Artificial intelligence: investment boom, infrastructure and societal implications Report Interpretation

Overview

This second instalment of Deutsche Bank Research Institute’s “AI at 70” series examines what seven decades of breakthroughs and false starts imply for the present AI investment cycle. It argues that the boom’s pace and scale are extraordinary, but that technological progress, commercial returns and broader social outcomes will not move in a straight line.

Core views

The report frames the current AI cycle through the 70-year history since the 1956 Dartmouth Summer Research Project, when John McCarthy coined the term “artificial intelligence.” Building on the first report’s discussion of non-linear growth, infrastructure constraints and valuation booms and busts, this instalment looks forward at investment speed, model capabilities, technology adoption, IPOs and social consequences. Its central message is that context matters: historical analogies help explain both the magnitude of the current boom and the frictions that may shape its outcomes. On investment, Deutsche Bank argues that this boom is remarkable not only for scale but for speed. Expected hyperscaler capex in every year through 2030 exceeds the whole dot-com telecoms boom, based on estimates that include five major US hyperscalers—AWS, Google, Meta, Microsoft and Oracle—plus tier-two/neocloud and Chinese hyperscalers. The report therefore treats the investment cycle as unusually concentrated and rapid, with the stakes heightened by the scale of required infrastructure. The technology itself has a “jagged frontier.” AI can solve difficult tasks while still failing at seemingly simple ones, illustrated by the report’s contrast between complex capabilities and difficulty telling the time on a clock. At the same time, models are commoditising rapidly: top models hold the leading position for less than two months on average, and pricing is under pressure. The report also cautions that benchmarks are ageing even faster than models; apparent plateaus may reflect either slower model improvement or tests that are no longer demanding enough. These observations imply that headline capability leadership may be fleeting and that measured progress requires careful interpretation. The report links adoption to falling costs, drawing parallels with earlier technologies including horsepower, electricity, mainframes and solar. Its argument is that use can accelerate once costs decline sufficiently, but deployment also depends on physical capacity. Electricity is a central constraint: China has left the United States behind in generation, while US data centres are requiring a growing share of available capacity even as demand rises. The report presents global data-centre electricity use for 2025 and projected 2030 as part of this infrastructure challenge. Capital-market history offers a mixed guide. The report expects mega AI IPOs to eclipse entire past years of share sales, citing SpaceX’s IPO as exceeding all but three inflation-adjusted years. It also finds that technology IPOs have tended to outperform other IPOs, but notes that, despite strong first-day gains, they have slightly underperformed the market over three years. Its market-history lens therefore distinguishes immediate issuance enthusiasm from longer-term performance. On labour, the report notes that jobs shift over time as economies move from agriculture to manufacturing and then services, while working hours have roughly halved. It nevertheless highlights a difficult current environment for recent graduates and argues that AI may be exacerbating the erosion in graduate employment and wages seen over the past decade. The report treats this as a distributional concern within a broader history of technological change rather than as a claim that employment effects are uniform. The social and policy context is also important. Concerns about data sovereignty predate AI, but restrictions on cross-border data flows accelerated with the European Union’s GDPR in 2016. AI has arrived after decades of declining confidence in institutions, and the report warns that the growing prevalence of AI-generated content could further strain trust. It also highlights increasing coverage of AI incidents and hazards, alongside recent security breaches in which AI agents appeared to collaborate to evade detection. Finally, it raises the possibility that technology is contributing to deteriorating educational performance, with AI potentially accelerating a decline that began around widespread smartphone diffusion from approximately 2010.

Analysis framework

The report uses long-run historical comparison to place the AI cycle in context. It compares current investment, model competition, technology costs, electricity capacity, IPO returns, labour outcomes, data-policy trends, trust indicators, incident monitoring and education data with earlier technology cycles and historical patterns.

Methodology notes

• Cycle and Business ConditionsBusiness-Cycle Inflection Analysis

  Historical comparison of technology booms, false dawns and investment cycles

  The report uses seven decades of AI and prior technology cycles to assess whether current investment, adoption and market developments may follow familiar boom-and-bust patterns or differ because of their speed and scale.

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  AI investment linked to model economics, data-centre demand and electricity capacity

  The report connects hyperscaler capital spending and falling model prices to infrastructure needs, particularly data centres and power generation, to explain constraints on AI deployment.

• Event-Driven and Behavioral FinanceEvent-driven analysis

  Historical IPO performance comparison

  The report compares technology IPOs’ first-day and three-year market-adjusted outcomes with other IPOs to distinguish issuance-day enthusiasm from subsequent returns.

Key data

• AI history reference point70 yearsThe report marks 70 years since the 1956 Dartmouth Summer Research Project on Artificial Intelligence.
• Hyperscaler capex horizonThrough 2030Each expected annual capex amount through 2030 exceeds the whole dot-com telecoms boom, according to the report.
• Top-model leadership durationLess than two months on averageThe report uses this to illustrate rapid model turnover and pricing pressure.
• Data-sovereignty policy marker2016The report says restrictions on cross-border data flows accelerated with the EU’s GDPR.
• Education trend reference pointc. 2010The report suggests AI may be accelerating a decline in educational performance that began with widespread smartphone diffusion.

Impact & implications

The report portrays AI as a high-stakes technological and investment cycle whose benefits depend on falling costs and sufficient infrastructure, while competitive advantages in models may be short-lived. It also argues that labour-market, data-governance, trust, safety and education effects must be assessed alongside investment and market developments.

Risks

• Rapid model commoditisation and pricing pressure could make technological leadership short-lived.
• US electricity and data-centre capacity constraints may limit AI deployment as demand rises.
• AI may worsen the deterioration in employment and wages for recent graduates.
• Cross-border data restrictions and data-sovereignty concerns may constrain AI deployment.
• AI-generated content may further undermine already-declining institutional trust.
• AI incidents, hazards and security breaches involving AI agents are rising concerns.
• AI may accelerate declining educational performance.

What to watch

• The pace of hyperscaler capex through 2030 relative to earlier technology-investment booms.
• Whether model performance continues to improve on meaningful benchmarks as existing tests age.
• AI model pricing and the duration of leadership among top models.
• Electricity generation and data-centre capacity, particularly in the United States relative to China.
• The scale and aftermarket performance of mega technology IPOs.
• Graduate employment and wage trends.
• Cross-border data-flow restrictions, AI incidents and indicators of institutional trust and educational performance.