Neurophos Secures $110 Million to Develop Compact Optical AI Processors

Two decades ago, Duke University professor David R. Smith created a real-world “invisibility cloak” using artificial composite materials known as metamaterials. Unlike its fictional counterpart in Harry Potter, this prototype had limited functionality, only concealing objects from a specific microwave wavelength. However, these foundational advances in material science paved the way for subsequent breakthroughs in electromagnetism research.

Today, Austin-based startup Neurophos—a photonics company spun out from Duke University and the Metacept incubator founded by Smith—is pushing this research forward. Their goal is to tackle one of the most critical challenges for AI labs and hyperscale data centers: scaling computational power without a corresponding surge in energy consumption.

The company has developed a “metasurface modulator” with unique optical properties, enabling it to function as a tensor core processor. It specializes in matrix vector multiplication, the core mathematical operation behind much of today’s AI work, particularly inference. Currently, this task is handled by specialized silicon-based GPUs and TPUs. By densely packing thousands of these modulators onto a single chip, Neurophos claims its “Optical Processing Unit” (OPU) achieves significantly higher speeds and far greater energy efficiency for AI inference compared to the mainstream GPUs powering data centers today.

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

The concept of photonic computing is not new. In theory, chips that use light instead of electricity promise superior performance because light generates less heat, travels faster, and is less affected by temperature fluctuations and electromagnetic interference.

However, practical challenges have persisted. Optical components are typically much larger than silicon transistors and are difficult to manufacture at scale. Furthermore, they require data converters to switch between digital and analog signals, components that are often bulky and power-hungry.

Neurophos asserts that its proprietary metasurface technology solves these problems simultaneously. The key is the component’s dramatically reduced size—reportedly about 10,000 times smaller than conventional optical transistors. This miniaturization allows thousands of units to be integrated onto a single chip, enabling massively parallel calculations and delivering efficiency leaps over traditional silicon.

Disrupt 2026: Limited-Time Ticket Offer

Tickets are now available! Secure your spot and save up to $680 with our exclusive launch rates. Be among the first 500 registrants to receive 50% off a +1 guest pass. TechCrunch Disrupt gathers top executives from Google Cloud, Netflix, Microsoft, Box, a16z, Hugging Face, and many others for over 250 sessions aimed at accelerating growth and honing your competitive advantage. Network with hundreds of groundbreaking startups and participate in curated sessions designed to spark deals, provide key insights, and deliver inspiration.

Disrupt 2026: Limited-Time Ticket Offer

Tickets are now available! Secure your spot and save up to $680 with our exclusive launch rates. Be among the first 500 registrants to receive 50% off a +1 guest pass. TechCrunch Disrupt gathers top executives from Google Cloud, Netflix, Microsoft, Box, a16z, Hugging Face, and many others for over 250 sessions aimed at accelerating growth and honing your competitive advantage. Network with hundreds of groundbreaking startups and participate in curated sessions designed to spark deals, provide key insights, and deliver inspiration.

“By shrinking the optical transistor, we can perform vastly more computations in the optical domain before needing to convert data back to the electronic domain,” explained Dr. Patrick Bowen, CEO and co-founder of Neurophos, in an interview with TechCrunch. “To achieve high speed, you must first solve the energy efficiency challenge. Making a chip 100 times faster typically means it consumes 100 times more power. True speed is only attainable once you’ve drastically improved efficiency.”

According to Neurophos, the outcome is an optical processor that dramatically outperforms current leaders like NVIDIA’s B200 AI GPU. The startup states its chip operates at 56 GHz, delivering a peak performance of 235 Peta Operations Per Second (POPS) while drawing 675 watts of power. In comparison, they claim the B200 delivers 9 POPS at 1,000 watts.

Bowen revealed that Neurophos has already secured multiple design-win customers (though their names remain confidential) and that companies like Microsoft are evaluating the startup’s technology with serious interest.

Nevertheless, Neurophos is entering a fiercely competitive market dominated by NVIDIA, whose silicon GPUs have become the backbone of the modern AI boom. Other companies are also exploring photonic computing, with some, like Lightmatter, shifting focus to optical interconnects. Neurophos itself is still in development, targeting the first commercial shipments of its chips for mid-2028.

Despite the timeline, Bowen is confident that the performance and efficiency gains from the company’s metasurface modulators will create a substantial competitive barrier.

“The approach of others, including NVIDIA, involves incremental, evolutionary improvements tied to the progress of silicon foundries like TSMC. On average, a new TSMC process node improves energy efficiency by about 15% every couple of years,” he said.

“Even when we project NVIDIA’s architectural improvements forward to our 2028 launch date, we maintain a massive advantage. We are starting from a baseline of being approximately 50 times more energy-efficient and faster than the current Blackwell architecture in both raw speed and efficiency.”

To overcome the historical manufacturing hurdles of photonic chips, Neurophos states its design is compatible with standard silicon foundry materials, tools, and processes, enabling cost-effective volume production.

The new capital will accelerate development of the company’s first integrated photonic computing system. This includes data center-ready OPU modules, a complete software stack, and early-access hardware for developers. The funding will also support the opening of a new engineering site in San Francisco and the expansion of its headquarters in Austin, Texas.

“Modern AI inference requires staggering amounts of power and compute resources,” said Dr. Marc Tremblay, Corporate Vice President and Technical Fellow of Core AI Infrastructure at Microsoft. “We need a computational breakthrough on the same scale as the leaps we've seen in AI models. This is precisely what Neurophos’s technology and their deeply talented team are building.”