The interplay of investment, valuation and leverage behind 2025’s most powerful market theme

Setting the Scene

Equity markets continue to climb on a mix of resilient earnings, rising AI-linked optimism, and a shift in interest-rate expectations. US and global indices hover near record highs, driven disproportionately by a narrow group of mega-cap technology firms. This concentration is what keeps the bubble debate alive: strong index performance is resting on an increasingly expensive, increasingly small cohort.

But the picture is more complex than a simple “bubble or not” judgment. Much of today’s rally still reflects real factors — robust earnings from dominant platforms, aggressive share buybacks, and a macro backdrop that supports higher multiples. The AI-productivity narrative adds another layer, pushing investors to reward companies seen as central to the next computing cycle.

This dynamic fuels the recurring question: “Is it too late to buy?” As commentators note, trying to call the top is a losing battle. Market cycles driven by technological inflection points can extend far longer than sceptics expect, often unwinding only when liquidity conditions or macro narratives shift meaningfully.The AI discussion introduces a second complication — what some analysts describe as a “double bubble”. The first is an industrial bubble: an enormous wave of capital expenditure on chips, data centres, cloud infrastructure, and energy capacity. As with railways or fibre-optic networks, some overbuilding is inevitable, but much of the investment will leave behind lasting economic assets.

Hyperscaler (e.g. Google, Microsoft, Oracle) capex trend, (%), 2022-2025, quarterly

Source: FT, Barclays

The second is a financial bubble: valuations in AI-exposed equities and private companies increasingly rely on optimistic assumptions about future dominance, margins, and monetisation. The gap between corporate AI rhetoric and current cash flows creates fertile ground for speculative excess.

Because these two bubbles overlap, AI has become a systemic, macro-relevant theme, not a niche technology story. Even if valuations correct, the aftershocks will extend well beyond the tech sector.

What Counts as a Bubble? 

The surge in AI investment has revived a classic debate: when does aggressive capital deployment become a bubble, and can some bubbles be beneficial?

Some industry leaders frame today’s environment as what Jeff Bezos once described as a “good bubble”— a period of short-term excess that accelerates long-term innovation. Overinvestment may appear irrational in real time, yet it can accelerate technological progress by pushing capital, talent, and infrastructure into emerging areas faster than normal market dynamics would allow.

Under this view, today’s cycle resembles an industrial bubble, which typically leaves durable assets behind. Advanced chips, data centres, and energy infrastructure may be expensive today, but they form the backbone for future growth. By contrast, a purely financial bubble inflates asset prices without building anything lasting.

Yet credible scepticism persists. Investor James Anderson, for example, draws parallels to the dot-com era, questioning whether AI adoption is scaling fast enough to justify commitments of this size — such as Nvidia’s long-term exposure to OpenAI. The concern is whether expectations are running ahead of fundamentals. Valuations add another layer. CAPE ratios multiples for leading AI players have reached levels reminiscent of the early 2000s.

Cyclically adjusted P/E10, or Shiller CAPE dynamics, 1970-2025, daily

Source: Shiller Data

However, unlike their dot-com predecessors, today’s giants operate with strong balance sheets, significant cash flow, and clear revenue pipelines. The emerging consensus is not that AI is definitively a bubble, but that it sits at the intersection of innovation-driven optimism and pockets of speculative excess. The key question is whether current investment translates into lasting productivity — or simply inflates prices.

Capital Flows and Private-Market Frenzy

Capital is pouring into AI at a scale reshaping both private and public markets. AI start-ups have raised approximately $161 billion in 2025, making up more than two-thirds of all US venture funding.

The rise of venture capital  investment in AI, 2016-2025

Source: FT, PitchBook

This capital is highly concentrated: the ten largest loss-making AI firms, including OpenAI, Anthropic and xAI, have added nearly $1 trillion in valuation over the past year despite limited revenues and, in some cases, negative gross margins.

Top 10 AI start-up valuation gains, $bn, Oct 2024 vs Oct 2025

Source: FT, Company reports

A defining feature of this cycle is the reflexive link between private and public markets. Late-stage private rounds influence expectations for listed semiconductor firms and cloud providers. Long-term compute-capacity agreements between private AI developers and public technology leaders, (e.g., Nvidia, AMD, Broadcom, Oracle) transmit private-market optimism directly into public-market valuations — a dynamic reminiscent of the dot-com era.

