Neurophos Raises $110M to Build Optical AI Chips That Challenge Nvidia's GPU Dominance

Neurophos Secures $110 Million Funding to Revolutionize AI Hardware

Neurophos, an Austin, Texas-based AI chip startup, has closed a $110 million Series A funding round led by Gates Frontier, Bill Gates' venture firm, with participation from Microsoft's M12, Aramco Ventures, Bosch Ventures, and others 1

1

. The company, spun out of Duke University and Metacept, aims to tackle what may be the biggest challenge facing AI labs and hyperscalers: scaling computing power while keeping energy consumption in check 1

1

. This brings the startup's total funding to $118 million to date 5

5

.

Optical Processing Unit Built on Breakthrough Photonics Technology

At the heart of Neurophos' innovation is its optical processing unit, which integrates more than one million micron-scale optical processing elements on a single chip 5

5

. The company has developed proprietary metasurface modulators with optical properties that enable the chip to serve as a tensor core processor for matrix vector multiplication—math at the heart of AI inference work currently performed by specialized GPUs and TPUs 1

1

. These optical transistors are approximately 10,000 times smaller than traditional optical components available from silicon photonics factories today 2

2

3

3

.

CEO and co-founder Patrick Bowen explained the significance: "When you shrink the optical transistor, you can do way more math in the optics domain before you have to do that conversion back to the electronics domain" 1

1

. The startup claims its approach addresses the fundamental challenge that has plagued photonic computing: optical components have traditionally been much larger than silicon counterparts and difficult to mass-produce.

Replace Electrons With Photons for Massive Performance Gains

Neurophos' first-generation chip, the Tulkas T100, features a single photonic tensor core measuring 1,000 by 1,000 processing elements—roughly 15 times larger than the typical 256 x 256 matrix used in most AI GPUs 2

2

3

3

. This optical AI chip operates at 56 GHz, yielding a peak 235 Peta Operations per Second (POPS) and consuming 675 watts, compared to Nvidia's B200, which delivers 9 POPS at 1,000 watts 1

1

. The company claims its dual reticle design will be capable of 470 petaFLOPS of FP4/INT4 compute—about 10 times that of Nvidia's newly unveiled Rubin GPUs—while using roughly the same amount of power 3

3

.

The chip uses photons instead of electrons to transmit data, bypassing the speed and energy limitations inherent in traditional electrical systems. Early tests have demonstrated performance exceeding 300 trillion operations per second per watt, far surpassing existing standards for energy-efficient AI hardware 5

5

.

Challenging Moore's Law and Nvidia's Market Dominance

Bowen argues that the AI industry cannot afford to wait for Moore's Law to keep pace with demand. "What everyone else is doing, including Nvidia, in terms of the fundamental physics of the silicon, it's really evolutionary rather than revolutionary," he stated 1

1

. He noted that TSMC node improvements average about 15% in energy efficiency every couple of years, while Neurophos claims a 50x advantage over Nvidia's Blackwell architecture in both energy efficiency and raw speed by its 2028 launch 1

1

.

The startup enters a market dominated by Nvidia, the world's most valuable public company with a market capitalization exceeding $4 trillion . Other competitors like Lightmatter have pivoted to focusing on interconnects, while companies like Cerebras Systems are reportedly raising funds at a $22 billion valuation .

Path to Production and Early Customer Adoption

Neurophos expects its first chips to hit the market by mid-2028, with the Tulkas T100 equipped with 768 GB of HBM consuming 1 to 2 kilowatts of power under load 3

3

. The company is actively working on a proof-of-concept chip to validate its claimed compute density and power efficiency 3

3

. Initial production volumes are expected to be in the thousands rather than tens of thousands of chips 3

3

.

The startup envisions its first chip being used primarily as a prefill processor for LLM inference—the compute-intensive stage where input tokens are processed 3

3

. Bowen suggests pairing one rack of 256 Neurophos chips with something like an NVL576 rack from Nvidia 3

3

. The company has already signed multiple customers and says Microsoft is "looking very closely" at its products 1

1

. The startup is also partnering with Norwegian data center operator Terakraft to launch a real-world pilot of its optical AI accelerator in 2027 5

5

.

Manufacturing Advantage Through Standard Processes

A critical advantage for Neurophos is that its chips can be manufactured using standard silicon foundry materials, tools, and processes 1

1

. Bowen confirmed the company received its first silicon back in May, demonstrating compatibility with standard CMOS processes and existing foundry technologies 3

3

. This addresses a historical challenge with photonic chips, which have been difficult to mass-produce. The company could potentially tap fabs like Intel or TSMC for production 2

2

.

Microsoft's Marc Tremblay, corporate vice president and technical fellow of core AI infrastructure, noted: "Modern AI inference demands monumental amounts of power and compute. We need a breakthrough in compute on par with the leaps we've seen in AI models themselves" 5

5

. The fresh funding will accelerate development of the company's first integrated photonic compute system, including datacenter-ready OPU modules and a full software stack 1

1

.