years.
The road to $900M · Enfabrica timeline
On the Series C, Sankar struck a measured, almost engineerly note. "This Series C fundraise fuels the next stage of growth for Enfabrica as a leading AI networking chip and software provider," he said, adding that the round "speaks to the commercial viability and value of our ACF SuperNIC silicon" and that the company was "well positioned to advance the state of the art in networking for the age of GenAI."
The exitNvidia's side door
Then came the deal that put Sankar's name in the trade press for a week. In September 2025, reports emerged that Nvidia would pay more than $900 million - in cash and stock - to license Enfabrica's technology and hire Sankar along with much of his team. He joined Nvidia's networking organization.
The structure was notable. Rather than a straight acquisition, it was an "acquihire": a license plus a hiring wave, a playbook big tech companies have used to bring in talent and intellectual property without triggering the full weight of a merger review. For a company whose earlier strategic investor was Nvidia itself, the arrangement had a certain symmetry. Nvidia had helped fund the plumbing; now it owned the right to use it and the people who built it.
We're well positioned to advance the state of the art in networking for the age of GenAI. Rochan Sankar
There is a founder lesson buried in all of this, and it is not the usual one about grit. Sankar did not build a flashier version of the thing everyone else was building. He found the constraint the market was ignoring - the unglamorous layer of wires and interfaces and memory access - and he had spent his entire career becoming the person best equipped to fix it. The gold rush went to the GPU makers. Sankar quietly sold the pipes underneath.
The characterA quiet operator
For someone at the center of a nine-figure deal, Sankar keeps a low public profile. There is no torrent of hot takes, no personal brand machine. What surfaces instead is a systems-level way of thinking - the habit of looking at the whole computing stack and asking where the real friction lives, rather than chasing whatever is loudest. He writes occasionally on Enfabrica's blog and Medium, usually to explain the technical guts of what the company is doing: memory tiering, CXL, RDMA networking, the mechanics of getting past the memory wall.
His co-founding partnership with Shrijeet Mukherjee reflects the same pragmatism. Between them they carried decades of network infrastructure experience from Broadcom, Google and Cisco - the kind of unglamorous, deep-in-the-plumbing resumes that turn out to matter enormously when the plumbing suddenly becomes the story.
What comes next for Enfabrica's remaining product and people, now that its founder and much of its tech sit inside Nvidia, is an open question. But Sankar's arc already makes a clean point. He spent twenty-six years learning one narrow thing - how to move data through silicon - and then waited, patiently, for the world to need it more than almost anything else. When AI turned data movement into the whole ballgame, he was ready.