Neurophos Secures $110 Million to Create Powerful Tiny Optical Processors

Austin: A professor from Duke University, David R. Smith, created a type of invisibility cloak twenty years ago using special materials called metamaterials. These materials didn’t work like the ones in movies, but they helped lead to advancements in how we understand electromagnetism.

Now, a new Austin-based startup called Neurophos is building on that research. They are trying to tackle a big issue in the AI world: how to improve computing power without using too much energy. Their solution is something called a “metasurface modulator.” This device can work like a special processor for math used in AI, called matrix vector multiplication. Currently, this math is mostly done by GPUs and TPUs, which use traditional bits of technology made from silicon.

Neurophos believes their special modulator can fit thousands of them on a chip, making it much faster and more efficient than the usual silicon GPUs found in AI data centers. To help pay for these chips, Neurophos raised $110 million from various large investments, including Gates Frontier, which is run by Bill Gates.

Even though photonic chips, or those made from light, aren’t new, they offer many benefits. They can perform better than silicon chips because light doesn’t generate as much heat and works faster. However, these optical parts are often larger and harder to produce in large numbers. They also need converters, which can take a lot of power.

Neurophos says they have invented a tiny metasurface that is about 10,000 times smaller than regular optical transistors. This small size allows many calculations to happen at once, improving efficiency. Dr. Patrick Bowen, the CEO of Neurophos, explained that smaller optical transistors let them do more math quickly.

According to their claims, their optical processing unit can perform much better than Nvidia’s B200 AI GPU. They say their chip can operate at 56 GHz, achieving a peak of 235 Peta Operations per Second while only using 675 watts. In comparison, the B200 GPU can achieve 9 POPS at 1,000 watts.

Bowen mentioned that Neurophos has already found several customers interested in their products. Companies like Microsoft are paying close attention too. However, the startup will enter a tough market, mostly led by Nvidia, which has been essential in the AI field. Other companies are also exploring photonics, but Neurophos plans to start selling its chips by mid-2028.

Bowen feels confident about the advantages of their new modulators. He believes that while others, including Nvidia, are making small improvements to silicon technology, Neurophos is creating something genuinely innovative.

To make their chips, Neurophos is using regular materials and equipment from silicon manufacturing. This approach could help solve the production challenges faced by optical chips. The funding from investors will support the creation of their photonic computing system, which will include components ready for data centers.

Dr. Marc Tremblay from Microsoft said that modern AI needs a lot of power and computing strength. He believes Neurophos has the potential to create methods of computing that match the impressive improvements seen in AI models.