Mega-cap companies amplify these effects. Nvidia’s rise to a $5 trillion market capitalization — comparable to the combined GDP of Italy and France — illustrates the unprecedented concentration of value in AI infrastructure. Its massive order backlog, including an estimated $100 billion exposure to OpenAI, has raised questions about circular financing flows between venture capital, start-ups, and their suppliers.

Leverage is an increasingly relevant part of this ecosystem. Global data-centre spending is projected to exceed $3 trillion by 2029. Companies such as Meta have raised more than $26 billion in bonds to fund AI infrastructure. Start-ups including CoreWeave and OpenAI are tapping high-yield credit markets to finance growth. The short economic lifespan of GPUs — often under three years — compresses the period over which leveraged investments must pay off, increasing systemic vulnerability.Major institutions, from Goldman Sachs to the IMF, warn that valuation distortions are emerging, particularly in late-stage private rounds and among speculative public names. Yet industry leaders remain divided.

GDP vs Market Capitalization volumes, $tn, Oct 2025

Source: Goldman Sachs GIR, IMF

The Hidden Leverage

Beneath the surface, the AI boom is supported by an increasingly intricate financing structure. Start-ups rely on cloud providers for compute, while those same cloud providers invest in the start-ups — creating circular funding loops where growth relies heavily on continued investor confidence rather than end-user demand.

At the same time, traditional financing is often too slow for the pace of AI expansion. Firms are turning to private credit, vendor financing, and bespoke lending structures that offer speed but little transparency. Hyperscalers frequently provide credit to partners who then purchase their services, embedding leverage across the value chain.

Because much of this borrowing sits off-balance sheet, the true level of system-wide leverage is obscured. Private-credit vehicles, securitised loans, and structured financing arrangements mean liabilities accumulate in less regulated areas. Liquidity appears abundant — until it isn’t.

A deeper structural issue is the duration mismatch between long-term financing and the short life cycle of AI hardware. Chips and accelerators depreciate quickly; new architectures arrive roughly every two years. This mismatch raises the risk of stranded assets if computational demand stabilises or efficiency improvements reduce the need for cutting-edge hardware. These intertwined pressures leave the sector highly sensitive to funding slowdowns, technological shifts, or credit tightening.

AI Funding Loop scheme

Source: FT

The Cargo-Cult Effect

A growing tension in the AI cycle is the gap between corporate narratives and actual productivity gains. The 2025 McKinsey AI Adoption Survey found that while 88% of companies use AI in at least one business function, two-thirds are still in early experimentation. Most have not yet seen material improvements in operating income, though cost benefits are emerging in functions such as software engineering, manufacturing, and corporate finance.

The survey highlights that the companies most likely to benefit are those with clear strategies, defined efficiency goals, and rapid scaling plans. Superficial adoption — driven by peer pressure or hype — rarely produces lasting value.

This gap is what some commentators describe as the “cargo-cult problem”. The original concept referred to Pacific islanders who, after observing US military operations in WWII, attempted to recreate them through symbolic imitation. Applied to modern technology, it describes companies adopting AI tools without understanding how to deploy them effectively — mimicking the form without achieving the function.

Unsurprisingly, this divergence fuels investor anxiety. Even AI leaders acknowledge excesses: Alphabet CEO Sundar Pichai recently described current investment levels as containing “some irrationality,” noting that no company — including Alphabet — is immune to a potential correction.

Macro Overview: Looking Ahead to 2026–2028

The AI expansion is no longer just an industry story — it is now visible in macro data. The AI build-out has become a macro-relevant force, adding a measurable boost to GDP growth but also showing early signs of deceleration. Analysts outline a bear-case in which global data-centre capex falls by around 20%, driven by power constraints, slower permitting, and efficiency gains that reduce the need for continuous hardware expansion. The risk is not collapse but a cooling of an investment cycle that has been running at exceptional speed.

The macro exposure is unusually large. Over the past year, data-centre construction and AI-related equipment spending have contributed roughly 0.8–1.0 p.p. to US real GDP growth, an extraordinary impulse for a single capex category. Annual domestic spending of $150–170 billion has flowed into semiconductors, cloud infrastructure, and high-density facilities. A partial reversal of this impulse would soften sequential GDP prints heading into 2026, particularly if it coincides with weaker manufacturing momentum.

Balance-sheet strength provides important buffers. Hyperscalers operate with substantial cash reserves, giving them the flexibility to adjust spending without triggering broader stress. This distinguishes the current cycle from past tech booms, where aggressive borrowing amplified downturns.

