AI: Between an Investment Boom and Pressure on Returns
PERSPECTIVES | AI Special

- Falling usage costs are opening up new applications for AI. Broad-based productivity gains, however, have yet to materialise.
- Infrastructure investment is already supporting the US economy while increasing its dependence on the investment cycle.
- For investors, the key question is which companies can translate substantial spending into sustainable returns.
The economic success of artificial intelligence is unfolding on two different timelines. Data centres are being built, chips ordered and software developed. This spending is already generating substantial demand. Yet most of the earnings needed to justify these massive investments over the long term still lie in the future. The opportunities and risks of the AI boom emerge precisely from the gap between these two timelines.
For capital market investors, the crucial questions are therefore how technological benefits can be monetised and which companies will profit. A technology can achieve widespread adoption while leaving behind providers whose investments fail to pay off. The data in our latest Assenagon analysis reveal AI’s considerable potential for application and disruption. They also show how much depends on its successful implementation.
Falling Costs Broaden the Range of Applications
AI is becoming cheaper and more capable. A comparison of models shows that a level of performance achievable at the end of 2024 can be delivered in 2026 at costs up to 70 times lower. Although this analysis is based on a standardised performance test whose results cannot be applied to every use case, the economic direction is clear: processes for which AI previously offered little financial benefit may become commercially viable. [1]

However, lower costs can also generate additional demand. Even as AI becomes cheaper to use, increased consumption can absorb the savings. This principle, known as the Jevons paradox, explains why more efficient models and further infrastructure expansion can go hand in hand. The investment boom itself contributes to growth.
AI infrastructure expansion is already supporting economic activity, even without necessarily delivering measurable productivity or efficiency gains. Purchases of equipment, software development and data centre construction all increase investment demand. The short-term boost to growth therefore comes before any potential longer-term improvement in productive capacity.
The US data examined illustrate the scale of this effect. From the first quarter of 2025 to the first quarter of 2026, the selected categories of IT investment accounted for an average of around 40 per cent of reported real GDP growth. After adjusting for net IT imports, the share was approximately one quarter. The scope of these categories matters: they also include investment unrelated to AI. These figures therefore measure the contribution of selected IT investments rather than a separately identifiable AI effect. [2]

Sources: Federal Reserve, FEDS Notes, 17 July 2026; BEA/FRED; own calculations. Contributions to growth: quarterly data from 2022 to Q1 2026, annualised; GDP data as at June 2026. IT investment as a share of GDP: annual figures through 2025, data as at 30 September 2026. The categories also include spending unrelated to AI.
Massive investment in AI is creating an economic dependency. If companies scale back their expansion plans, an important source of demand could weaken. This would not require the technology itself to fail. More cautious investment decisions alone would be enough to slow momentum. At the same time, bottlenecks in energy supply, construction and equipment could constrain the delivery of planned capacity. Local opposition is also growing. During the US midterm elections, voters supported restrictions on data centre expansion in 46 out of 47 local ballot measures. [7] Alongside physical bottlenecks, permitting and location risks could therefore delay planned capacity or increase its cost.
Historical Comparisons Highlight the Potential Downside
Nominal private investment in IT equipment and software reached approximately 4.5 per cent of US GDP in 2025, slightly exceeding the 2000 peak of 4.4 per cent. Following the technology boom at that time, this share fell by almost one percentage point of GDP by 2002. Data centre construction is not included in this time series. [2]
This historical comparison provides a sense of the significance of today’s investment cycle. However, a similar investment share does not establish either a comparable degree of excess or an imminent downturn. Such a conclusion would also require comparable business models, financing structures and earnings prospects. Above all, historical experience shows that even a technological transformation that succeeds over the long term can involve substantial fluctuations in investment activity.
Adoption Is Growing Faster than Measurable Benefits
AI adoption is advancing rapidly, outpacing measurable productivity gains. In the euro area, the proportion of employees using AI at work rose from 26 per cent in 2024 to 52 per cent in 2026. At the same time, the vast majority of companies surveyed in an international study reported no productivity impact over the preceding three years. [3]

These findings are not necessarily contradictory. AI can speed up individual processes, while redesigning entire workflows and integrating the technology take longer. Companies need to integrate applications, train employees and check results. Time saved in individual tasks therefore does not automatically translate into greater efficiency and higher output across the company. Meanwhile, a European Investment Bank study has already identified initial efficiency gains from AI, amounting to 4 per cent higher revenue per employee. Positive productivity effects are also being mentioned increasingly often in listed companies’ earnings calls. [3, 4]
Falling usage costs could make AI viable in more processes and spread efficiency gains more widely. This, in turn, could create earnings potential to support further AI investment. The eventual scale of the macroeconomic impact remains uncertain.
Early Changes Are Emerging for Young Workers
The studies examined suggest that young workers face particular pressures in the US labour market. In occupations with a high proportion of tasks that AI can automate, employment among 22- to 25-year-olds is falling, primarily because of reduced hiring. No comparable decline is evident among experienced workers. [5]
These findings do not yet establish widespread AI-driven job losses. A central question for the future will be whether, and to what extent, AI replaces employees altogether or assists them in their work. Public acceptance of further expansion is also likely to depend on this distinction. The limits of that acceptance are already becoming apparent in local votes on data centres. A greater focus on complementing human labour and delivering measurable productivity gains could strengthen public support and, over the long term, the earnings underpinning further expansion.
High Investment Increases the Pressure on Future Earnings
The financial dimension is particularly apparent among the major cloud providers Alphabet, Amazon, Microsoft, Meta and Oracle. In 2025, their capital expenditure averaged approximately 32 per cent of revenue. In June, the Bank for International Settlements (BIS) estimated that this ratio would reach around 48 per cent in 2026. It also projected gross debt issuance by the five companies of approximately USD 175 billion for 2026. Credit default swap (CDS) spreads in the AI sector have recently developed less favourably than those of BBB-rated peers. If future earnings disappoint, these spreads could widen further. [4]

