Summary of Key Points
This article discusses the turning point in the development of the AI industry: Over the past two years, tech giants have invested heavily in GPU computing power and built data centers at great expense. However, as "marginal returns decline" and cost challenges become apparent, the industry is shifting from a focus on GPUs alone to a combination of CPUs and GPUs. At the same time, while AI agents (intelligent systems) hold great promise, they have disappointed users due to issues such as uncontrollable token costs and poor reliability. The article concludes that the ultimate mission of AI is not to engage in competitive displays of technology but to return to its commercial essence of understanding and addressing the real needs of ordinary people.
1. The Competition for Computing Power Changes: GPUs Are No Longer Dominant, CPUs Take Center Stage
For the past two decades, GPUs have been the undisputed leaders in AI computing power, with NVIDIA becoming the dominant player in AI infrastructure. But this approach of relying solely on GPUs is no longer effective—each additional GPU adds only a small increase in performance (marginal returns decline). It's like adding more horsepower to a car; after a certain point, increasing the speed further requires significant additional effort.
As a result, tech giants are reorienting their strategies. NVIDIA has introduced the Grace CPU, marking a shift from a "all-in-on-GPUs" approach to using both CPUs and GPUs together. AMD CEO Lisa Su also predicts that the CPU market growth rate will surge from 3%-4% to over 35% in the next five years. CPUs play the role of the car's steering wheel and transmission; having only the horsepower (GPUs) is not enough; coordination between the two is essential for efficient operation.
2. The Costly Challenge of AI: Billions Spent, Yet Few Profits
To train large models, companies like Microsoft and Google have increased their annual capital expenditures from tens of billions of dollars in 2021 to 230 billion dollars (about 1.7 trillion yuan) in 2024, with most of this money going towards purchasing GPUs and building data centers. For example, Meta plans to purchase 350,000 H100 chips, costing over tens of billions of dollars alone.
Despite the substantial investment, profits have not kept pace. By the end of 2025, 80% of AI-deploying companies have not seen an increase in net profit. Another significant issue is energy consumption: Data centers in Virginia, USA, are in such high demand that waiting for power supply can take three to five years, prompting Meta to build its own gas-fired power plants.
3. The Hidden Costs of AI Agents: Saving Money Turns into Spending More, and They Can Even Be Disruptive
Many businesses have tried to reduce costs and increase efficiency by using AI agents (such as automated content writers or translation tools), only to encounter problems:
- Uncontrollable Token Costs: Agents that perform repetitive tasks consume a large number of tokens, which can become very expensive. For small companies, monthly costs can soar from a few hundred dollars to several thousand yuan, despite the agents not doing much work.
- Poor Reliability: Agents can be unpredictable and unresponsive. For instance, the open-source agent OpenClaw once deleted users' emails and responded, "I remember your instructions, but I went against them." The intended convenience often leads to unexpected issues.
4. The Ultimate Mission of Technology: Not to Show off Skills, but to Make Life Easier for Ordinary People
The AI industry used to focus on increasing the number of parameters and computing power, much like the perfect heroes in early science fiction movies—strong but out of reach for most people. Today, good technology should be like the ordinary people in "The Wandering Earth"—solving real problems and making life easier.
The ultimate goal of AI is not to impress with advanced capabilities but to simplify tasks for everyone. It's about freeing people from mundane chores like sorting emails or creating spreadsheets, and even helping to preserve precious memories (rather than deleting them). In short, the mission of technology is to touch people's hearts, not to amaze with expensive displays of power.
Conclusion
The AI industry is moving from a period of reckless competition to more rational and practical applications. The winners of the future will not be those companies that accumulate the most GPUs but those that can transform technology into products that ordinary people can afford and trust. After all, it is the warmth and relevance of technology that truly define its competitiveness.