Two decades ago, David R. Smith, a professor at Duke University, created a real-world invisibility cloak using artificial composite materials known as "metamaterials." Although this early version could only partially hide objects from a single microwave frequency—far from the magic of fiction—the foundational research gradually influenced broader electromagnetism studies. Now, Neurophos, a photonics startup originating from Duke and the Metacept incubator led by Smith, is building on that legacy to address a critical challenge for AI developers and large-scale data centers: expanding computational capacity without a corresponding surge in energy use.

The company has developed a "metasurface modulator" whose optical characteristics allow it to function as a tensor core processor, executing matrix vector multiplication. This mathematical operation is central to many AI tasks, especially inference, which is currently handled by specialized GPUs and TPUs relying on conventional silicon-based transistors. Neurophos states that by integrating thousands of these modulators into a single chip, its "optical processing unit" operates much faster and with greater efficiency during inference—a typically resource-intensive process—than the silicon GPUs widely deployed in AI data centers.

To advance its chip development, Neurophos has secured $110 million in a Series A funding round. The investment was led by Gates Frontier, the venture firm founded by Bill Gates, and included participation from Microsoft’s M12, Carbon Direct, Aramco Ventures, Bosch Ventures, Tectonic Ventures, Space Capital, and other backers.

Photonic computing is not a novel concept. In principle, chips that use light instead of electricity can achieve higher performance, as light generates less heat, moves faster, and is less affected by temperature fluctuations and electromagnetic interference. However, optical components have historically been bulkier than silicon parts, posed manufacturing challenges at scale, and required power-hungry digital-to-analog converters. Neurophos argues that its proprietary metasurface overcomes these hurdles simultaneously, being roughly "10,000 times" smaller than traditional optical transistors.

This miniaturization allows thousands of units to be packed onto a chip, dramatically boosting efficiency by enabling vastly more simultaneous calculations. Dr. Bowen of Neurophos explained, "When you shrink the optical transistor, you can do way more math in the optics domain before you have to convert back to the electronics domain." He emphasized, "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 power. So you earn the privilege of speed only after tackling efficiency."

According to the startup, the outcome is an optical processing unit that significantly outpaces Nvidia’s B200 AI GPU. Neurophos claims its chip operates at 56 GHz, delivering a peak performance of 235 Peta Operations Per Second (POPS) while consuming 675 watts. In comparison, the B200 is rated for 9 POPS at 1,000 watts. Bowen noted that Neurophos has already secured several customers—though none were named—and that companies including Microsoft are evaluating its technology closely.

Nevertheless, the company is stepping into a competitive arena dominated by Nvidia, whose hardware has largely fueled the current AI expansion. Other firms are also exploring photonics, with some, like Lighmatter, shifting focus toward interconnect solutions. Neurophos does not anticipate having production-ready chips until mid-2028. Bowen remains confident, however, that the performance and efficiency gains from its metasurface modulators will create a durable competitive edge.

"What others, including Nvidia, are doing with silicon physics is evolutionary, not revolutionary, and it’s tied to the pace of TSMC’s manufacturing advances," he observed. "TSMC’s nodes typically improve energy efficiency by about 15% every couple of years. Even projecting Nvidia’s architectural progress, by our 2028 launch we expect to maintain a massive advantage—starting with a 50x lead over Blackwell in both energy efficiency and raw speed."

To circumvent the production difficulties often associated with optical chips, Neurophos says its design can be fabricated using standard silicon foundry materials, tools, and processes. The new capital will support development of its first integrated photonic computing system, which will include data-center-ready OPU modules, a complete software stack, and early-access hardware for developers. The company is also expanding its engineering presence with a new site in San Francisco and growing its headquarters in Austin, Texas.

Dr. Marc Tremblay, Corporate Vice President and Technical Fellow of Core AI Infrastructure at Microsoft, commented, "Modern AI inference requires monumental amounts of power and compute. We need a breakthrough in computing that matches the leaps we’ve seen in AI models, which is exactly what Neurophos’ technology and highly skilled team are building."