For more on hyperscalers’ record debt issuance and its impact on global bond indices, see our latest Assenagon Credit study.
Financing and Sales Can Depend on the Same Customer
Close financial ties within the industry add another dimension. Under so-called circular financing arrangements, a chip or cloud provider may supply capital to a customer while also selling that customer hardware or computing capacity. In the BIS sample of 1,246 AI companies examined, such structures accounted for approximately 16 per cent of deals between AI companies from 2021 to 2025. Their share of disclosed deal value was 46.4 per cent. [6]
One example of circular financing is the agreement between Microsoft, Nvidia and Anthropic announced in November 2025. Microsoft and Nvidia committed to investments of up to USD 5 billion and USD 10 billion respectively. At the same time, Anthropic committed to purchasing USD 30 billion of cloud capacity from Microsoft and additional computing capacity of up to one gigawatt based on Nvidia systems. Both investors are therefore also suppliers: Microsoft through its cloud services and Nvidia through the chips powering that capacity. [8]
Such agreements can enable expansion and are not, in themselves, evidence of unsustainable business arrangements. Circularity nevertheless creates risks. Providers effectively finance their customers and benefit from their purchases. Both sales and the value of their equity stakes depend on the same customer’s success. If that customer’s earnings subsequently disappoint, orders and equity values could fall simultaneously, while financing becomes more expensive. These feedback effects can amplify the investment cycle.
For Capital Market Investors, Sustainable Earnings Are What Count
The data paint a nuanced picture. AI has the potential to improve companies’ productive capacity. Its expansion is already supporting demand. At the same time, the revenue providers will need to generate in future is rising. A successful technological transformation can therefore produce very different investment outcomes within the industry.
For equities, the key questions are which companies can generate lasting additional profits from using or providing AI, and which expectations are already reflected in their share prices. Companies applying AI to improve existing processes may also benefit. The investment and productivity data examined here do not support a blanket assessment of the attractiveness of AI stocks.
For corporate bonds, the focus shifts to the sustainability of cash flows, financing needs and commitments to business partners. From a macroeconomic perspective, the main question is whether the investment impulse broadens or whether growth becomes increasingly dependent on a small number of very large spending programmes. In addition, substantial new bond issuance by hyperscalers coincides with already high government financing requirements. This could intensify upward pressure on long-term bond yields.
In this environment, diversification requires an understanding of shared economic dependencies. Different companies and asset classes may be exposed to the same investment cycle. For example, an equity investment in an equipment supplier and a bond issued by a cloud provider may both benefit from continued capacity expansion. The extent to which these risks interact within a portfolio requires a separate assessment in each case.
The decisive test of the AI boom therefore lies in execution: lower usage costs must translate into economically viable applications, and substantial investment must generate sustainable earnings. How successfully this happens is likely to determine the resilience of today’s growth impulse and where long-term opportunities emerge for investors.
Sources and data reference dates
[1] Epoch AI, model costs and GPQA Diamond benchmark; own calculations. Prices as at 3 August 2026; model releases included up to July 2026.
[2] Federal Reserve, FEDS Notes, 17 July 2026; BEA releases of 25 June and 30 September 2026, accessed via FRED; own calculations. GDP data as at June 2026; IT investment as a share of GDP: data as at 30 September 2026.
[3] ECB, Consumer Expectations Survey, 26 August 2026; Yotzov et al., Firm Data on AI, March 2026, survey conducted between November 2025 and January 2026; EIB Working Paper 2026/02.
[4] BIS, Annual Economic Report 2026; S&P Global Market Intelligence; SEC filings/company reports. Data as at June 2026.
[5] Stanford Digital Economy Lab/ADP, Canaries Dashboard, 23 September 2026, and study published in August 2026; Yale Budget Lab, 15 September 2026; Indeed/FRED, Federal Reserve, accessed on 15 October 2026.
[6] BIS, sample of 1,246 AI companies covering the period from 2021 to 2025; own presentation. Data as at October 2026.
[7] Ballotpedia, 2026 ballot measures on data centres, as at 13 November 2026; own analysis.
[8] Microsoft, NVIDIA and Anthropic, “Microsoft, NVIDIA and Anthropic announce strategic partnerships”, Official Microsoft Blog, 18 November 2025.
