Deutsche Bank: AI superintelligence can be expected in 2030, electricity and cost are key constraints

The Zhitong Finance App learned that in its latest report, Deutsche Bank outlined a scenario: if AI models can improve themselves and computing power continues to expand, artificial intelligence may reach “superintelligence” in 2030. However, the bank also warned that scientific uncertainty, power constraints, and cost pressures could derail this path.

Can superintelligence be expected in 2030? Setting a “hard threshold” for electricity, data and costs

The Deutsche Bank Research Department stated in the report that whether the above results can be achieved will depend on whether progress in large-scale AI technology can continue, and whether the system can independently improve the technology used to construct future models. The report also lists a number of constraints, including electricity supply, data availability, and the time required to build the calculation infrastructure.

The report predicts that Alphabet (GOOGL.US), Amazon (AMZN.US), Meta (META.US), Microsoft (MSFT.US), and Oracle (ORCL.US) will spend close to 800 billion US dollars this year, mainly for data centers; by 2030 and beyond, the annual capital expenditure will exceed 1 trillion US dollars.

Deutsche Bank estimates that training a hypothetical general artificial intelligence (AGI) would require 300 billion to 500 billion US dollars of infrastructure and about 6 gigawatts of electricity. Using Nvidia (NVDA.US) Vera Rubin chips, completing this type of training within four months would require approximately 7.8 million chips and 26 gigawatts of facility electricity, costing up to 900 billion US dollars at current prices. According to the report, if the current trend of improving computing efficiency continues, the cost of similar clusters could drop to about 200 billion US dollars by 2030.

AI investment logic and risk: adopt acceleration and be wary of “AI pitfalls”

Deutsche Bank said that the investment logic of AI does not depend entirely on whether superintelligence can be achieved. Businesses are adopting AI tools at about three times the speed of previous waves of mobile, internet, and cloud computing. The AI economy's revenue outside of China is likely to exceed 200 billion US dollars in 2026, and less than 1 billion US dollars in 2022.

However, Deutsche Bank warns that the financing of AI infrastructure still relies mainly on expectations of future returns rather than current revenue. The report also notes that as capital expenditure rises, the cash flow of hyperscale cloud vendors is under pressure.

The report quoted Jasjeet Sekhon, chief strategy officer of Google DeepMind, as saying that the market needs to be wary of the risk of an “AI pocket” (air pocket), that is, expenses have already occurred, but expected revenue has not been realized.