AI as a Macro Force

MFO Insights
back to Perspectives
08/2026

AI as a Macro Force

Artificial intelligence has moved from narrative to national accounts. It has become a dominant force behind global capital markets, partly masking commodity shocks and far-reaching geopolitical shifts. Its most visible macroeconomic impact is not yet a broad-based productivity boom, but one of the largest and fastest technology-driven investment cycles in US history.

This concentration has been reflected in the stock market through the dominance of tech mega caps, which account for a disproportionate share of US market capitalization and earnings. The eight largest stocks in the S&P 500 Index, all AI-exposed tech mega caps, account for about 35 percent of the S&P 500 Index's market capitalisation.

At the same time, the AI-related capex1 boom has become a real macroeconomic force and one of the main drivers of US investment demand. Because the AI-buildout heavily relies on imports it has contributed to soaring imports and a wider US trade deficit, dampening the net effect on GDP. However, Morningstar estimates that AI investments still contributed 0.4 to 0.6 percentage points to US GDP growth of 2 percent in 2025 depending on how IT equipment investment is measured.

The contribution is expected to grow further. Apollo notes that consensus expects hyperscaler2 capex to rise from 1.4 percent of US GDP in 2025 to around 2.4 in 2026 and 3 percent from 2027 to 2029. That is more than twice the peak of the telecom and fibre buildout, which topped out at 1.2 percent of GDP in 2000, but still less than half the housing boom, which peaked at 6.6 percent of GDP in 2005. In absolute terms, commonly cited estimates put US hyperscaler capex near $800 billion in 2026, while Goldman Sachs Research estimates global AI-related investment at just over $1 trillion.

‍Hyperscaler capex expected to reach 3 percent of GDP

Hyperscaler capex expected to reach 3 percent of GDP
Source: Apollo

‍

The current spending spree has therefore delivered a powerful demand-side boost. In addition, AI-related equity gains have indirectly supported consumption through a positive wealth effect, as higher stock market valuations made households feel richer. This is consistent with the decline in US household saving rates.

Conversely, an end to the investment boom would slow growth directly through weaker investment demand and indirectly through falling equity prices and weaker consumption as the wealth effect reverses. Given today's low saving rates, households appear particularly vulnerable to such valuation shocks. BCA estimates that a 10 percent decline in US equity wealth could reduce consumption by 0.9 percent of GDP. Because US technology stocks dominate global equity markets, a correction would likely have global repercussions, including for European investors.

‍

From a demand boom to a productivity boom?

The first-order effect of AI is capital deepening and a positive demand shock. The productivity acceleration may come later, while the capex is happening now.

Estimates of potential future productivity gains vary widely. Optimists see AI as a general-purpose technology comparable to electricity or the internet. Federal Reserve researchers argue that generative AI has characteristics of past transformational technologies including broad applicability, continual improvement and the ability to generate complementary innovations. The OECD estimates annual productivity gains of 0.4 to 1.3 percentage points in highly exposed economies with broad adoption, such as the US and UK, and lower gains in countries with a less AI-exposed sector mixes or slower diffusion.

Daron Acemoglu is more cautious. He estimates that AI may raise total factor productivity by no more than 0.66 percent over ten years, and potentially less than 0.53 percent once harder-to-automate tasks are considered. His argument is not that AI is useless, but that many applications improve efficiency without transforming production at scale.

Current evidence supports elements of both views. Micro studies show meaningful productivity gains in coding, customer support and knowledge work, but aggregate productivity data have not yet shown a clear AI-driven step change. Possible explanations include slow adoption, early-stage implementation, measurement issues and dilution effects when displaced workers move from high-productivity sectors into sectors less suited to digitization.

‍

Rational exuberance

The scale of the investment boom stands in sharp contrast to the more cautious estimates of future productivity gains, as well as to doubts about whether frontier model providers, hyperscalers and other layers of the AI ecosystem will be able to monetise their investments. Some data also suggest that token demand growth is slowing after the initial phase of "tokenmaxxing," when companies experimented with broad and often expensive AI usage. The risk of a bubble therefore looks high.

Falling Token Expenditures: A Sign of Slowing Demand?

