The Computing Power Competition in the AI Era: The Game of Dreams and Costs

Big companies chase their AI dreams without rest

This week, financial reports from tech giants like Microsoft, Google, and Meta were released. Except for Meta, Microsoft and Google's financial reports performed well, boosting their stock prices. These companies are continuously enhancing their cloud revenue growth and overall profit margins through a "revenue-expense-revenue" positive cycle logic. AI is driving rapid growth in the scale of businesses such as advertising and search. At the same time, the "arms race" for Computing Power among top cloud providers is in full swing, and the likelihood of a slowdown in capital expenditures in the short term is almost zero. In the early days of the AI era, top companies can only survive through continuous investment amidst fierce competition. According to the Computing Power service provider CoreWeave, the balance point of AI Computing Power supply and demand may not arrive until 2030.

AI applications have yet to bear fruit

The market is looking forward to seeing "grounded" AI applications that can truly enhance productivity or facilitate life. This year has already witnessed the stunning performance of OpenAI Sora and the rise of the domestic large model Kimi. In the future, we can expect continuous updates of large models such as GPT, Gemini, and Llama, as well as giants like Microsoft and Apple introducing large models to PCs, mobile phones, and other endpoints. The road to AGI is winding and long, and it may take a considerable amount of time to incubate AI applications that truly change society.

Focus on "High Cost-Performance Computing Power"

Looking back at the 4G and 5G eras, the decrease in mobile data costs brought about the golden age of mobile apps. Similarly, the decrease in the cost of unit Computing Power will be a necessary condition for humanity to enter the AGI era. Affordable, accessible, and stable AI Computing Power is the foundation for the birth of significant AI applications. Taking Sora as an example, it will take several more months to open to the public, one important reason being the need to optimize the Computing Power required for inference. Leading cloud providers have the strength to invest heavily to maintain competitiveness, but this does not represent the overall market demand. What many small and medium-sized enterprises truly need is high cost-performance Computing Power.

Breaking down the Computing Power costs, aside from the approximately 10% electricity operating costs, the rest are almost all fixed asset investments, including GPUs, network equipment, cooling, etc. In terms of GPUs, TSMC is expanding the CoWoS capacity required for AI Computing Power, and it is expected that the monthly capacity will approach 40,000 wafers by the end of the year, which is an increase of over 150% compared to the total capacity in 2023. Regarding network equipment, NVIDIA's GB200 uses a large number of copper cables for short-distance connections, indicating that it not only pursues ultra-high performance but also considers cost control for customers. Optical module manufacturers are also actively promoting high cost-effective optical connection solutions such as LPO. In terms of cooling, as the power density of single cabinet increases, the cost-effectiveness of liquid cooling solutions will surpass that of air cooling.