In recent years, driven by the AI boom, the global industry has been aggressively scaling up computing power.
GPUs are in extremely short supply, and NVIDIA's market capitalization has surged. Major tech giants are aggressively stockpiling and grabbing capacity. Meanwhile, some emerging chip startups are seizing the opportunity to rise and grab a share of the market.
While scaling up computing power, the industry has gradually realized a problem: as chip manufacturing processes approach physical limits, Moore's Law is failing. Chip power consumption is increasing, heat dissipation is becoming more challenging, and costs are becoming unbearable.
The performance improvement of traditional chips has reached a dead end. How should we face the ever-expanding computing power black hole? Is there a new technological path to break this deadlock?
Currently, the most promising solution lies in an old friend we are very familiar with—light. Optical computing technology, centered on photons, is opening a door to a new world with its unique advantages.
█ Why Has Electrical Computing Reached a Bottleneck?
To understand the value of optical computing, we must first figure out what has gone wrong with the electrical computing system we have used for decades.
Whether it is a CPU, GPU, or NPU, all traditional chips essentially rely on electron movement to perform logical calculations.
As we all know, traditional computing chips are composed of countless semiconductor transistors. Simply put, a transistor is a micro-switch. When current flows through it, it represents "1"; when it is disconnected, it represents "0".
By controlling the on/off states of transistors and the flow of electrons, 0s and 1s can be expressed and recorded, enabling binary addition, subtraction, multiplication, division, and logical judgments. The accumulation of massive logical calculations can accomplish various complex data processing tasks.
This classic computing architecture, born in the 1940s, has supported the rapid development of the entire information age and propelled the first half of the AI wave.
Today, this architecture has hit the ceiling and faces a series of insurmountable physical bottlenecks:
· Slower Speed
The speed of light in a vacuum is 3×10^8 m/s. The actual drift velocity of electrons in semiconductors is far less than the speed of light. Under the normal operating electric field of room-temperature silicon semiconductors, the electron drift velocity is only 10^3 to 10^5 m/s.
According to basic physics, an applied voltage creates an electric field. The slow, directional movement of electrons along the electric field is what we call carrier drift velocity. The weaker the electric field, the lower the velocity.
As electrons travel, they continuously collide with silicon lattice atoms and impurity atoms. Their directional movement is frequently interrupted and reversed, preventing the macroscopic directional drift rate from increasing.
It is important to note that the propagation speed of electrical signals in chips ≠ the speed of electron movement. In wired metal medium communications, the speed of electrical signals is determined by the magnetic permeability and dielectric constant of the medium surrounding the conductor. In copper wires, this speed is about 77% of the speed of light (230,000 km per second).
However, in reality, the signal delay in CPUs and RF chips is determined by the electric field propagation delay, not the electron drift velocity.
Taking a water pipe as an example: when we turn on the faucet, water flows out instantly. This is not because the water from the source instantly reaches the end, but because the water pressure (electric field) instantly propagates throughout the entire pipe. Under the influence of water pressure, the water molecules (electrons) in each part of the pipe only move forward slowly.
It doesn't matter if the principle isn't fully understood; just remember one point—the speed of electrical signals in chips is actually not as fast as we imagine. Slow signals mean high latency and low efficiency.
· Extremely High Power Consumption
Everyone knows about electrical resistance. Resistance occurs when electrons move directionally in a conductor and collide with lattice atoms, converting kinetic energy into lattice vibrations, thereby generating Joule heat (thermal energy).
Electron movement inevitably generates resistance, and resistance leads to heat generation and energy loss. This is an unchangeable physical characteristic.
As the transistor density of chips continues to approach physical limits, this "heat generation" has evolved from simple energy loss into a "thermal wall" that hinders further leaps in computing power.
Power consumption not only limits the improvement of chip computing power but also brings huge cost pressures (power consumption and cooling costs).
· Limited Density
The core logic of traditional computing power upgrades is stacking transistors and piling up chips.
However, the physical area of a chip is limited, and transistors cannot be stacked infinitely.
The higher the integration density, the greater the difficulty in wiring, cooling pressure, and signal interference, leading to continuously diminishing marginal effects in computing power improvement.
In short, as chip manufacturing processes continue to shrink, problems such as electron migration delay, energy loss, surging power consumption, signal crosstalk, tunneling effects, and skin effects are becoming increasingly severe.
The traditional computing architecture relying on electrons can no longer meet the computing power demands of the AI era. The entire industry urgently needs a brand-new computing paradigm to break the shackles. Optical computing is currently the most reliable path to break through this deadlock.
