Summary of Key Points
Recently, there has been a adjustment in the AI hardware sector, triggered by Meta's decision to make some of its computing resources available. However, this is not due to a decline in demand for AI computing power but rather an industry-wide transition period from using older technologies to newer ones. Global tech giants are still investing heavily in the development of the next generation of computing solutions (such as Blackwell/Rubin), while older models like H100/H200 are being repurposed for tasks that require lower performance, such as inference and fine-tuning of models. The future of AI competition will shift from who has the most computing power to who can use it most efficiently and profitably. The current market adjustments reflect an upgrade in market understanding, not a reversal in the industry's fundamental logic.
1. Meta Selling Computing Power: Not Because of Excess, but Because Old Equipment Needs New Uses
Many people are concerned when they hear that Meta is making its computing resources available, wondering if there is a surplus of power. But think about it this way: If Meta truly had enough power, why would it have invested $125-145 billion in infrastructure in the first half of the year, signed a $60 billion GPU contract with AMD, and purchased nearly $50 billion worth of CoreWeave computing resources?
The truth is that AI computing power, like smartphones, also goes through generations. H100/H200 were once very popular, but now larger models require more powerful new technologies like Blackwell for training. Although older models are not at the top of the spectrum, they are still useful for tasks like answering questions in ChatGPT and customizing AI solutions for businesses. By making these old resources available to others, Meta is essentially upgrading its infrastructure without wasting them and generating additional revenue—this is a case of asset optimization, not a sign of declining demand.
2. No Overall Surplus of Computing Power; Just a Transition from Old to New
To determine if there is a surplus of computing power, look at where the giants are investing their money. In 2026, Amazon, Microsoft, Google, and Meta plan to invest a combined total of $725 billion in AI infrastructure, a 77% increase from the previous year! Even Google's own investment of $180 billion is not enough, so it is looking for additional resources. Anthropic (the parent company of Claude) has signed a long-term agreement with AWS to secure computing power worth $100 billion—this certainly doesn't indicate a surplus.
The real issue is a lack of interest in investing in new technologies; rather, there is a fierce competition for new computing power and a redistribution of existing resources. It's similar to the transition from traditional cars to electric vehicles: fuel cars haven't disappeared, but they have shifted to uses like ride-sharing and logistics—the structure has changed, not the demand.
3. AI Competition Enters an Era of Efficiency
In the past two years, the focus has been on who has the most GPUs and data centers. Now, the game has changed; having computing power alone is not enough; you need to be able to make money from it. For example, with the same 1000 GPUs, some companies use them to train large models and generate revenue through advertising or subscriptions, while others leave half of their resources idle or use them for less profitable tasks. In the future, investors will pay more attention to how efficiently computing power is utilized and the proportion of revenue generated by AI services, rather than just the amount of power available.
4. How to Tell If AI Hardware Is Really on the Decline?
The current market adjustments are just a temporary blip. To determine if there has been a fundamental shift in the industry's logic, look for these three signs:
- Cloud computing giants like Microsoft Azure, Google Cloud, and Amazon AWS reduce their infrastructure investments (e.g., not building new data centers).
- Companies extend the lifespan of their servers (e.g., using them for five years instead of three).
- Leading AI companies (like OpenAI and Anthropic) stop competing for new computing resources or even return existing ones.
Since none of these signs have appeared yet, there's no need to panic. The overall trend in the AI hardware industry is still one of growing demand, with a shift in how that demand is being met.
In Conclusion
AI hardware is not on the decline; it has just moved from a phase of rapid growth to a period of more focused development. The focus has shifted from who can acquire the most resources to who can use them most effectively. Investors should not panic but consider which companies can turn computing power into actual profits—these will be the real opportunities in the future.