Neurophos Raises $110M to Build Optical AI Chips for Inferencing
▼ Summary
– Twenty years ago, Duke University research on metamaterials for invisibility cloaks laid the groundwork for today’s photonic computing advances.
– Neurophos, a startup from that research, is developing an optical processing unit using a “metasurface modulator” to perform AI matrix math more efficiently than silicon chips.
– The company claims its chip is far faster and more energy-efficient for AI inferencing than current GPUs, citing a 56 GHz speed and 235 POPS at 675 watts compared to Nvidia’s B200.
– Neurophos has raised $110 million in Series A funding, led by Gates Frontier, to develop its technology, with its first chips expected to hit the market by mid-2028.
– While entering a market dominated by Nvidia, Neurophos believes its revolutionary optical approach and use of standard silicon manufacturing processes provide a significant performance and efficiency advantage.
A new wave of computing technology is emerging to tackle the immense energy demands of artificial intelligence. Neurophos, an Austin-based photonics startup, has secured $110 million in Series A funding to advance its optical AI chips, aiming to deliver a revolutionary leap in processing speed and efficiency for AI inferencing tasks. The round was led by Gates Frontier, with participation from Microsoft’s M12, Aramco Ventures, and Bosch Ventures, among others.
The company’s origins trace back to pioneering metamaterials research at Duke University. Two decades ago, Professor David R. Smith demonstrated an early “invisibility cloak” using these artificial composites. While that application was limited, the underlying science has now evolved into a potential solution for the data center’s most pressing issue: skyrocketing power consumption from running AI models.
At the core of Neurophos’s approach is a “metasurface modulator”, an optical component so small the company claims it is roughly ten thousand times smaller than traditional optical transistors. This device functions as a tensor core processor, performing the matrix vector multiplication that is fundamental to AI work. By densely packing thousands of these modulators onto a single optical processing unit (OPU), Neurophos asserts it can achieve performance far beyond today’s silicon-based GPUs.
“If you want to go fast, you have to solve the energy efficiency problem first,” explains Dr. Patrick Bowen, CEO and co-founder of Neurophos. “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 startup’s bold claims include a chip operating at 56 GHz, delivering a peak 235 Peta Operations Per Second (POPS) while consuming just 675 watts. This performance is positioned as vastly superior to current leaders, such as Nvidia’s B200 GPU, which Neurophos states delivers 9 POPS at 1,000 watts.
While photonic computing promises advantages like faster signal propagation and reduced heat, it has historically faced significant commercialization hurdles. Optical components are typically bulky, difficult to manufacture at scale, and require power-hungry digital-to-analog converters. Neurophos argues its ultra-compact metasurface technology overcomes these barriers by allowing vastly more calculations to be performed in the optical domain before conversion is needed.
Furthermore, the company states its chips can be fabricated using standard silicon foundry tools and processes, potentially sidestepping the mass-production challenges that have plagued other photonics ventures. The fresh capital will fuel the development of a complete photonic compute system, including datacenter-ready OPU modules, a software stack, and early-access hardware. Neurophos is also expanding its engineering presence in San Francisco and its headquarters in Austin.
The market entry will be challenging, given Nvidia’s dominant position and the existence of other photonics companies, some of which have shifted focus to interconnects. Neurophos does not expect its first chips to reach the market until mid-2028. However, Bowen is confident the fundamental physics of their technology creates a durable advantage.
“What everyone else is doing is, in terms of the fundamental physics of the silicon, it’s really evolutionary rather than revolutionary,” he said, noting that traditional silicon improvements are tied to the incremental progress of foundries like TSMC. “By the time we come out in 2028, we still have massive advantages… because we’re starting with a 50x over [Nvidia’s] Blackwell in both energy efficiency and raw speed.”
The company has already signed multiple undisclosed customers and reports that major players like Microsoft are evaluating its technology. Microsoft’s Dr. Marc Tremblay highlighted the need for a computational breakthrough to match the advances in AI models, stating that Neurophos’s technology is developing precisely that. As AI’s hunger for compute intensifies, the industry is watching to see if light-based processors can finally deliver on their long-held promise.
(Source: TechCrunch)