Computing is typically done on electronic circuits - by moving digital bits through transistors, switches and memory. Its highly efficient, and has faithfully followed Moore’s Law in terms of scalability, size, power and speed.

The problem is that data centers deal with massive quantities of optical data. Traditionally, this data had to be converted to the electrical domain (photons to bits) to enable switching of data into different channels, and to perform compute operations for on-line search, and learning and training of AI data streams. Conversion from optics to electronics and back to optics is highly lossy (OEO), and dramatically increases energy and water consumption in data centers - something that is expensive in terms of money, politics and company image. OCS (Optical Circuit Switching) is making tremendous inroads into minimizing this expensive OEO conversion through the use of silicon photonics and meta-materials for switching optical beams. Computing though needs some help. Neurophos is doing just that. The company uses silicon photonics circuits and meta-materials technology to perform computations in the optical domain.

Optical computing in general is not new.

It started in 1960 when researchers discovered that lenses and optical filters could perform Fourier transforms, a mathematical operation that separates signals or waves into individual frequencies.

1970s: spatial light modulators (SLMs) contributed further by enabling creation of patterns of light, altering how light is focused or even make a laser beam move in a specific pattern, all in a programmable way.

1980 - 2000 : The next 2 decades saw significant academic research and innovations. However, they never made commercial progress - primarily because electronic computing was progressing dramatically, and optics was still confined to the laboratory. SLM technology was adapted to perform vector matrix multiplication (VMM), a critical breakthrough.

2000-2025: The next 25 years witnessed the maturation of optics - optical semiconductors, silicon photonics, meta-materials, fiber-optics and free-space optics for applications in sensing, communications, materials processing and computing. It also saw the rise of hyperscale data centers which consume massive amounts of optical technology and have fueled the growth of multi-billion optics players like Lumentum, Coherent, Corning and IPG Photonics. Data centers now need optics not just to transport and switch data, but also to do things with the data - like compute. A hyperscale data center consumes 100MW-1GW of power - of which ~ 50% is used for compute. Optical computing can reduce energy consumption (and associated cooling needs) dramatically.

Neurophos is at the forefront of making optical computing practical. Based in Austin, Texas, i t recently raised $110M t o accelerate its progress in optical computing. The round was led by Gates Frontier, with participation from M12 (Microsoft’s Venture Fund), Carbon Direct Capital, Aramco Ventures, Bosch Ventures, Tectonic Ventures, Space Capital, and others. This has allowed the company to scale its workforce by 4X in 2026 . According to Dr. Patrick Bowen, CEO and Co-Founder of Neurophos. “Moore’s Law is slowing, but AI can’t afford to wait. Our breakthrough in photonics unlocks an entirely new dimension of scaling, by packing massive optical parallelism on a single chip. This physics-level shift means both efficiency and raw speed improve as we scale up, breaking free from the power walls that constrain traditional GPUs. Optical compute changes how much AI we can deliver from a given amount of power, space, and money".

A few basic points first. Computing in AI during the inference phase is of 2 types - “prefill” and “decode”. Prefill is the first phase of LLM (Large Language Model) inference. The LLM reads and analyzes the entire input prompt in parallel. Decode is where the model writes out the response, although this is highly sequential. Prefill dominates the compute load during inference in AI. There is also compute load associated with training the LLMs on billions of pieces of data. While this does consume a lot of power, it is a transient power draw during the initial training. Training updates consume much lower power.

Figure 1 compares classical optical computing (electrical and GPU based), the problems with it, and how Neurophos solves this through its OCU (Optical Compute Unit) architecture.

As mentioned earlier, in a data-center application, minimizing energy consumption is critical. Currently, computing occurs in the electrical domain - while most of the data transport in Scale Up and Scale Across racks is optical. Compute in the electrical domain with GPU (Graphical Processing Units) chips requires an Optical-to-Electric conversion for compute and then a Electric-Optic conversion for transport. Electron transport through metal conductors is lossy. Additionally, electrical devices Like GPUs use switches throughout the computing core as it performs calculations. Waiting for switches to settle eats into compute time. In the Neurophos optical core, calculations happen instantaneously, at the speed of light, because no switching is required. It performs millions of calculations at once, 56 billion times a second, at a much faster pace than the GPU of equivalent size. Apart from speed, OCUs also consume dramatically less power since no energy to power the switches is needed, and signal propagation losses are minimal. Electrical workflow and power consumption is concentrated at the inputs and outputs, for the electro-optical and opto-electronic conversions. As a result, an OCU can perform 25X more operations per second than a GPU for the same chip size and energy consumption.

Currently, there is no effective memory technology for photons like there is for electrons. Therefore, the the computation has to pause to lock in the intermediate results before moving to the next compute phase, clocking the math from one calculation to the next, discretely. For this reason, SRAM and DRAM have to be used. In the future, if the optical memory issue is resolved, no OEO conversion at the input and output of the OCU will be required, leading to even faster speeds, and lower power consumption, cost and size.

