Here is a fact about artificial intelligence that does not fit neatly on a keynote slide: a large chunk of the electricity a data center burns is not spent thinking. It is spent moving. Bits shuttle between thousands of GPUs, and every time they do, they pass through switches - racks of electrical hardware that convert light back into electrons, make a routing decision, and convert them back again. nEye.ai, a semiconductor company in Santa Clara, looked at that arrangement and asked a question that sounds naive until you sit with it: what if the light never had to stop?
That is, more or less, the entire pitch. nEye - legally nEye Systems, Inc. - builds an optical circuit switch, or OCS, and it builds it onto a single chip. The trick is to fuse three things that normally live in separate worlds: MEMS (micro-electromechanical systems, which is a fancy way of saying tiny mechanical mirrors), silicon photonics (steering light through a chip), and plain CMOS (the logic that makes it all controllable). Put them on one wafer and you get a switch that redirects beams of light directly, without the expensive round-trip into electrical hardware.
The reason anyone cares is timing. Optical anything has been "five years away" in the networking industry for about two decades. What changed is that AI made the electricity bill enormous. When your data center is measured in megawatts, a switch that claims to use a thousandth of the power stops being an academic curiosity and starts being a line item worth fighting over.
The SuperSwitch, and some big numbers
nEye's flagship product is called the SuperSwitch. It is a high-radix optical circuit switch - "radix" meaning how many things it can connect at once - built at wafer scale. The company makes a set of comparison claims against existing optical switches that are, to put it gently, attention-getting: roughly 100 times smaller, 1,000 times lower power, 10,000 times faster to reconfigure, and 10 times cheaper. These are the company's own figures, not an independent benchmark, and the right way to read them is as a statement of ambition. But even if reality lands at a fraction of each, the direction is the interesting part.
Figures above are nEye's stated claims versus existing optical switch solutions, not independently verified benchmarks.
What you can actually do with this, if it ships at volume, is rewire the economics of an AI cluster. nEye describes a world where GPU, CPU and memory stop being fixed to their racks and instead get pooled - connected on demand through a fast optical layer, so an operator can compose exactly the machine a given workload needs and recompose it an hour later. That flexibility is the sort of thing hyperscalers spend a great deal of money chasing, because the alternative is buying hardware for your peak and letting it idle the rest of the time.
A Berkeley lab with a very long fuse
nEye did not appear from nowhere in a burst of AI enthusiasm. It was founded in 2020 by researchers out of the University of California, Berkeley, and the science underneath it goes back much further. Co-founder and chief scientist Ming C. Wu runs Berkeley's Photonic Systems Integration Lab and has spent a career on exactly this problem. CTO Tae Joon Seok earned his PhD there and specializes in MEMS-based optical devices. Co-founders Kyungmok "Mogi" Kwon and Xiaosheng Zhang round out a founding team that reads like a photonics reading list.
The person tasked with turning that into a business is CEO Ashish Vengsarkar, who is not an academic but a commercialization veteran - more than two decades at Lucent and Alcatel-Lucent building optical components for the telecom industry. This is a useful pairing. The hard part of a company like nEye is rarely the physics; the physics is in the papers. The hard part is manufacturing - taking a beautiful wafer-scale demo and making millions of identical, reliable parts that a customer will bet their data center on.
Who is paying for this
In April 2026, nEye raised an $80 million Series C led by Sutter Hill Ventures, whose managing director Stefan Dyckerhoff joined the board. That followed a $72.5 million raise the year before, bringing the total to roughly $152.5 million. The cap table is worth pausing on: it includes both CapitalG, Alphabet's growth fund, and M12, Microsoft's venture arm, plus Socratic Partners. When the two companies that arguably operate the world's most demanding AI data centers both back the same photonics startup, that is a signal that has less to do with fashion and more to do with their own electricity bills.
| Round | Amount | Date | Lead / Notable |
|---|---|---|---|
| Series C | $80.0M | Apr 2026 | Sutter Hill Ventures |
| Series B | $72.5M | Apr 2025 | CapitalG, M12, Socratic |
| Total | ~$152.5M | — | — |
The stated use of the new money is refreshingly unglamorous: scale foundry-based, high-volume manufacturing. That is the tell that a hardware company is serious. Anyone can raise on a demo; the interesting sentence is the one about yield and volume. The optical circuit switching market, for what it's worth, is projected to grow past $3 billion within a few years - big enough to matter, small enough that whoever nails the chip early gets a real head start.
What you'd actually buy
It helps to be concrete about who nEye sells to, because "optical circuit switch" is not a phrase that shows up in a consumer's life. The buyer is an infrastructure operator - a hyperscaler, a cloud provider, an HPC lab - building a fabric that has to connect thousands of accelerators without melting or bankrupting the building. Today that operator stitches those accelerators together with tiers of electrical switches, optical transceivers, and a lot of cabling, each layer adding latency, heat, and cost. nEye's proposition is to collapse a chunk of that stack into a chip-scale optical layer that switches the light directly.
The practical payoff, if it lands, comes in three flavors. First is power: an optical switch that idles cool changes the math on a facility's total draw, which is the single most binding constraint on new AI capacity. Second is density: 100 times smaller means more switching in the same rack, which means shorter reaches and less cabling. Third - and this is the one operators dream about - is composability. A fast optical layer lets you treat GPU, CPU and memory as a shared pool rather than fixed furniture, wiring up the exact machine a job needs and tearing it down when the job ends. Utilization is where data-center money quietly leaks, and composability is a way to plug the leak.
nEye is roughly 45 people, which is small for the size of the problem and typical for the stage. The culture reads as a research group growing a manufacturing spine: a founding team steeped in a decade of Berkeley photonics work, now paired with a CEO whose whole career was about turning optical science into shippable telecom parts. The interesting organizational question for a company like this is not whether the ideas are good - they are published and peer-reviewed - but whether it can graft factory discipline onto a lab sensibility fast enough to meet demand it did not have to manufacture two years ago.
The catch, stated plainly
None of this is shipping in your favorite chatbot yet. nEye's own framing is that its recent milestone "validates" the technology and the work now is manufacturing - which is to say, the company is standing at the exact spot where many brilliant deep-tech startups have stumbled: the chasm between a stunning lab result and a boring, dependable supply chain. It also has company. Google runs its own in-house optical switching in its data centers, and a cohort of photonics-interconnect startups - Lightmatter, Ayar Labs, Celestial AI among them - are attacking neighboring pieces of the same problem. nEye's bet is specific: not just photonics, but the whole switch, on one chip. The distinction matters. Plenty of companies are making the links between chips faster; nEye is trying to own the intersection where those links meet and get redirected. If it works, that intersection becomes a chokepoint of the good kind - the piece everyone else has to route through.
Whether that bet pays off will be decided in a foundry, not on a slide. But the underlying observation is hard to argue with. AI's story is usually told as a story about models. It is really a story about physics - heat, power, and the speed of light - and nEye.ai has planted itself squarely at that physical edge, betting that the future of AI infrastructure runs on photons.