A small semiconductor startup in Austin, Texas has gained major funding based on a simple but radical premise: the fastest way to scale artificial intelligence may be to stop relying on electrons altogether. Neurophos, an AI chip company backed by Bill Gates’ Gates Frontier, says it has developed a new class of optical processors that could sharply reduce the energy required to run advanced AI models.
Founded in 2020, the company is building what it calls an optical processing unit, or OPU, that performs the core math behind AI workloads using light rather than electricity. By replacing conventional electronic transistors with micron-scale photonic components, Neurophos claims it can deliver far greater compute density while avoiding the heat and power constraints that increasingly limit modern GPUs.
One Extremely Large Core
Neurophos’ architecture centers on a single, unusually large optical matrix. According to the company, each chip contains one photonic tensor core measuring 1,000 by 1,000 processing elements. That is roughly 15 times larger than the matrix engines typically used in today’s AI accelerators. While most GPUs rely on hundreds of smaller tensor cores working in parallel, Neurophos is pursuing a different approach: one extremely large core operating at very high speed.
That speed is central to the company’s performance claims. The optical tensor core is designed to run at approximately 56 gigahertz, far above the clock speeds seen in CPUs and GPUs. Because the matrix multiplication itself happens optically, the main sources of power consumption are the conversions between electronic and optical signals. Neurophos says this allows its chip to deliver up to 470 peta-operations per second in AI workloads while consuming power comparable to leading accelerators from NVIDIA.
Those claims arrive as the AI industry confronts a mounting energy problem. Large language models are extraordinarily power-hungry, pushing data centers toward higher electricity consumption and rising operating costs. Many chipmakers are now exploring alternatives to traditional scaling as Moore’s Law slows and efficiency gains become harder to extract from conventional silicon.
To work around those limits, Neurophos’s photonic transistors, which it calls metamaterial optical modulators, are roughly 10,000 times smaller than the photonic components typically produced in silicon photonics fabs today. That size reduction, the company says, makes it possible to pack more than one million optical processing elements onto a single chip.
Could be Produced by Established Foundries
Neurophos claims it has created its new design to fit standard CMOS manufacturing techniques. That compatibility means the chips could eventually be produced by established foundries such as Intel or TSMC, rather than requiring specialized fabrication processes.
Even so, commercialization remains a stretch. The company does not expect volume production until around 2028, and initial deployments are likely to be limited in scale.
The startup recently closed a $110 million Series A funding round led by Gates Frontier, with participation from Microsoft’s venture fund and several strategic investors. Neurophos plans to use the capital to complete a proof-of-concept chip and deliver early-access hardware to select partners.
In its first incarnation, the OPU is expected to target AI inference rather than training, particularly the compute-intensive prefill phase of large language model inference. That strategy is in keeping with a broader industry shift toward splitting up AI workloads, pairing specialized accelerators with conventional GPUs to improve efficiency.
