Recently, Musk stated that TSMC's concerns about excess chip capacity are "correct." He predicted that the limiting factor in the AI industry will shift from chip manufacturing to "activating the chips," and the core bottleneck of this transition lies in power supply, transformer configuration, and cooling system deployment.

On January 6th, Musk had an in-depth conversation with Peter Diamandis, Executive Chairman of Singularity University, and David Blundin, founder of Link Ventures, about the future of AI. He pointed out that current chip capacity is growing exponentially, while the energy infrastructure supporting it can only expand linearly. When these two curves intersect, a large number of high-performance AI chips will be unable to be put into use due to a lack of supporting power conversion equipment and cooling systems.

This judgment directly points to the serious underestimation of current AI infrastructure construction. For investors, this means the focus of the AI computing power race is shifting from chip procurement to the capability of building energy infrastructure.

Power becomes the "rate limiter" for AI deployment

In the conversation, Musk further revealed the specific form of the power bottleneck in AI infrastructure. He emphasized that deploying AI chips is not as simple as "shipping GPUs to a power plant," but rather requires simultaneously solving three core issues: gigawatt-level power supply, high-voltage power conversion, and efficient cooling systems.

Musk specifically pointed out that the entire data center industry is undergoing a critical transition from air cooling to liquid cooling, and warned that this process carries significant risks. He said:

"This is a fundamental transformation for data centers, which have long relied on air cooling. Once a liquid cooling pipe ruptures — for example, if a water pipe bursts in a data center — it could result in a loss of up to $1 billion."

He used xAI's "Colossus-2" project in Memphis as an example to illustrate the actual challenges: although the project site is located near multiple 300 kV high-voltage lines, it still took approximately one year to complete the connection. To achieve the launch of a 1-gigawatt training cluster by mid-January 2025, the team had to temporarily combine multiple gas turbines ranging from 10 to 50 megawatts as a transitional power source and use a large number of Megapack battery groups for power balancing.

The "crossing of curves" between chip capacity and power supply

When asked whether he agreed with TSMC's concerns about overcapacity, Musk gave a positive response: "Although I'm not sure about their reasons, the conclusion is correct."

He pointed out that the key is to identify the "limiting factor of each period" and predicted that by the third quarter of 2026 (approximately 9-12 months later), the core bottleneck would shift from chip manufacturing to the ability to "get the chips running."

This judgment stems from the misalignment of two development trajectories: AI chip capacity is expanding at an exponential rate, while the power infrastructure supporting it can only grow linearly. Musk emphasized: "If chip production grows exponentially, while power supply can only increase slowly in a linear fashion, the two curves will inevitably intersect." This means that the speed of chip production may far exceed the speed at which they can actually be deployed and powered on.

In response, David Blundin presented a different view, arguing that even if TSMC increases GPU production from 20 million to 40 million units, the market will find ways to solve the power supply issue. However, Musk maintained that any missing link in the power conversion or cooling system would prevent the chips from being truly activated, thereby fundamentally suppressing actual demand and purchasing behavior.

This article is from Wall Street Watch. Welcome to download the app to read more.

Original: toutiao.com/article/7592533394014126627/

Statement: The views expressed in this article are those of the author himself.