The prefill compute involves a significant amount of matrix multiplication, since the LLMs are creating correlations between the prompt and the trained engine in high dimensional space. In a GPU, a large area of the chip is dedicated to perform these calculations. In the Neurophos chip, the multiplication is achieved through 3 operations:

  1. Data is copied using optical splitters, which splits signals along different paths in a certain ratio. It is a passive component that can we created within a photonics circuit to split incoming photon streams to multiple streams along different waveguides. It uses interference couplers, adiabatic tapers and wavelength sensitive gratings to achieve this.
  2. Multiplication is achieved optically using reflective and transmissive structures that can combine intensity of multiple light beam. The actual optical elements could be interferometers and phase shifters.
  3. Addition is achieved using optical combiners, which are elements that a that merges multiple optical signals into a single output path. It can combine amplitude, wavelengths, phase or polarization properties of individual optical signals.

Optical compute’s primary application is matrix-matrix multiplication. In optics, one matrix is stored in a grid of optical modulators, while the other is sent through as a series of input vectors, with each number in a vector encoded in the amplitude of its own light beam. Copying is done by splitting each beam into about a thousand beams, spread across one row of the modulator grid. Multiplication in optics is done by reflection or transmission: here each copied beam reflects off a modulator, and the reflected light is the incident beam times the value stored in that modulator. Addition is done by focusing the light reflected from each column of the grid onto a single photodetector, which adds up that light into one element of the output. Because the modulators hold their matrix while thousands of input vectors pass through, the energy of loading it is spread across all of them, which is why an optical tensor core is built for matrix-matrix rather than vector-matrix multiplication.

These operations occur on an optical chip (like the one shown in Figure 1), which consists of tiny waveguides etched into materials like silicon, SiN or InP. Waveguides move optical beams within the chip. Laser sources and optical detectors are integrated to enable the OEO conversion. Modulators change different properties of light, such as intensity, phase, polarization, and wavelength, to encode and process information. Apart from modulators, waveguides also integrate splitters, reflective and transmissive surfaces, and combiners.

Neurophos’s AI inference chips use a meta-surface architecture with tunable optical elements. It targets the central constraint in AI infrastructure: manufacturing, powering, cooling, and deploying enough compute for hyperscale inference. Microsoft is engaged deeply and is an investor. Other commercial traction includes hyper-scalers. The company has completed its first compute demonstration with a meta-surface array and is taping out additional devices with improved reliability and performance. The supply-chain front includes leading foundries, custom ASIC designers, optics suppliers, and packaging partners.

A key challenge in optical computing and matrix multiplication is that standard modulators is large, about 2 mm in length. This means fewer modulators can be packed on to a chip of a given size, restricting compute speed and limiting energy efficiency (compute operation/W). An array of a thousand modulators on a side does a million multiplications and a million additions for only two thousand conversions (a thousand in and a thousand out), so each conversion is amortized over about a thousand operations. Neurophos secret sauce is the use of metamaterials to build modulators 10,000 times smaller in area than today’s state of the art, which puts a million of them on a chip the size of a fingernail instead of one the size of a card table. That makes a thousand-by-thousand array practical on a single chip, and with it the efficiency that optical compute has always promised.

The complete Neurophos optical tensor core is one assembly of three parts. First, a silicon photonics chip creates the input vectors and collects the output vectors. It is clocked at 56 GHz, and that clock sets the compute speed. It uses conventional large modulators, but only about a thousand of them, not a million, so it is more important that they be fast than small. Second, a meta-surface chip holds the weight matrix. Because the matrix stays the same across many input vectors, it is more important that these modulators be small than fast. Finally, an optical projection system connects the two chips. It copies each input beam from the silicon photonics chip across a row of the meta-surface, then sums the light reflected from each column by focusing it back onto a single photodetector on the silicon photonics chip. Packaged with HBM and a digital chip for control and memory, this optical tensor core becomes a drop-in replacement for a GPU. Per reticle-sized die, it is 50 times faster than today’s state-of-the-art GPUs, at about 35 times their fundamental energy efficiency.

Optics is already a big component of data centers, and the amount of optical content is expected to grow over the next 5 years as Scale-Up racks increasingly employ OCS (Optical Circuit Switching) vs EPS (Electrical Packet Switching). The OCS market is expected to grow ~5X over the next 5 years to $11B/year. The other large market is in the area of computing where OCUs (Optical Compute Units) are demonstrating significant improvements in compute speed, chip size and power consumption relative to electrical approaches like GPUs. According to the Yole Group, the first OCUs are expected to deploy in 2028, with the market expected to grow to ~$2.7B (1 million units) by 2034. The potential market is huge, however, as OCUs increasingly replace the GPU datacenter market for computing. This is estimated at ~$1T/year today.