The US economy: A cold hard look
Following the government shutdown on Wednesday, and the start of a new Federal Reserve rate-cutting cycle, the US equity market finds itself at a crossroads.
On one hand, there is valid context for the need of monetary easing. The Dallas Fed manufacturing survey showed contracting employment, while consumer confidence fell again in September with the labour differential narrowing, pointing to rising joblessness. The August JOLTS report speaks further to this – see Figure 1. Job openings edged up slightly (but only in non-cyclical sectors), while quits fell to a cycle low of 1.9%, and the vacancies-to-unemployed ratio slipped back below parity. Together, these indicators reconfirm the lingering “low hiring, low firing” regime.
Figure 1 – A continuation of the “low hiring, low firing” regime
Parallel to this, Fed Chair Jerome Powell recently described US stocks as “fairly highly valued,” a euphemistic phrasing for a dilemma that stems from a single driver: AI. For the most part, a full economic slowdown has been avoided simply due to the ongoing AI boom supporting growth via capital expenditures (“capex”). Tech-related stocks now represent nearly half of the S&P 500’s market capitalisation, singlehandedly pulling the broader index higher. Conseuqently, this has had significant household wealth effects, a likely reason for the resilient US consumer.
There is little doubt that AI will determine the direction of US equity markets over the next 6–12 months. The question is whether this direction can be predicted on current data. Markets can often look steady even as the pressure builds beneath the surface. History proves this, as a retrospective look at bear markets rarely define a single catalyst.
Lessons from the past
On a cyclically adjusted basis, the P/E ratio for US growth shares is not much lower than its prior peak in March 2000, a time when telecom capex drove valuations to unsustainable levels. When investment eventually slowed, the lack of earnings caused the bubble to burst – see Figure 2.
Figure 2 – The so-called “fairly highly valued” market
On the surface, this repeating peak makes sense as the setup rhymes: Telecom capex reached its apex during the race to build out fibre-optic networks. At the time, spending by wireline, wireless, and cable broadband providers reached 1.1% of US GDP. Today, capex by AI hyperscalers is about to surpass that level as a share of GDP – see Figure 3.
Figure 3 – AI hyperscaler capex (as a share of US GDP) is set to surpass the prior record highs of the telecom industry
There is, however, a key difference between the two regimes. The telecom buildout primarily created domestic infrastructure, with most of the benefits accruing within the US. By contrast, a large share of the capital equipment required for today’s AI data centres is imported. This creates an offset: As shown in Figure 4, periods of strong private investment in computers and communication equipment have generally boosted real GDP growth, but often alongside a negative contribution from net exports.
Figure 4 – Fixed investment in AI-related hardware are partly offset by imports
Telecom infrastructure eventually delivered significant productivity gains and GDP growth, as the networks laid across US soil formed the backbone of the digital age. By contrast, much of today’s AI investment is concentrated in imported capital equipment, from high-end semiconductors to server components. While this still supports technological innovation, the domestic multiplier for the US may be weaker. As could be expected, the economic benefits of AI will begin to rely less on the physical act of building data centres and more on how effectively US firms and households integrate AI into workflows.
A survey from the Census Bureau suggests that the adoption rate among companies is weakening. However, larger companies do expect to change this within the next six months – see Figure 5. There are clear use cases for AI (programming, customer service etc.), but most companies are still in the exploratory stage of leveraging GenAI. A recent study by McKinsey noted that while nearly 80% of companies have deployed some form of generative AI, the same percentage has seen no benefit to their bottom line from it. Another study by researchers at MIT found that 95% of companies have seen no material gain from their AI investments.
Figure 5 – AI adoption rates have stalled, but big companies aim to change that
When will the AI capex boom end?
McKinsey expects that between 2025-2030, investment in AI-related data centres will total approximately $5.2 trillion. The capex intensity is a structural component of the AI trend, meaning that at least for the short to medium-term, there is no way to pinpoint when it should end. Capex during the telecom boom was tempered by the decades-long lifespan of fibre cables, and an understood terminus: Once the last mile of cable was laid. The same is probably not true for the latest chips bought for training in the chase for Artificial General intelligence (“AGI”). Supply remains constrained, and the rapid pace of innovation shortens the effective lifespan of GPUs and CPUs to such an extent that they are often redundant by the time it is installed in the last crop of data centres.