Earnings sensitivity, however, is concentrated. A capex slowdown would disproportionately affect upstream suppliers — chips, accelerators, networking, and power-management hardware — where AI-related revenues can exceed 40–60% of total sales. Analysts expect most of the EPS impact to emerge in 2026–2027, particularly for firms tied directly to data-centre volume growth, while software and cloud platforms would see more moderate effects.

The question for 2026–2028 is whether enterprise productivity gains can absorb the cooling of infrastructure capex and transform the AI wave into sustained economic value creation. McKinsey estimates that Generative AI could create $2.6–4.4 trillion in annual global value across 63 use cases.

AI’s potential impact on the global economy, $tn

Source: McKinsey Global Institute

Productivity gains are expected in customer operations, marketing, R&D, and software engineering. If adoption continues at its current pace, AI could lift labour productivity by 0.1–0.6% annually by 2040. For 2026–2028, this suggests movement toward the upper end of a “Productive Boom” scenario.

BCG reports that enterprises are already reallocating IT budgets: increasing investment in AI, GenAI, and cloud while reducing spend in legacy categories. “Agentic AI” — automated intelligent assistants — is expanding into marketing, supply chain, and R&D. Some high-maturity firms are even repatriating workloads to on-premise servers to reduce cloud-cost volatility. These shifts point to a scenario between “Productive Boom” and “Air Pocket”, where firms push efficiency but remain cost-sensitive.PwC’s Global AI Jobs Barometer  highlights the macro stakes: AI could raise long-term global GDP by up to 15%, but this outcome depends on responsible adoption and regulatory trust. Without these, gains could fall to around 1%. This underscores that governance, ethics, and credibility will shape the next phase of AI far more than raw compute investment.

Investor Playbook

For investors, the central challenge is distinguishing between structural growth and speculative excess. A practical way to do this is through a “Bubble Barometer” — a set of signals that capture how economic fundamentals, financing conditions, and narrative momentum evolve in real time. Familiar faces provide a backdrop against which investors are better able to judge whether capital is being deployed into real productive capacity or is merely chasing frothy valuations.

For a start, hyperscaler capex remains one of the most helpful leading indicators. Because the AI boom is driven by the likes of Microsoft, Amazon, Meta and Google, rising capex generally signals strong underlying demand for computing power and data-centre capacity. Yet even a modest disruption can rapidly spread to vendors, start-ups and private credit lenders. Moreover, how companies characterise their spending — with phrases like “disciplined,” “front-end loaded,” or “customer-driven” — offers additional insight, since this language often reveals the optimism or unease that numbers alone cannot.

Along similar lines, the rise of private credit financing for data-centre projects offers another useful indicator. Growing reliance on this type of funding may signal that traditional capital sources are tightening, pushing companies toward increasingly complex lending structures. A jump in private credit can point to looming overcapacity, while a pullback is more often a sign of strain. Examining the terms around advance rates, covenants and lender protections helps reveal how comfortable lenders truly are with the risks they are taking on.

Moreover, energy-demand projections from the US Department of Energy are becoming ever more important, as energy is emerging as a key constraint on AI growth. When DOE projections rise in parallel with infrastructure expansion investors can have more confidence that projects have a long runway. But if forecasts decline or grid constraints emerge in certain geographies, it casts a shadow of doubt over the assumptions that underpin both investment and financing models. Because many financial models rely on the high utilisation of compute resources, a change in the energy outlook can significantly influence the expected returns.

Beyond the numbers, the broader narrative backdrop is also relevant. By comparing the extent to which companies invoke the AI theme during their earnings calls with the actual slice of revenue that can be tied directly to it, investors can assess whether performance continues to support the story. When AI mentions accelerate faster than the revenues that underpin them, sentiment has often gotten ahead of itself — a phenomenon that frequently precedes a de-rating.

AI has become a force powerful enough to drive markets, distort valuations and shape macro data, but the cycle is far from settled. Industrial overbuild, financial optimism and real productivity potential are all unfolding at once. The coming years will test whether the sector can transition from capital-intensive acceleration to steady, monetisable growth. Investors who focus on fundamentals — not just narratives — will be best positioned to judge whether AI is evolving into a durable engine of economic value or approaching its first true stress test.

Written by:

• Mariia Chasovnikova
• Claudio Di Nubila
• Demet Coşkun
• Viola Castoldi
• Davide Marsetti
• Andrea Rusconi
• Carlo Conte
• Samuele Scattolon