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

Twenty years ago, David R. Smith, a professor at Duke University, used synthetic composite materials called “metamaterials” to make a real invisibility cloak. While this cloak didn’t really work like Harry Potter’s, as it showed a limited ability to hide objects from light of a single microwave length, those advances in materials science eventually crossed over into electromagnetic research.

Today, Neurophos, an Austin-based photonics startup that spun out of Duke University and Metacept (an incubator run by Smith), is advancing this research to solve what may be the biggest problem facing AI and hyperscale labs: how to scale computing power while keeping power consumption in check.

The startup has come up with a “supersurface modulator” with optical properties that enable it to act as a basic tensor processor to do matrix vector multiplication — the mathematics that is at the heart of much AI work (particularly inference), and which is currently performed by specialized GPUs and TPUs that use traditional silicon gates and transistors. By fitting thousands of these modulators onto a chip, Neurovos claims, its “optical processing unit” is much faster than the silicon GPUs currently used en masse in AI data centers, and much more efficient at inference (running trained models), which can be a fairly expensive task.

To fund the development of its chips, Neurovos has just raised $110 million in a Series A round led by Gates Frontier (Bill Gates’ venture firm), with participation from Microsoft M12, Carbon Direct, Aramco Ventures, Bosch Ventures, Tectonic Ventures, Space Capital, and others.

Now, optical chips are not new. In theory, photonic chips offer higher performance than traditional silicon because light produces less heat than electricity, can travel faster, and is less susceptible to temperature changes and electromagnetic fields.

But photonic components tend to be much larger than their silicon counterparts, and can be difficult to produce in large quantities. They also need converters to convert data from digital to analog and back, which can be large and consume a lot of power.

However, Neurofoss posits that its metasurface could solve all these problems in one fell swoop because it is about “10,000 times” smaller than conventional phototransistors. The startup claims that the small size enables it to fit thousands of modules onto the chip, which results in much greater efficiency than traditional silicon because the chip can perform many calculations simultaneously.

“When you shrink an optical transistor, you can do more calculations in the optics before you have to make that conversion back to the electronics,” 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 take a chip and make it 100 times faster, it burns 100 times more energy. So you get the privilege of going fast after you solve the energy efficiency problem.”

The result, Neurophos claims, is a visual processing unit that can significantly outperform Nvidia’s B200 AI GPU. The startup says its chip can operate at 56 GHz, producing a peak of 235 beta operations per second (POPS) and consumes 675 watts, compared to the B200, which can deliver 9 POPS at 1,000 watts.

Neurophos has already signed up several customers (though he declined to name any), Bowen says, and companies including Microsoft are “looking closely” at the startup’s products.

However, the startup is entering a crowded market dominated by Nvidia, the world’s most valuable public company, whose products have more or less fueled the entire AI boom. There are also other companies working in photonics, although some, such as Lighmatter, have focused on interconnects. Neurophos is still a few years away from production, and its first chips are expected to hit the market by mid-2028.

But Bowen is confident that the advances in performance and efficiency offered by its metal roof modifiers will prove a sufficient moat.

“What everyone is doing, and this includes Nvidia, in terms of the fundamental physics of silicon, is actually evolutionary rather than revolutionary, and it’s tied to TSMC’s progress. If you look at the optimization of TSMC nodes, on average, they improve in power efficiency by about 15%, and that takes a few years.”

“Even if we map out the improvement Nvidia has made in architecture over the years, by the time we exit in 2028, we still have huge advantages over everyone else in the market because we start with a 50x advantage over Blackwell in both power efficiency and raw speed.”

To address the mass manufacturing issues that photonic chips have traditionally faced, Neurofos says its chips can be manufactured using standard silicon foundry materials, tools, and processes.

The new funding will be used to develop the company’s first all-optical computing system, including data center-ready OPUs, a full suite of software, and early access developer hardware. The company is also opening an engineering site in San Francisco and expanding its headquarters in Austin, Texas.

“Modern AI inference requires enormous amounts of power and computing,” Dr. Marc Tremblay, corporate vice president and technical fellow for core AI infrastructure at Microsoft, said in a statement. “We need a breakthrough in computing on par with the leaps we’ve seen in AI models themselves, which is what Neurofoss’s technology and high-caliber talent team are working to advance.”