This fact has benefited the pockets of chipmakers immensely and Figure 6 shows how eerily similar the pattern is between the revenues of major chipmakers, and the capex of the major hyperscalers. Evidently, while the market is enjoying paper gains, the actual cash flow benefits of AI have remained largely locked within this feedback loop.
What could break this circle?
Figure 6 – Chipmaker sales are synonymous with hyperscaler capex
Risk one: The adoption paradox
The late-1990s IT boom was a strong tailwind for jobs in technology. Employment in computer manufacturing, networking, and IT services grew rapidly until the dotcom crash. Afterward, jobs in computer production never recovered, and IT services employment took nearly a decade to regain momentum.
The current AI regime has so far skipped the labour boom entirely. Employment in IT services briefly accelerated during the pandemic but has since rolled over. Since late 2022, coinciding with the release of ChatGPT, IT services payrolls have stagnated. Even private payrolls most directly tied to data centres have been roughly flat for two and a half years. Unlike the 1990s, where the technology boom warranted a greater labour demand, today’s AI investment is capital-intensive and labour dilutive.
The recent hiring data underscores this fragility. The economy added just 22,000 jobs in August, while June’s figure was revised down to a net loss of 13,000. July’s 79,000 gain looks strong in comparison with the subpar readings of May, June, and August, but remains far below the averages of prior years – see Figure 7. On a three-month moving average basis, job creation has slipped to 29,000 from 35,000.
Figure 7 – The August readings suggest recession risk
The newest data point adds further credence to the historical pattern mentioned in our previous synopsis: Since the 1960s, every time annual growth in nonfarm payrolls has slipped under 1%, the economy has either already been in recession or was about to enter one.
While this hypothesis remains simply conjecture on limited data, it would be an ironic outcome if a stimulative regime serves primarily to boost AI capex even higher, drive labour demand even lower, and stretch the already struggling consumer even further. Companies have made impressive gains in earnings through AI so far, but is this sustainable if the actual consumers are being displaced to achieve this? It begs the question: Would the short-term US economy benefit from pausing AI demand and hiring humans again?
Risk two: Does AI have a power bottleneck?
AI requires vast amounts of electricity in both its training and use by the public. In 2023, US data centres accounted for 4.4% of total electricity generation, and the Department of Energy projects this share could triple by the end of the decade – see Figure 8.
Figure 8 – Power demand for data centres is on a one-way trajectory upward
Yet despite these soaring requirements, the expected energy investment boom has not materialised. Utilities and independent power producers have been slow to expand generation capacity, and the transmission bottlenecks remain unresolved. This mismatch raises a key risk: If economic activity rebounds, energy supply could tighten, leading to higher costs and potential delays in the accelerated adoption of AI across the workforce. In contrast to the telecom boom of the 1990s, where infrastructure spending paved the way for productivity gains, today’s AI expansion risks being throttled by insufficient energy investment – see Figure 9.
Figure 9 – No major shift in energy capex yet
Investment takeaway
US equities lean heavily on the AI capex boom, echoing the late-1990s telecom cycle. History shows that markets often falter once capex intensity peaks, as earnings momentum fades and investment slows. Today’s AI cycle may be even more fragile: Chips are short-lived, largely imported, and labour demand remains weak, potentially as a structural consequence of AI adoption. If the capex loop between hyperscalers and chipmakers break, so too could the rally underpinning US equities – see Figure 10.
Figure 10 – A market driven by capex, will fall with capex
If you are interested in finding out more about how cognisance of the macroeconomic backdrop impacts our investment decision making process, connect with Integrity Asset Management and let us help you navigate your investing journey.
For more information on this synopsis or to discuss solutions provided by Integrity Asset Management, please contact us at:
Tel: (021) 671 2112
Cell: 072 513 2684 / 084 601 1025
E-mail: nic@integrityam.co.za / herman@integrityam.co.za
Source: Bloomberg, 30 September 2025