The semiconductor startup selling the least visible, most necessary thing in an AI datacenter: the wires between the chips.
A logo the size of a business card for a company that fits inside a rack you will never open. Two friends started it in Yerevan in 2017. The plan was databases. The market had other ideas.
Here is a fact about artificial intelligence that nobody puts on a keynote slide: a large model is not trained on one chip. It is trained on thousands of them, and those thousands of chips spend an enormous amount of their time waiting to talk to each other. The compute is the glamorous part. The talking is the expensive part. Grovf builds the talking.
In February 2017, two friends and former colleagues in Yerevan, Armenia - Khachik Sahakyan and Artavazd Khachatryan - started a company around a deliberately unglamorous idea. What if the bottleneck in big data was not how fast you could compute, but how fast you could move data in and out of storage? Their answer was to do it in hardware, on FPGA chips, rather than in software. The first product was an FPGA-accelerated time-series database aimed at the Industrial Internet of Things - factory sensors generating more data than anyone could reasonably store.
It is worth pausing on how unfashionable this was. In 2017, the money was chasing apps. Grovf was writing register-transfer logic for programmable silicon. Their first funding was a $50,000 grant. That was, more or less, the entire company.
The through-line from that database to what Grovf sells today is a single idea repeated at larger and larger scale: move data in hardware, not software, and you win. A database is one place where that matters. It turns out the network between AI accelerators is a much bigger one. As the industry pivoted toward enormous models, the skill Grovf had been quietly compounding - low-latency data movement in silicon - stopped being a niche and started being the whole game.
So Grovf did not so much pivot as stand still while the market walked toward them. The FPGA database became FPGA IP for smart network cards. The IP for smart NICs became RDMA chiplets and turnkey networking IP for AI datacenters. The company that started by offloading factory-sensor databases now describes itself as building AI infrastructure backend solutions for trillion-parameter model training. Same core skill. Bigger room.
The word that keeps appearing in Grovf's technical materials is RDMA - Remote Direct Memory Access. It is one of those acronyms that sounds like plumbing because it is. RDMA lets one machine read and write another machine's memory directly, without dragging the operating system and its CPU into every transfer. In a normal network, moving data means copying it, interrupting the processor, and waiting. In an RDMA network, the data mostly just moves. When you are training a model across thousands of accelerators that need to swap gradients constantly, the difference between those two worlds is the difference between a cluster that scales and one that chokes.
Grovf's specific flavor is RoCE v2 - RDMA over Converged Ethernet - which matters because it means the fabric runs on ordinary Ethernet rather than exotic, single-vendor cabling. That is a strategic choice as much as a technical one. A standards-based fabric is a fabric a customer can adopt without betting the datacenter on one supplier. It also happens to be the layer where the industry's largest players are now spending enormous sums, which is either terrifying or validating depending on your temperament.
There is a version of this story that is purely about geography, and it is worth telling honestly. Grovf is registered in California and lists a Newark address, but the engineering center of gravity has always been Yerevan. Armenia has a deep bench of semiconductor talent - a legacy of Soviet-era microelectronics that never fully evaporated - and Grovf is one of the clearest examples of that talent building something aimed squarely at the global frontier rather than at outsourced contract work. The two-continent structure is not a curiosity. It is the business model: world-class silicon engineering at a cost structure that let the company survive on grants and a small seed round while it waited for its market to arrive.
“Grovf develops AI infrastructure backend solutions enabling trillion-parameter model training and deployment at scale.”
Grovf sells the same underlying thing - fast, low-latency data movement between accelerators - packaged three different ways depending on how much of the stack you want to own.
Advanced RDMA (scale-out) and scale-up chiplets delivered over the industry-standard UCIe interface, so you can drop networking straight into custom AI silicon.
Turnkey 400G/800G networking IP for scale-up and scale-out connectivity between accelerators - the hard part, pre-built and licensable.
PCIe cards delivering high-performance scale-out connectivity across accelerator servers, integrating into existing rack architectures.
Announced 2021: a low-latency FPGA IP core letting smart-NIC makers and integrators build RNIC use cases without writing RDMA from scratch.
The company's first product - hardware acceleration of core database functions for Industrial IoT big-data processing.
Integrate networking fabric directly into your chip via UCIe chiplets - skip building an interconnect team.
Connect accelerator servers at 800Gbps per port with sub-2-microsecond RDMA latency across 1M+ nodes.
License RDMA RoCE v2 IP instead of writing the protocol yourself - the part that usually eats a year.
Standards-based fabric means you are not locked into one vendor's accelerators or networking stack.
Background in Microwave Engineering and Photonics; lectured at Yerevan State University on computer modeling of physics problems before co-founding Grovf.
Friend and former colleague of Sahakyan; leads the silicon and technology side of the company from its Yerevan engineering base.
Sahakyan and Khachatryan start the company around FPGA-accelerated databases for the Industrial IoT.
Wins a $50,000 IMG grant for an FPGA-accelerated time-series / key-value database.
Selected for UC Berkeley's SkyDeck accelerator; becomes the first company from Armenia and the Caucasus to win an EU Horizon 2020 SME Instrument grant. Pre-seed led by SmartGate VC and Granatus Ventures.
Closes a $240K seed round, with Berkeley SkyDeck Fund among the backers.
Releases a low-latency RDMA RoCE v2 FPGA IP core for smart-NIC producers and system integrators.
Repositions around UCIe chiplets and 400/800G networking IP for AI datacenter scale-up and scale-out.
Grovf raised roughly $360K in total across grants and seed capital - a striking amount of restraint for a company entering one of the most capital-intensive markets in technology.
The interconnect layer of AI is suddenly crowded with giants. NVIDIA (via Mellanox and its ConnectX RDMA NICs), Broadcom, and Marvell own the incumbent networking silicon; newer interconnect players like Ayar Labs push optical chiplets; and open consortia around UCIe and UALink are racing to standardize AI scale-up fabric. Grovf's bet is that a standards-based, vendor-neutral fabric leaves room for a focused IP-and-chiplet supplier - the part of the market that does not require you to be a hyperscaler to matter.
Competing against NVIDIA on networking is, on paper, a suicidal proposition. But the interconnect market is not winner-take-all in the way GPUs sometimes are. System integrators, custom-silicon teams, and second-source-hungry hyperscalers all have reasons to want an alternative supplier of RDMA IP that is not tied to a single accelerator vendor. The whole point of UCIe and RoCE is that they are open standards; open standards create room for specialists. A company does not need to beat NVIDIA to build a durable business selling the networking block that goes into someone else's chip. It needs to ship IP that works, at latency numbers that hold up, to customers who would rather license than build. That is a narrower ambition than dethroning a trillion-dollar incumbent, and narrower ambitions are how small hardware companies survive long enough to matter.
The honest caveat: this is deep tech, and deep tech is slow, capital-hungry, and unforgiving. Grovf has been at it since 2017 on comparatively little money. Whether a lean, two-continent IP company can convert an early lead in a suddenly-hot category into real revenue is genuinely unknown - and the public record thins out considerably after the 2021 seed round. What is not in doubt is that the company placed its bet on the right layer of the stack years before that layer became fashionable. In hardware, being early and being patient are frequently the same virtue.
Profile compiled from public sources. Figures such as revenue and employee count are approximate. Contact: info@grovf.com