A Duke-born photonics startup rebuilding the processor around metamaterials and photons - and betting it can out-efficiency the GPU.
Every large AI model, underneath the chat window, is doing the same unglamorous chore over and over: multiplying matrices. Today that math runs on silicon GPUs, moving electrons through transistors and paying for it in heat and electricity. Neurophos, a photonics startup based in Austin, Texas, thinks the answer is to stop moving electrons at all.
The company builds an Optical Processing Unit (OPU) - a chip that performs matrix-vector multiplication with light instead of current. Its core trick is a metasurface optical modulator, a micron-scale metamaterial component the company describes as roughly 10,000 times smaller than conventional optical transistors. Shrinking the part that much is what lets Neurophos fit more than a million optical processing elements onto a single chip - the density that makes photonic computing practical rather than a lab curiosity.
The pitch to a data-center operator is simple: the OPU is meant to be a drop-in, energy-efficient alternative to the GPU for AI inference - the part of AI that answers queries, as opposed to training. Neurophos frames the whole business around one metric it argues actually matters: dollars per inference.
The physics origin story is unusually good. Neurophos spun out of the Duke University lab of Prof. David R. Smith, the same metamaterials research that produced early-2000s "invisibility cloak" demonstrations. The technology that once bent radar around objects is now being pointed at the most expensive arithmetic in the data center.
AI's constraint is increasingly not intelligence but electricity. As models scale, GPU-based data centers hit limits on power draw and cooling, and the cost of every generated answer climbs. Neurophos argues the fix has to happen at the level of the device itself.
CEO Patrick Bowen's contention is that GPU power consumption grows roughly linearly with chip area, so scaling compute 100x would demand unsustainable power and cooling. Neurophos claims its optical design changes that curve - with power scaling closer to the square root of area - so that efficiency and speed improve together as the system grows. If that holds in production, it inverts the economics that constrain today's racks.
Neurophos's own published targets put its OPU well ahead of a current flagship GPU on raw operations per second - at lower power. These are company figures for pre-production silicon, so read them as ambition, not benchmark.
A photonic AI inference chip integrating 1M+ micron-scale metamaterial modulators to do matrix-vector math with light. Designed as a drop-in GPU alternative for data centers.
The metamaterial building block - roughly 10,000x smaller than a conventional optical transistor - that makes dense, manufacturable optical computing possible.
A folded-pipeline optical systolic array clocked near 56 GHz that streams data through analog in-memory optical compute, targeting exaflop-scale throughput on one chip.
PyTorch models and early-access hardware let customers compile their models and measure throughput and efficiency before commercial chips ship.
This physics-level shift means both efficiency and raw speed improve as we scale up, breaking free from the power walls that constrain traditional GPUs.
The obvious incumbent is Nvidia, whose GPUs are the default for AI work. Neurophos isn't trying to build a faster GPU - it's changing the substrate, computing with photons rather than electrons. That's the core differentiator, and it's also the core risk: optical computing has been "a few years away" for two decades.
Within photonics, the distinction matters. Several peers - Lightmatter, Ayar Labs and others - have leaned toward optical interconnect, using light to move data between chips faster. Neurophos is attempting the harder thing: doing the actual computation optically. Other novel-compute challengers include Lightelligence and Celestial AI.
Neurophos's claimed edge is manufacturability. By making its metamaterial modulators small enough to pattern with standard semiconductor processes, it argues photonic compute can finally be built at data-center scale on existing fab lines rather than in a research setting.
Its market is data-center AI inference - hyperscalers, cloud providers and enterprises running large models at volume. The company sells hardware and systems with a software stack, and courts early customers through its developer program ahead of a commercial ramp targeted for 2028.
On January 22, 2026, Neurophos announced a $110 million Series A led by Gates Frontier, bringing total funding to roughly $117-118 million. The investor list is unusually broad for a chip startup.
Software (M12), energy (Aramco), industrials (Bosch) and climate (Carbon Direct) rarely agree on a single chip. Their shared bet isn't really on a product - it's on the future cost of electricity for AI. The round funds data-center-ready photonic systems and engineering expansion in Austin and San Francisco.
David Smith's lab pioneers the metamaterial research behind early "invisibility cloak" demonstrations - the physics that later underpins Neurophos.
Patrick Bowen and Andrew Traverso commercialize metasurface optical computing out of Smith's Duke lab, incubated by Metacept.
The company advances its Optical Processing Unit and opens a developer partner program with PyTorch models.
Gates Frontier leads a $110M round; total funding reaches ~$117-118M to fund data-center-ready photonic systems.
Neurophos targets first commercial OPU systems and production ramp.
Physicist and entrepreneur with a Master's in Micro-Nano systems from ETH Zurich and a PhD from Duke under metamaterials pioneer David Smith. Co-founder of the Metacept commercialization center.
Duke researcher and co-founder who helped translate the lab's metasurface work into Neurophos's optical processing hardware.
Around 55 people work at Neurophos, drawing chip and photonics veterans from Nvidia, Apple, Samsung, Intel, AMD, Meta, ARM, Micron, Mellanox and Lightmatter - a deep-tech, physics-first bench assembled before the company had a commercial chip to sell.
Search these to go deeper on Neurophos's technology and story:
▶ Patrick Bowen interviews ▶ OPU / product demos ▶ SemiWiki CEO interviewAn Optical Processing Unit (OPU) - a photonic AI inference chip that uses light and metamaterial modulators to perform the matrix math behind AI models, positioned as an energy-efficient alternative to GPUs.
Co-founded by CEO Patrick T. Bowen and Andrew Traverso, it spun out of Duke University's metamaterials lab and the Metacept incubator, and is headquartered in Austin, Texas.
About $117-118 million total, including a $110 million Series A announced in January 2026 led by Gates Frontier, with M12, Carbon Direct, Aramco Ventures, Bosch Ventures and others participating.
Instead of moving electrons through silicon, Neurophos computes with light using micron-scale metamaterial modulators, claiming much higher operations-per-watt and performance that improves as the chip scales up.
Neurophos offers early-access developer hardware and software now, but first commercial systems and production ramp are targeted for 2028.