Source: Bloomberg, MFO

‍

However, this does not mean that the AI buildout is necessarily irrational. The Bank for International Settlements (BIS) models the AI boom as a winner-take-most contest in which firms overcommit capital to secure early scale. As a result, it estimates that AI investment may exceed the socially efficient level by around 1.5 in a conservative baseline and by around three times under less elastic demand. In other words, the investments may be individually rational for each firm but collectively excessive.

If AI turns out to be a bubble, that does not mean that it will leave no lasting economic impact. In Speculative Growth and the AI "Bubble" Caballero (2026) argues that AI valuations need not be viewed in binary terms, either fully justified by fundamentals or purely irrational. Installed capital can leave a productive legacy even if asset prices later correct and some investments are written off. But this depends on sufficient capital being accumulated before beliefs change, valuations fall, and the buildout ends.

History suggests that this sequence is familiar. Railways, electricity and the internet all generated overinvestment before their full productivity effects became visible. The San Francisco Fed draws a parallel with electrification: the technology existed long before it transformed productivity because firms first had to redesign production processes around electric power. AI may follow a similar path. Automating existing tasks is only the first step; larger productivity gains will likely require firms to reorganise workflows, business models and labour allocation.

The current exuberance therefore contains both the potential for a lasting productive legacy and the potential for disappointment, at the company level and at the macro level, if expectations for AI-related earnings and productivity gains fail to materialize. Or, as Wachter and Wachter (2026) put it: "If the boom fails to materialize, the current buildout will be the largest misallocation of capital in history."

Risks: shifting from earnings to balance sheets

The AI-boom has started to penetrate all corners of financial markets. Apollo estimated in May that about half of all investment-grade bond issuance, more than a third of high-yield issuance and 87 percent of VC financing in 2026 to date (as of May) was AI-related. And financing needs are only beginning to rise. Goldman Sachs expects global AI-related investment to exceed $1 trillion in 2026 and to continue rising thereafter, resulting in cumulative additional capital investment of $4 to $8 trillion over the next five years.

This has important implications. First, financing needs are growing rapidly, which could eventually push up the cost of capital and interest rates. Second, a growing share of global capital is being absorbed by a single investment theme, making it less available for other sectors and projects. Third, AI-related investments could turn into a systemic financial risk.

AI Is Penetrating Financial Markets

Source: Apollo

‍

For now, the risks to the financial system still appear limited and are primarily equity-related valuation risks. But the situation could change as financing shifts from internal cash flows and equity toward public debt, private credit, vendor financing and long-term lease structures.

The US corporate bond market is about $11.7 trillion in size, according to SIFMA. If half of the projected $4 to $8 trillion AI buildout were financed through public debt issuance, AI-related issuance could increase its share of the US corporate bond market several-fold from its current mid-single-digit percentage. The effect could be even stronger in the smaller private-credit market.

Financing is expanding rapidly and increasingly shifting from equity financing, previously funded largely through operating cash flows, towards debt and vendor financing. These arrangements often take the form of complex, circular and nested structures involving off-take agreements, long-term lease commitments, loss guarantees and securitisations. As a result, part of the risk associated with the AI infrastructure buildout is being transferred onto the balance sheets of chip vendors such as Nvidia and Broadcom, hyperscalers such as Amazon, Microsoft, Alphabet and Meta, and ultimately credit investors.

Large technology companies are leveraging their strong balance sheets to accelerate the deployment of capital and reduce financing costs. In doing so, they provide frontier model developers such as Anthropic and OpenAI with access to data-center capacity, while helping data-center developers pre-finance both chip purchases and the construction of new facilities.

If the boom succeeds, these firms will benefit both as vendors and as investors. If it falters, they could lose on both fronts, as part of the risk has effectively been transferred onto their balance sheets. However, debt levels in the IT sector remain very low relative to earnings and are, in fact, lower than those of the broader developed equity market. Hyperscalers' balance sheets are generally large, diversified and healthy and appear capable of absorbing even a complete write-off of their stakes in Anthropic and OpenAI should frontier AI monetization ultimately disappoint.

The bigger risk is strategic. A failed AI monetisation would imply weaker future cloud revenue, slower enterprise AI adoption, stranded data-centre capacity, lower GPU demand and stress in leasing, project-finance and private-credit structures. Nvidia's key vulnerability is that today's extraordinary GPU demand normalises. And because many commitments are implicit, including backstops, lease obligations and loss guarantees, stated debt levels may understate effective leverage in parts of the ecosystem.

