Summary of Key Points
Meta's plan to sell its idle AI computing power has caused panic in the global market (with chip stocks such as NVIDIA plummeting), but this does not indicate an overall surplus of computing power in the industry. Instead, it is a structural issue stemming from Meta's own strategic mistakes: over the past two years, the company has spent heavily on purchasing GPUs and acquiring companies along the AI supply chain. However, due to a lack of cloud services to utilize this power, lagging model development, and difficulties in integrating acquisitions, the computing power has remained unused. The market has misinterpreted Meta's individual challenges, while the actual industry demand is still growing rapidly. Meta's adjustment represents a shift from a reckless arms race to more pragmatic business operations, marking a transition from an inflated bubble to a more mature reality.
Why Is Meta Suddenly Selling Computing Power?
Meta's actions in AI over the past two years can be described as "reckless spending": it expects $145 billion in capital expenditures for 2026, signed a five-year contract for 60 billion chips with AMD, and invested $21 billion in CoreWeave. It has also acquired companies in data annotation, voice interaction, hardware, and nearly the entire AI supply chain. However, these investments have not yielded tangible benefits; instead, they have created three major problems:
1. Lack of cloud services to utilize computing power: Google and Microsoft can earn money by renting out their GPUs through cloud services, while Meta's main revenue comes from advertising, which has a limited demand for additional computing power (once algorithms are optimized to a certain extent, more GPUs do not significantly increase click-through rates), turning excess capacity into pure cost.
2. Slow model development: Meta's Llama series models have been criticized for being "cheated," and the next generation of models has been delayed. While other companies are iterating on their models, Meta was busy with acquisitions, resulting in its core products falling behind.
3. Difficulties in integrating acquisitions: The acquired companies each use different technologies and cultures, making coordination challenging (for example, how to integrate data annotation with hardware, and who will set the rules?). These issues have combined to create an idle capacity that Meta must sell to reduce its financial burden.
Where Does the Market's Panic Go Wrong?
The market panicked when it saw Meta selling computing power, assuming the AI bubble was about to burst, but this logic is too simplistic:
- **Meta's surplus is "structural": It simply does not have enough use for all the power it has acquired; it's not that the entire industry no longer needs it. For instance, OpenAI's o1 model is still making breakthroughs in inference capabilities, and domestic companies like DeepSeek are approaching top levels at lower costs. The industry's technology is not stagnant—Meta has just fallen behind.
- Demand has not declined: Tokens (the "energy" consumed by AI applications) are a good indicator of demand. In March, China's daily average token usage reached 140 trillion, a 1,000-fold increase from the beginning of 2024, indicating that AI applications are still booming, and many small and medium-sized teams are still waiting for GPUs.
- Meta's computing power is quickly sold out: If there were an oversupply, it would not be so easily acquired. In reality, the power is being transferred from Meta (an inefficient user) to companies that can make effective use of it—a normal market mechanism for optimizing resource allocation.
Is There Really a Surplus of Computing Power in the Industry?
The computing power market is similar to the real estate market: high-end GPUs (core areas) are in high demand, while older models (outer suburbs) are harder to sell.
- High-end power is still in short supply: The rental price of H100 GPUs has increased by 40% in the first half of the year, and they are often sold out quickly. Meta is selling the previous generation of H100 GPUs, which are indeed partially idle, but top-tier training power remains in high demand.
- Differential demand: Large model companies need the latest GPUs, while smaller teams may use older models, but overall demand has not decreased. Treating Meta's sale of older GPUs as a sign of an industry-wide surplus is like saying the entire housing market has collapsed just because houses in the suburbs are cheaper.
Does Meta's Sale of Computing Power Mean Defeat?
Meta's move is more of a realization of its limitations rather than defeat:
- Strategic correction: It realized that spending heavily was not effective and is now focusing on cost-effectiveness. Selling computing power helps to monetize idle resources and reduce losses.
- Pragmatism: Meta is no longer pretending to be a leading AI player but is shifting to a role as a provider of computing power, concentrating its resources on profitable areas (such as advertising). A more cautious Meta might have a longer-term competitive advantage.
For the Industry: Bubble Correction, Not a Collapse
This shake-up clarifies that the computing power market is moving from a state of "infinite scarcity" to one of "tiered pricing." Meta's withdrawal does not change the long-term demand for AI computing power; instead, it promotes a more rational approach:
- Bubble correction: People are now more discerning about which computing powers are valuable and buying them based on actual needs.
- Sign of maturity: The industry is shifting from an arms race to practical applications, with only truly valuable AI companies surviving. The superficial ones will be eliminated.
In summary, Meta's issues are internal, but the AI industry continues to grow, albeit less recklessly.
(End of translation)