In-Depth Analysis: The "Power Revolution" Behind the Surge in AI Computing Power and the Chip Price Hike
Hello everyone, I'm your financial journalist. Recently, there's been a hot topic in the tech community called the "AI slowdown theory." In simple terms, many people are concerned about whether the development of artificial intelligence (AI) is happening too quickly and whether it might suddenly come to a halt. This concern indeed caused some instability in the semiconductor stocks in the U.S. stock market in mid-September.
But if you look closely at what's been happening, you'll discover a more fundamental logic: AI isn't slowing down; instead, due to its high power consumption, it's triggering a complete transformation from the power grid to the servers themselves. The main players in this transformation are no longer the well-known CPUs or GPUs, but rather the power semiconductors that are responsible for delivering and converting electricity.
Today, we'll break down this news in plain language to understand the opportunities and risks involved.
---
1. Demand Isn't Estimated, It's Forced: The Anxiety of Cloud Providers
First, let's clarify one thing: Are AI data centers still being built? The answer is yes, and they're being built even more urgently than expected.
The president of Infineon's Greater China region, Pan Dawei, mentioned something crucial: "The demand from customers three months ago was one number, and three months later, it was another number." This indicates that demand is not static but is exploding dynamically.
In the past, we thought that servers could be bought, plugged in, and used right away. But now it's different. For cloud service giants like Amazon, Microsoft, and Google, their biggest bottleneck is no longer the lack of chips; it's the limit on their physical construction capabilities.
For example, imagine you want to open a popular hot pot restaurant. You've bought enough ingredients (chips) and hired all the chefs (engineers), but the speed of renovating your restaurant (data center) or the approval process for power connections is too slow. In this situation, the company that can deliver electricity stably and efficiently in a short time will gain the upper hand. So, although there's concern about an AI bubble, the construction of infrastructure, due to its high fixed costs (buildings, equipment), is difficult to stop once it's underway. Instead, the surge in demand has made the supply chain even more strained.
2. Why the Sudden Shift to 800V DC Power? Because Traditional Power Supply Methods Can't Keep Up
This is the core technical reason behind the news. Our computers and servers used to run on low-voltage alternating current (AC). However, AI chips, such as NVIDIA's GPUs, now consume a tremendous amount of power:
- In the past: A GPU used to consume a few hundred watts, and a large server rack would be considered high-powered if it used 100 watts.
- Now: A single GPU can consume 2000W or even 4000W, and in the future, it could reach 10 kilowatts. The power density of entire racks is increasing to 200W, 500W, and the next generation could reach 1 megawatt (1000W).
This hits the limitations of traditional power supply systems. AC conversion efficiency is low, and the wiring is complex and heavy, making heat dissipation more difficult. If we continue using old methods, data centers will become huge "furnaces" with exorbitant electricity costs.
Therefore, NVIDIA is leading the call for a switch to 800V high-voltage direct current (HVDC).
Why use DC? Because DC requires fewer conversions, resulting in less energy loss. It's like taking a high-speed train from Beijing to Shanghai directly instead of making three stops, which is faster and more efficient. This not only improves the computing power generated per kilowatt of electricity but also simplifies the data center design. It's about saving money and increasing efficiency.
3. The Arrival of Solid State Transformers (SSTs): The End of Traditional Transformers?
Since the power supply method is changing, the devices responsible for converting electricity also need to evolve.
Traditional transformers, which you might see in residential areas, work on the principle of electromagnetic induction and consist of copper coils and iron cores. They have drawbacks: they're large, relatively inefficient, and not well-suited for handling high-voltage DC.
The current trend is towards solid state transformers (SSTs), which can be thought of as electronic versions of transformers. They use power semiconductor devices (highly efficient electronic switches) for voltage conversion.
- Traditional transformers: Like old-fashioned pumps that rely on mechanical rotation, they have frictional losses.
- SSTs: Like smart valves that control water flow precisely with electronic signals; they're smaller, more efficient, and respond faster.
Infineon is already developing 20kW and 30kW SST solutions, and many cloud providers have started testing them. This means that traditional power equipment manufacturers may face technological challenges, while those with core semiconductor technology will enter a new market opportunity.
4. The Third Generation of Semiconductors (SiC and GaN) Become a Must-Have
Here are two terms to know: silicon carbide (SiC) and gallium nitride (GaN). These are the so-called third-generation semiconductors.
Why are they suddenly so important? Because traditional silicon (the material used in CPU chips) isn't efficient enough under high voltage, high frequency, and high temperatures:
- Scenario 1: In 800V DC systems, high-efficiency power conversion is needed, and SiC and GaN devices are more efficient and generate less heat.
- Scenario 2: Server power supply modules (PSUs) are also being upgraded. Infineon's roadmap shows that when a single PSU exceeds 30kW in power, SiC and GaN are essential; traditional silicon materials are no longer sufficient.
It's like car engines. In the past, gasoline engines (silicon) were fine for ordinary cars, but for supercars (AI servers), more efficient hybrid or electric systems (SiC/GaN) are required.
Therefore, power semiconductors are no longer just supporting components; they are the core devices that enable the implementation of AI data centers. Without these efficient electronic switches, the 800V DC architecture wouldn't be possible, and AI computing power wouldn't be able to be released effectively.
5. The Price Hike Is Here: The Struggle Between Capacity Expansion and Cost Pressures
Finally, let's talk about money. The news mentions that companies like Infineon, STMicroelectronics, and Texas Instruments are all raising prices. Infineon has even increased prices twice.
Why do they dare to raise prices? There are two main reasons:
1. Demand is surging, and capacity can't keep up: Although Infineon started production at its new factory in Dresden, Germany, and invested an additional 500 million euros in expansion, the demand driven by AI is growing exponentially. When supply falls short of demand, sellers have more power.
2. Cost pressures are immense: Not only the chips themselves, but the entire semiconductor supply chain is facing rising costs—energy, raw materials, transportation, and even labor services. Building new factories and purchasing new equipment (especially for producing 12-inch GaN wafers) requires significant capital, which is reflected in the price of each chip.
What does this mean for ordinary people or investors?
- Short-term: Server manufacturers and data center operators will face increased costs, which may squeeze their profit margins.
- Long-term: This is a necessary part of the industry's technological upgrade. Companies that control core materials (SiC, GaN) and advanced packaging technologies will have more pricing power and market share.
- Risk Warning: Pan Dawei also noted that future prices are difficult to predict because they're linked to raw material prices. If global inflation continues or AI demand suddenly slows down, the newly expanded capacity could become idle, leading to price wars.
Summary
This news might seem to be about fluctuations in semiconductor stocks, but it's actually about the energy revolution behind AI infrastructure.
The development of AI has shifted from competing on computing power (buying more GPUs) to competing on energy efficiency (how to supply power more efficiently). In this process, the 800V DC architecture is the direction, solid state transformers are the means, and third-generation semiconductors like SiC and GaN are the key.
For non-experts, remember this: The future bottleneck for AI computing power may not lie in the chips themselves, but in how to deliver electricity to the chips efficiently. Whoever masters this technology will control the "switches" of the AI era.