‍

A bumpy road ahead

The enormous infrastructure buildout tied to the multi-layered AI ecosystem has become a key driver of US growth and financial markets. It may create lasting social value through future productivity gains, but investor returns are likely to be uneven. This transformation is marked by extreme uncertainty, signs of overinvestment in both scale and speed, and a path that is unlikely to be linear.

Additional risks, including tighter regulation and rising societal scepticism, appear only partly priced in. Investor expectations are therefore likely to shift repeatedly as they realign with reality, accompanied by market shakeouts and volatility.

The correction in semiconductor and other technology stocks through June and July may have been only a taste of what lies ahead. For now, markets remain supported by strong and surprisingly resilient earnings, reflected in record-high earnings per share for the S&P 500 Index. But this earnings momentum is unlikely to last forever.

The long-term, cyclically adjusted price-to-earnings ratio, which divides stock prices by average earnings over the previous ten years, climbed above 40 for the S&P 500 Index in August, approaching historic extremes and underscoring the market's vulnerability to a slowdown in earnings momentum.

S&P 500 Earnings and Long-Term Valuations Near Historical Highs

Source: Bloomberg, MFO

‍

The massive expansion of data centres continues to drive demand for chips and related infrastructure. However, doubts about the sustainability of AI-related capex could slow investment, dampen semiconductor demand and weigh on earnings. If that happens, current valuations could quickly become unsustainable. In addition, reported earnings are, in some cases, flattered by valuation gains on private AI investments, which could make any reversal self-reinforcing. We may not be in a valuation bubble, but we are very likely in an earnings bubble.

In a benign scenario, this could lead to a digestion phase. Investment slows but does not collapse. Earnings expectations are revised lower, while enterprise adoption gradually catches up with installed capacity. Valuations compress, but the infrastructure ultimately proves productive. This is the benign version of Caballero's rational bubble: some investments disappoint, investors earn less than expected, but society still benefits from a large and lasting capital stock.

In a more severe scenario, the AI boom ends as a classic boom-bust cycle. AI demand falls short of expectations, capacity catches up with fading demand, compute scarcity disappears, semiconductor margins normalise and hyperscaler capex slows sharply. The impact would extend from suppliers to data-centre developers, power and cooling infrastructure providers, utilities and ultimately the broader equity market. The macroeconomic effects would follow through weaker investment demand, lower corporate profits and a reversal of the positive wealth effect that has supported consumption.

If the boom continues unabated, it could evolve into a debt-powered bubble. Financing structures become increasingly aggressive, circular financing expands, private-credit exposure rises and investors extrapolate future AI cash flows far beyond realistic adoption paths. A confidence shock would then trigger not only an equity-market correction but also broader credit tightening. The BIS explicitly warns that debt financing, circular ownership structures and network linkages can amplify stress and allow problems at one firm to propagate through the system. This scenario still requires a further expansion of credit, but the early signs are already visible.

Which path ultimately materialises will depend less on technological progress than on monetisation and the evolution of investor beliefs. AI is already transforming investment spending, capital markets and economic growth. The unresolved question is whether future productivity gains and revenues will be sufficient to validate the trillions of dollars currently being committed.

‍

1 Capital expenditures‍
2 Hyperscalers are large technology companies that build and operate vast networks of data centres to provide cloud computing, data storage and networking services on a global scale. They are also the primary investors in AI infrastructure. The major US hyperscalers typically include Amazon, Microsoft, Google, Meta and Oracle.

‍

‍

Download

Nadja Bleuler

Chief Economist, Partner
About MFO
Marcuard Family Office was founded in Zurich in 1998 and is an independent, partner-led multi-family office.
Why?
As a family office, we take care of all relevant wealth issues, leaving you time for other things.
When?
Whether you are selling your business, restructuring the current situation, or overcoming difficult family circumstances, the associated challenges can be overwhelming. As experienced advisors, we professionally guide you through transition phases as well as through calmer periods.
Legal Notices
Copyright © 2025 Marcuard Family Office in Zurich. All rights reserved
Data privacy | Imprint | Disclaimer