TSMC held its third quarter legal meeting yesterday (16th). Judging from the questions raised by legal persons, in addition to the company’s business, the most concerning thing is whether the AI bubble bursts. Currently, Chairman Wei Zhejia holds a positive view, but it seems unable to alleviate the market’s concerns about the AI bubble.
Foreign media "Financial Times" (FT) pointed out that there is a wonderful sign of AI equipment, that is, GPUs are expensive to buy, but very cheap to rent. For example, when the NVIDIA B200 was first launched at the end of 2024, a single unit cost more than $500,000. However, by the beginning of 2025, the hourly rental price was only $3.20, and the minimum rent dropped to $2.80 per hour last month.
Since NVIDIA upgrades its chip architecture every two years, data center operators with the most financial resources can lease "non-latest generation" chips at extremely low prices to bind customers.
However, although the steady decline in GPU rental prices is a common situation of "price-cutting competition", the report believes that the reality is more complicated.
Judging from price data collected by RBC Capital Markets, hourly rental prices for NVIDIA H200 and H100 chips have fallen by 29% and 22% respectively this year, but prices have not changed much among the four cloud giants (this refers to Amazon, Microsoft, Google and Oracle), making the price gap between small operators and ultra-large-scale cloud service operators ever widening.
The reason why the overall GPU rent has fallen is largely due to new entrants, which is different from the actual price war. Generally speaking, price wars are mostly caused by large companies using large quantities and low prices, forcing small companies to follow up in price-cutting competition.
Reports speculate that this may be related to the type of user using the GPU. The first is AI startups or research institutions. They need to conduct a large number of model trainings in a short period of time, so they require huge computing power. These companies or institutions are already existing users of cloud giants, so they are more accustomed to using the same platform and are willing to pay even if the price is higher.
The second type is general companies that want to build website chatbots, summary tools, or similar AI gadgets. They may have chosen to rent GPUs in the cloud in the past, but as relevant technologies gradually become available, these companies may directly use ready-made large-scale language models such as OpenAI and Anthropic to build chatbots, and pay by "token" instead of by the hour.
The last type is those who have been sucked away by customers but still remain in the market and rent GPUs. These may be amateur players or small teams such as academics with limited funds, investment teams who want to do financial models, etc. According to reports, in the face of most of the customer base being sucked away by large companies, we can only continue to lower prices to attract the remaining users.
Looking at the current price, an entry-level NVIDIA DGX A100 cluster (containing eight GPUs) will cost about US$199,000 when it is launched in 2020. Assuming a chip life of about five years and full year-round operation, each GPU would have to generate at least about $4 an hour in revenue to break even.
But by 2020, the average A100 rental price was US$2.40 per hour, and now it has dropped to US$1.65 per hour; in contrast, very large cloud operators still rent for more than US$4 per hour, and some small competitors are as low as US$0.40 per hour. Although many important factors are ignored, it can still determine whether the current pricing is reasonable.
The final report believes that it can be inferred that the NVIDIA GPUs purchased during the epidemic may not be repaid before they are taken to the second-hand market; even if customers are attracted by low prices, they lack the ability to pay higher fees; ultra-large cloud companies choose to wait and see, waiting for low-price market competitors to go bankrupt; it is expected that the wave of data center consolidation is coming, which will eliminate a large number of AI startups that cannot afford real computing costs; if ordinary companies only want to use OpenAI or Anthropic Help build chatbots and summary tools. The actual value may be less than $3 trillion, and the size of the GPU market may be overestimated.
What GPU pricing can tell us about how the AI bubble will pop Further reading: AI demand is strong, US factory gross profit dilution reduced! JPMorgan maintains TSMC outperforms and maintains target price of NT$1,550 To accelerate high-capacity HDD verification, Witten Electronics launches expansion of SIT laboratory