The startup teaching artificial intelligence to do its arithmetic with light - and betting that photons, not transistors, are how AI stops eating the power grid.
Here is a fact that sounds made up but isn't: modern AI spends an enormous share of its energy not on thinking, but on moving numbers around. Electrons shuffle between memory and processor, hit a wall physicists have politely named the "memory wall," and generate heat the whole way. Opticore's argument is that this is not a problem you fix by building a bigger, hotter GPU. It's a problem you fix by changing what does the moving. Their answer is light.
Opticore, based in Berkeley, California, builds what it calls an Optical Processing Unit - an OPU, if you want the acronym that the company would very much like to make as familiar as GPU. The idea is that the same matrix multiplications that GPUs grind through electronically can instead be performed with photons traveling through waveguides on a chip. Light is fast, light generates little heat, and light can carry a lot of information in parallel. If you can get it to compute, you get - the company claims - up to 100 times lower energy consumption and roughly 25 times the computing density of a leading GPU.
Those are big numbers, and the correct response to big numbers from a pre-product startup is a raised eyebrow. But the eyebrow should come down a little when you learn that the underlying physics was demonstrated and peer-reviewed before the company existed. The core efficiency result ran in Nature Photonics in 2023. The theoretical proposal goes back to a 2019 paper in Physical Review X. Opticore is, in a sense, the commercialization arm of a decade of optics research that happened to finish just as generative AI made everyone desperate for it.
The era of computing with light has arrived - processing billions of parameters in a single chip using 100-fold lower energy.— Zaijun Chen & Ryan Hamerly, Co-Founders
An optical processing unit that performs GPU-class computation using light and waveguides instead of electrical components - the product the whole company is built around.
A patented trick that reuses the same optical hardware across time slices, so a single chip can encode as many as a trillion parameters at roughly 100 TOPS/W.
Silicon photonics stitched to optoelectronic packaging and high-bandwidth memory, aimed squarely at the electronic "memory wall" that throttles AI training.
The headline figure is a ~100x energy-efficiency improvement over leading GPUs. It's a company claim, drawn from lab demonstrations, not an independent benchmark of a shipping product - so read the chart as the thesis Opticore is selling, not a verdict.
Opticore is the kind of company that starts in a lab and hires a business plan later. Its founders come out of MIT, USC and UC Berkeley photonics research, with an advisory line back to MIT's Prof. Dirk Englund.
PhD in physics (Max Planck / LMU Munich), former MIT quantum-photonics postdoc and USC research assistant professor. Awarded the 2023 Optica Foundation Challenge and SPIE AI/ML awards.
Photonics researcher with roots in the MIT work on optical neural networks and photonic AI accelerators that seeded the company's core technology.
Professor and integrated-photonics researcher whose academic work underpins Opticore's chip design and fabrication approach.
Core theoretical proposal published in Physical Review X - years before the AI boom made it commercially interesting.
The ~100x efficiency and area-density result appears in Nature Photonics; the team receives a DARPA NaPSAC award.
Opticore launches publicly with its seed round, led by Sagax Capital, Jetha Global and Bioeconomy.XYZ.
Further validation published in Science Advances; the company adds an INSPIRED award to its research resume.
Raises an additional $7.5M - co-led by Jetha Global and Origin Ventures - bringing total funding to roughly $14.5M.
The target customer is not you and it's not your laptop. It's the data center - specifically the racks that train and run large language models, where electricity bills and cooling have quietly become the dominant cost of doing AI. Opticore's pitch to those operators is simple arithmetic: if a chip can do the same work at a fraction of the energy, the economics of large-scale AI change, and so does the environmental footprint.
That framing is why it's fair to call Opticore a semiconductor company and a climate company at the same time. The mission statement leans hard into the second reading - "scalable, energy-efficient computing solutions that address the global energy and environmental challenges posed by AI." When a technology's main selling point is that it uses 100x less power, the line between chip startup and energy startup gets blurry, and that's clearly by design.
The competitive neighborhood is crowded and well-funded. Lightmatter has raised hundreds of millions; Ayar Labs, Lightelligence and Celestial AI are all working adjacent corners of the optical-compute problem. Opticore's differentiator, as it tells the story, is an end-to-end OPU aimed at memory-bound work rather than just optical interconnects between conventional chips.
It is also, notably, small. Roughly 16 people and about $14.5M raised is a slingshot compared to some rivals' war chests. The bet baked into that number is a very deep-tech bet: that in photonics, defensible physics and patents matter more than headcount, and that a tight team sitting on a decade of peer-reviewed results can move faster than its funding suggests.