█ What Exactly Is Optical Computing?
As the name suggests, optical computing uses "photons" to perform calculations.
The industry defines optical computing as follows: Optical computing is a new computing paradigm that uses photons as information carriers and utilizes optical devices (such as lasers, modulators, detectors, optical waveguides, etc.) to achieve data calculation, storage, transmission, and processing.
When it comes to light, the first reaction is often optical communication. Over the past few decades, the development of optical communication technology has been obvious to all. The reason humanity uses light for communication is precisely because photon transmission has advantages that electron transmission can hardly match.
The speed of light is the fastest known transmission speed. Photon propagation has almost no resistance, generates no heat, and has no loss. Moreover, multiple beams of light can be transmitted in parallel in the same space and channel without interfering with each other.
Given the outstanding performance of light in the communication field, experts have considered whether light can be used for computing.
Our current computing power system is "optical transmission and electrical computing," where data transmission relies on light, and data calculation and processing rely on electricity. Optical computing means that data calculation and processing (including storage) also rely on light, achieving "optical transmission and optical computing." This is true all-optical computing, an "ultimate form."
Compared with traditional silicon-based electronic computing, the advantages of optical computing are mainly reflected in four aspects:
· Ultimate Speed Enhancement
Electron transmission speed is limited by the medium, while the speed of light is the cosmic speed limit. The computing speed of optical computing can be improved by hundreds or even thousands of times compared to traditional chips.
· Cliff-like Drop in Power Consumption
The biggest waste in electronic computing is resistance heating, whereas photon transmission loss is extremely low. Under the same computing power, the power consumption of optical computing is only a fraction or even one-hundredth of that of traditional chips. This can significantly reduce computing costs.
· Ultra-High Parallel Computing Capability
Electrical signal transmission is prone to mutual interference and crosstalk, so the parallel computing capability of electronic chips is limited.
Light, however, can achieve multi-wavelength and multi-channel parallel transmission, where optical signals in different optical paths and with different wavelengths do not interfere with each other. A single optical computing chip can achieve synchronous parallel computing of massive data, with a computing power density far exceeding that of traditional electronic chips.
· Strong Anti-Interference and Stability
Electronic chips are easily affected by electromagnetic interference, and their performance drops significantly in high-temperature and complex environments.
Optical signals, on the other hand, are immune to electromagnetic interference and can still work stably in extreme environments, making them suitable for high-end scenarios such as military, aerospace, and autonomous driving.
█ How Exactly Is Optical Computing Implemented?
The underlying principle of optical computing lies in how to express 0s and 1s.
The underlying data of computers consists only of 0s and 1s. Electrical computing uses high and low voltage levels of electrical signals to represent 0 and 1. Optical computing, however, uses the on/off, intensity, and phase changes of light to carry 0 and 1 data, achieving binary operations. The logic of PN junctions, holes, and gate voltages in transistors no longer exists in optical computing.
Optical computing uses photons as information carriers and utilizes optical phenomena such as light propagation, interference, diffraction, resonance, and superposition to express 0s and 1s, thereby achieving logical computing functions.
Optical chips are actually the same concept as silicon photonics (Photonic Integrated Circuit, or PIC) in the field of optical communication. Hundreds of optical devices (such as lasers, modulators, waveguides, and detectors) are integrated on a single chip to form a complete optical circuit system.
For example, the light output from a laser, when encountering a modulator, will have its intensity adjusted according to the input data: high light intensity corresponds to "1", and low light intensity corresponds to "0".
After that, photons pass through a pre-designed computing array. Taking the MZI (Mach-Zehnder Interferometer) technology route as an example, the key basic computing unit of the array is the MZI, which is also a phase adjuster. The phases of light are different, and the phase adjustment in each part varies. The combination of signals with different phases leads to constructive or destructive interference, thereby changing the light intensity. This achieves the effect of matrix computation.
The optical signal after interference carries the computation results. Finally, a detector receives the optical signal and converts it to output the computation results.
█ Can Optical Computing Immediately and Comprehensively Replace Electronic Chips?
The answer is obviously no.
Traditional electrical computing has a development history of nearly a century, and the entire hardware and software ecosystem is deeply rooted. It is absolutely impossible to be comprehensively replaced overnight. Many technologies in optical computing are still in the laboratory verification stage, and there is still a long way to go before large-scale commercial deployment.
