Twenty years ago, a Duke University professor, David R. Smith, used artificial composite materials called “metamaterials” to make a real-life invisibility cloak. While this cloak didn’t really work like Harry Potter’s, exhibiting limited ability to conceal objects from the light of a single microwave length, those advances in material science did eventually trickle down to electromagnetism research.
Today, Austin-based Neurophos, a photonics startup spun out of Duke University and Metacept (an incubator run by Smith), is taking that research further to solve what may be the biggest problem facing AI labs and hyperscalers: how to scale computing power while keeping power consumption in check.
The startup has come up with a “metasurface modulator” with optical properties that enable it to serve as a tensor core processor for doing matrix vector multiplication — math that is at the heart of a lot of AI work (particularly inferencing), currently performed by specialized GPUs and TPUs that use traditional silicon gates and transistors. By fitting thousands of these modulators on a chip, Neurophos claims, its “optical processing unit” is significantly faster than the silicon GPUs currently used en masse at AI data centers, and far more efficient at inferencing (running trained models), which can be a fairly expensive task.
To fund the development of its chips, Neurophos has just raised $110 million in a Series A round led by Gates Frontier (Bill Gates’ venture firm), with participation from Microsoft’s M12, Carbon Direct, Aramco Ventures, Bosch Ventures, Tectonic Ventures, Space Capital, and others.
Now, photonic chips are nothing new. In theory, photonic chips offer higher performance than traditional silicon because light produces less heat than electricity, it can travel faster, and is far less susceptible to changes in temperature and electromagnetic fields.
But optical components tend to be much larger than their silicon counterparts, and can be difficult to mass-produce. And they also need converters to transform data from digital to analog and back, which can be large and take up a lot of power.
Neurophos, however, posits that the metasurface it has developed can solve all of those problems in one swoop because it is about “10,000 times” smaller than traditional optical transistors. The small size, the startup claims, enables it to fit thousands of units on a chip, which results in far more efficiency than traditional silicon because the chip can do many more calculations at once.
“When you shrink the optical transistor, you can do way more math in the optics domain before you have to do that conversion back to the electronics domain,” Dr. Patrick Bowen, CEO and co-founder of Neurophos, told TechCrunch. “If you want to go fast, you have to solve the energy efficiency problem first. Because if you’re going to take a chip and make it 100 times faster, it burns 100 times more power. So you get the privilege of going fast after you solve the energy efficiency problem.”
The result, Neurophos claims, is an optical processing unit that can wildly outperform Nvidia’s B200 AI GPU. The startup says its chip can run at 56 GHz, yielding a peak 235 Peta Operations per Second (POPS) and consuming 675 watts, compared to the B200, which can deliver 9 POPS at 1,000 watts.
Bowen says Neurophos has already signed multiple customers (though he declined to name any), and companies including Microsoft are “looking very closely” at the startup’s products.
Still, the startup is entering a crowded market that’s dominated by Nvidia, the world’s most valuable public company, whose products have more or less underpinned the entire AI boom. There are also other companies working on photonics, though some, like Lighmatter, have pivoted to focusing on interconnects. And Neurophos is still a few years away from production, expecting its first chips to hit the market by mid-2028.
But Bowen is confident the performance and efficiency advances of the metasurface will prove a sufficient moat.
“What everyone else is doing is, and this includes Nvidia, in terms of the fundamental physics of the silicon, it’s really evolutionary rather than revolutionary, and it’s tied to the progress of TSMC. If you look at the improvement of TSMC nodes, on average, they improve in energy efficiency about 15%, and that takes a couple years,” he said.
“Even if we chart out Nvidia’s improvement in architecture over the years, by the time we come out in 2028, we still have massive advantages over everyone else in the market because we’re starting with a 50x over Blackwell in both energy efficiency and raw speed.”
And to address the manufacturing issue optical chips have traditionally faced, Neurophos says its chips can be made with standard silicon foundry materials, tools, and processes.
The fresh funding will be used for the development of the company’s first integrated photonic compute system, including datacenter-ready OPU modules, a full software stack, and early-access developer hardware. The company is also opening a San Francisco engineering site and expanding its HQ in Austin, Texas.
“Modern AI inference demands monumental amounts of power and compute,” Dr. Marc Tremblay, corporate vice president and technical fellow of core AI infrastructure at Microsoft, said in a statement. “We need a breakthrough in compute on par with the leaps we’ve seen in AI models themselves, which is what Neurophos’ technology and high-talent density team is developing.”
Originally published at TechCrunch
