Deutsche Bank’s report warns that companies will divide into winners and victims in the era of artificial intelligence, weakening the viability of traditional portfolio diversification.
Conditions for achieving superintelligence by 2030
The $900 billion bill and power and chip requirements
Three investment challenges ending the era of traditional diversification
The possibility of superintelligence failing to materialize and air pocket risks
Frequently asked questions
Conditions for achieving superintelligence by 2030
Artificial intelligence could reach the stage of “superintelligence” by 2030, according to a new research paper published by Deutsche Bank Research. However, the German bank sets a major condition for this to happen; this scenario will only occur if model scaling continues, AI systems learn how to self-develop and self-improve, and the industry manages to overcome severe constraints imposed on electrical power, infrastructure, and financial funding.
The research paper was titled “AI Boom or Bust: What Will Superintelligence Look Like and When Will We Get There?,” authored by Deutsche Bank analyst Adrian Cox in collaboration with external expert Azeem Azhar, to examine what the “San Francisco consensus” on reaching human-surpassing intelligence by 2030 means for companies and investors.
The report’s core idea is that investors may not actually need superintelligence to arrive for AI expenditures to bear fruit, but they certainly need revenue growth to keep pace with steep costs.
The $900 billion bill and power and chip requirements
Deutsche Bank estimates that training a hypothetical general-intelligence-like model could require infrastructure costing between $300 billion and $500 billion and about 6 gigawatts of power. If this training is executed over four months using Nvidia’s Vera Rubin chips, it could require roughly 7.8 million chips and 26 gigawatts of utility power, at a cost approaching $900 billion in today’s prices, though continued improvements in computing efficiency could lower the cost of a similar cluster to about $200 billion by 2030.
The bank expects Alphabet, Amazon, Meta, Microsoft, and Oracle to spend nearly $800 billion on capital expenditures this year alone, noting that their annual spending could surpass the $1 trillion mark by 2030 and beyond.
Analyst Adrian Cox wrote: “Under our baseline scenario, the best-performing AI models by 2030 will possess superintelligent characteristics, meaning they will be better than human experts across all domains.”
Three investment challenges ending the era of traditional diversification
Cox argued that artificial intelligence could sharply divide companies into clear winners and clear losers, which will weaken the usual protection provided by portfolio diversification. He stated, “Spreading capital across more names will only spread exposure to risk rather than mitigate it,” and urged investors to sort their holdings into “AI adopters or AI casualties.”
Second, the report warned against interpreting quiet price action as lower risk, citing education company Chegg, whose stock lost half its value in a single day in May 2023 after management blamed slowing customer growth on the emergence of ChatGPT.
Third, Cox cautioned that dollar-cost averaging down when a stock drops may no longer be wise, explaining that “a lower price may be information that a company’s business model has become obsolete rather than a buying opportunity.”
The possibility of superintelligence failing to materialize and air pocket risks
The report left room for the possibility that superintelligence might not materialize at all. According to MarketWatch, owned by News Corp, the bank’s view suggests it may never be achieved, even though the U.S. economy will still be able to reap significant productivity gains driven by AI.
Corporate adoption of AI is moving at three times the speed of previous mobile, internet, and cloud computing waves, and AI economy revenues outside China could exceed $200 billion in 2026 compared to less than $1 billion in 2022.
Nevertheless, the bank warned that a large portion of infrastructure is currently funded based on projected future revenues, putting pressure on the cash flows of major cloud companies. It cited a warning from Jaspreet Sokhi, chief strategist at Google DeepMind, regarding the risk of an “AI air pocket” where heavy upfront spending occurs while projected revenues lag behind.
Frequently asked questions
Question: What is the estimated cost of training a superintelligence model in four months?
Answer: Deutsche Bank estimates that training over four months on Nvidia Vera Rubin chips requires 7.8 million chips and 26 gigawatts of electricity at a cost approaching $900 billion in today’s prices.
Question: How much will the top five technology companies spend on infrastructure this year?
Answer: The report expects Alphabet, Amazon, Meta, Microsoft, and Oracle to spend nearly $800 billion this year, with annual spending exceeding $1 trillion by 2030.
Question: Why does Deutsche Bank warn investors against relying on traditional stock diversification?
Answer: Because AI will sharply divide companies into winning adopters or casualties whose business models are becoming obsolete, making the purchase of declining stocks a risk rather than an opportunity.
Question: What is meant by the term AI air pocket in the report?
Answer: It is a warning issued by Jaspreet Sokhi of Google DeepMind pointing to the risk of deploying massive upfront investments and capital expenditures without actual revenues successfully catching up.