Therefore, the industry generally believes that the current main task of optical computing is not to comprehensively replace electrical computing, but to first integrate into the existing electrical computing system to achieve optoelectronic hybrid computing, and then consider all-optical computing in the future.
Schematic diagram of optoelectronic integration (Source: Optical Computing White Paper for New Intelligent Computing 2025)
In fact, the current focus of optical computing is on large-scale matrix operations in AI large model training and inference.
The parallel computing advantage of optical computing is highly suitable for processing linear algebra-intensive tasks such as matrix multiplication, convolution, and multiply-accumulate operations, enabling exponential leaps in computing power with extremely low energy consumption.
The industry is committed to building an optoelectronic hybrid AI cluster system.
Among them, tasks such as control logic and task scheduling, storage and data management, high-precision general-purpose computing, and complex branch computing will continue to use electrical computing. Meanwhile, tasks with high repetitiveness and high parallelism, such as AI matrix operations, parallel floating-point operations, image computing power, and signal processing, will adopt optical computing.
In this way, the respective advantages of both can be leveraged, complementing each other and working collaboratively to build a new generation of computing power architecture with faster speed, lower power consumption, better cost-effectiveness, and stronger stability.
Notably, optical computing can be better integrated with optical communication. This is also very important.
Currently, AI clusters are basically adopting the all-optical interconnection route. Vendors such as Huawei are using optical interconnection. Although NVIDIA is still using electrical interconnection (copper connections), it is also actively embracing optical interconnection.
The highly popular NPO/CPO/LPO/XPO technologies in recent years, along with I/O and OCS technologies, are all optical communication technologies introduced to solve the bottlenecks of electrical interconnection in bandwidth and power consumption. These have paved the way for the seamless integration of optical computing and optical communication. In fact, technological research in optical communication also serves optical computing.
█ What Are the Research and Industrialization Progress of Optical Computing?
Optical computing is a field with great commercial value and strategic prospects. Global tech giants and top research institutions are accelerating their layouts, trying to seize the initiative in this field.
From the perspective of technology and manufacturing processes, optical computing is still in its early stages, and related technologies and processes are not yet mature enough. Optical computing chips require precise micro-nano optical structure design, ultra-high precision optical path packaging, light source control, and optoelectronic conversion modules, all of which are still under development. Compared with mature electrical chip processes, optical chips have low yield, high costs, and extreme difficulties in mass production.
From the perspective of the ecosystem, the gap is even larger. Currently, there is no mature software stack, compiler, toolchain, or community for optical computing, and the standard system is far from unified. The time required behind this is even longer.
Let's look at the specific progress at home and abroad.
Internationally, the main route is the silicon-based MZI (Mach-Zehnder Interferometer) coherent interference route mentioned earlier. Leading companies in the field of optical computing include startups such as Lightmatter and Luminous Computing. Tech giants like Intel and NVIDIA are also actively entering the field through mergers and acquisitions or in-house research.
Envise, launched by Lightmatter, is the world's first photonic computing platform (photonic accelerator card). It adopts a hybrid optoelectronic heterogeneous system combining mature CMOS processes with advanced silicon photonics integration technology. The computing power of a single card reaches 300 TOPS. It is mainly targeted at data center AI inference, has received investments from Google and NVIDIA, and is being piloted by North American cloud providers.
It is reported that the Envise 4S server, equipped with 16 Envise chips, has a total power consumption of only 3KW, which can reduce AI training energy consumption by 90%.
Domestically, the main route is "MRR (Microring Resonator) + photonic in-memory computing." Universities such as Tsinghua University, the Shanghai Institute of Optics and Fine Mechanics of the Chinese Academy of Sciences, Shanghai Jiao Tong University, and the University of Science and Technology of China are actively conducting research on optical computing.
Enterprises such as Lightbit Technology, Lightelligence, and Qisuan Guangqi have also made good progress in technological research and industrialization.
At the recent WAIC 2026, Lightbit Technology released a 256×256 photonic in-memory chip and plans to launch the second-generation computing card within the year, featuring a single-card computing power of 1,000 TOPS, benchmarking against mainstream AI inference GPUs.
█ Conclusion
The dividends of electronic computing have been exhausted. Optical computing is very likely the core key to the next-generation computing power revolution.
It is not a simple technological upgrade, but a brand-new transformation of the underlying logic of computing. Everything we are familiar with regarding computing could be completely overturned by optical computing.
Empowered by optoelectronics, computing power is reborn. The era of optical computing has just begun to unfold.