In the past, everyone always felt that AI volume training and GPU training clusters often cost several computer rooms and thousands of cards, and the cost was hundreds of millions. But in 2024, the wind direction suddenly changed. Everyone found that what really determines the efficiency of AI model implementation is actually the reasoning computing power.
Especially in the tide of generative AI, from ChatGPT to various AI drawings, intelligent customer service, automatic video generation... After the model is trained, it needs to be put into production and serve users. Who can run reasoning faster and more stably. And behind this, there is a key but less-mentioned role-Optical Transceiver, which is becoming more and more important.
1. Why AI reasoning is becoming more and more important
To put it simply-AI has stepped down from the altar and has begun to truly serve the business.
For example, if you are using an AI customer service, when you enter a sentence, the AI background must immediately understand what you mean, retrieve the data, generate an answer, and send the result back within hundreds of milliseconds. Behind this process is actually reasoning.
For services like this, there are thousands of simultaneous requests every day. The larger the model and the more requests, the higher the requirements for back-end reasoning capabilities. According to IDC data, as early as 2022, the proportion of reasoning in cloud AI computing power has exceeded that of training (58.5% vs 41.5%). By 2026, the proportion of reasoning is expected to increase to 62.2%. In other words, AI is entering the era of reasoning from the era of training. Models are no longer just made to see, but to run and make money.
2. The surge in data volume has brought about a rigid demand for Optical Transceiver
If there is too much reasoning, the high-speed flow of data is naturally indispensable.
Today's AI clusters are often composed of hundreds of GPU cards. These cards must be "strung" together, and high-frequency, high-speed, and low-latency data communication must be carried out between each other, otherwise the model cannot run smoothly.
The "string" here, to put it bluntly, relies on network interconnection, and one of the core devices responsible for these high-speed interconnection is-Optical Transceiver.
The role of Optical Transceiver can be simply understood as "building an ultra-fast information highway between GPUs, servers, and switches." Only when the communication is fast enough can the AI computing power really run. In particular, the current AI main force uses 800G or even 1.6 T high-speed Optical Transceiver, which is used for the internal interconnection of high-speed data centers. The efficiency of each link may directly affect the processing speed and stability of the entire platform.
So don't look at Optical Transceiver as a small one, but in scenarios such as AI clusters with extreme communication rates, its performance is directly related to: whether the reasoning response speed can be real-time, whether the model service stability is high enough, and the overall energy consumption of the data center Can it be controlled?
3. The more the AI model is, the greater the demand for Optical Transceiver
You may ask, will the demand for Optical Transceiver continue to rise in the future
The answer is: almost a certainty.
Because the AI industry is still evolving rapidly, the new generation of models (such as OpenAI's o1-preview) has a larger number of parameters and higher calls, and the requirements for bandwidth and connection capabilities are rising. In order to save the energy consumption cost of training and reasoning, AI manufacturers have increasingly begun to deploy solutions such as photoelectric fusion, high-speed optical interconnection, and liquid-cooled Optical Transceiver. Optical Transceiver is no longer just a "supporting role", but has become a part of supporting AI computing power infrastructure.
The extent of AI computing power in the future depends largely on: whether you can make massive data run efficiently and stably. AI cannot run by relying on large models alone. Model training depends on computing power, and model landing depends on reasoning. Reasoning is inseparable from high-speed interconnection, and interconnection depends on Optical Transceiver. At the bottom of this AI revolution, the "invisible" Optical Transceiver is silently pushing the entire industry forward.
Therefore, the next time you see an intelligent customer service reply in seconds or an AI video generated in a few seconds, it may be an Optical Transceiver that has completed tens of millions of data transfers in a few microseconds, which makes all this possible.
