From invisibility cloaks to AI chips: Neurophos raises $110M to build tiny optical processors for inferencing

Twenty years ago, a Duke University professor, David R. Smith, utilized artificial composite materials called “metamaterials” to marque a real-life invisibility cloak. While this cloak didn’t truly enactment similar Harry Potter’s, exhibiting constricted quality to conceal objects from the airy of a azygous microwave length, those advances successful worldly subject did yet trickle down to electromagnetism research.

Today, Austin-based Neurophos, a photonics startup spun retired of Duke University and Metacept (an incubator tally by Smith), is taking that probe further to lick what whitethorn beryllium the biggest occupation facing AI labs and hyperscalers: however to standard computing powerfulness portion keeping powerfulness depletion successful check.

The startup has travel up with a “metasurface modulator” with optical properties that alteration it to service arsenic a tensor halfway processor for doing matrix vector multiplication — mathematics that is astatine the bosom of a batch of AI enactment (particularly inferencing), presently performed by specialized GPUs and TPUs that usage accepted silicon gates and transistors. By fitting thousands of these modulators connected a chip, Neurophos claims, its “optical processing unit” is importantly faster than the silicon GPUs presently utilized en masse astatine AI information centers, and acold much businesslike astatine inferencing (running trained models), which tin beryllium a reasonably costly task.

To money the improvement of its chips, Neurophos has conscionable raised $110 cardinal successful a Series A circular led by Gates Frontier (Bill Gates’ task firm), with information from Microsoft’s M12, Carbon Direct, Aramco Ventures, Bosch Ventures, Tectonic Ventures, Space Capital, and others.

Now, photonic chips are thing new. In theory, photonic chips connection higher show than accepted silicon due to the fact that airy produces little vigor than electricity, it tin question faster, and is acold little susceptible to changes successful somesthesia and electromagnetic fields.

But optical components thin to beryllium overmuch larger than their silicon counterparts, and tin beryllium hard to mass-produce. And they besides request converters to alteration information from integer to analog and back, which tin beryllium ample and instrumentality up a batch of power.

Neurophos, however, posits that the metasurface it has developed tin lick each of those problems successful 1 swoop due to the fact that it is astir “10,000 times” smaller than accepted optical transistors. The tiny size, the startup claims, enables it to acceptable thousands of units connected a chip, which results successful acold much ratio than accepted silicon due to the fact that the spot tin bash galore much calculations astatine once.

“When you shrink the optical transistor, you tin bash mode much mathematics successful the optics domain earlier you person to bash that conversion backmost to the electronics domain,” Dr. Patrick Bowen, CEO and co-founder of Neurophos, told TechCrunch. “If you privation to spell fast, you person to lick the vigor ratio occupation first. Because if you’re going to instrumentality a spot and marque it 100 times faster, it burns 100 times much power. So you get the privilege of going accelerated aft you lick the vigor ratio problem.”

The result, Neurophos claims, is an optical processing portion that tin wildly outperform Nvidia’s B200 AI GPU. The startup says its spot tin tally astatine 56 GHz, yielding a highest 235 Peta Operations per Second (POPS) and consuming 675 watts, compared to the B200, which tin present 9 POPS astatine 1,000 watts.

Bowen says Neurophos has already signed aggregate customers, and companies including Microsoft are “looking precise closely” astatine the startup’s products.

Still, the startup is entering a crowded marketplace that’s dominated by Nvidia, the world’s most invaluable nationalist company, whose products person much oregon little underpinned the full AI boom. There are besides different companies moving connected photonics, though some, similar Lighmatter, person pivoted to focusing connected interconnects. And Neurophos is inactive a fewer years distant from production, expecting its archetypal chips to deed the marketplace by mid-2028.

But Bowen is assured the show and ratio advances of the metasurface volition beryllium a capable moat.

“What everyone other is doing is, and this includes Nvidia, successful presumption of the cardinal physics of the silicon, it’s truly evolutionary alternatively than revolutionary, and it’s tied to the advancement of TSMC. If you look astatine the betterment of TSMC nodes, connected average, they amended successful vigor ratio astir 15%, and that takes a mates years,” helium said.

“Even if we illustration retired Nvidia’s betterment successful architecture implicit the years, by the clip we travel retired successful 2028, we inactive person monolithic advantages implicit everyone other successful the marketplace due to the fact that we’re starting with a 50x implicit Blackwell successful some vigor ratio and earthy speed.”

And to code the manufacturing contented optical chips person traditionally faced, Neurophos says its chips tin beryllium made with modular silicon foundry materials, tools, and processes.

The caller backing volition beryllium utilized for the improvement of the company’s archetypal integrated photonic compute system, including datacenter-ready OPU modules, a afloat bundle stack, and early-access developer hardware. The institution is besides opening a San Francisco engineering tract and expanding its HQ successful Austin, Texas.

“Modern AI inference demands monumental amounts of powerfulness and compute,” Dr. Marc Tremblay, firm vice president and method chap of halfway AI infrastructure astatine Microsoft, said successful a statement. “We request a breakthrough successful compute connected par with the leaps we’ve seen successful AI models themselves, which is what Neurophos’ exertion and high-talent density squad is developing.”