AI equities: Why the scepticism may be warranted

Over the summer, we have seen some scepticism towards AI-related equities, leading to a notable rotation out of tech stocks and into cyclical, defensive and value-oriented sectors. Yet AI is no longer just an equity market story. Since the launch of ChatGPT, the NASDAQ has gained 41% and the Magnificent Seven have grown to represent more than 30% of the S&P 500. At the same time, AI has become a dominant force in capital markets, accounting for 87% of venture capital funding, 49% of year-to-date net investment-grade issuance, and 38% of net high-yield issuance, according to Apollo. With the AI buildout now driving a meaningful share of US growth, we examine the sustainability of the underlying business models and whether the recent market scepticism is warranted.

Why AI is not software

A useful starting point is the distinction between the two things that happen in an AI data centre. Training a model is a large, one-off investment. Running it, known as inference, is an ongoing cost that scales directly with usage, since every query consumes compute. This is where AI diverges fundamentally from classical software, where the marginal cost of an additional user is close to zero. Inference has become the dominant cost: Roughly 90% of the electricity consumed by AI data centres today goes to inference rather than training. Although hardware progress has pushed the price of inference per token down dramatically, usage has grown even faster: longer prompts, reasoning models that "think" before answering, longer outputs. The net effect is that inference costs remain the central economic challenge for AI developers, and a key reason why the leading model companies are, for now, not profitable.

Competition is the real constraint on pricing power

Data collected by the OECD illustrate how crowded the market has become. The number of developers offering text-based AI models rose from fewer than five in early 2023 to 68 by February 2025, while the number of active models increased from a handful to around 790. At the top of the market, performance curves for the leading providers track each other closely, and no single player has managed to establish a durable lead. This matters because switching costs in AI are unusually low. Unlike traditional enterprise software, where changing vendors means retraining staff and rebuilding workflows, the core AI product remains, in essence, a chat interface. When one provider releases a stronger model, users migrate, which keeps competitive pressure correspondingly high.

Frontier models are best understood as infrastructure with an unusually short useful life. The value must be extracted before the technology is obsolete. This is the opposite of the classical software playbook, in which a firm builds a product, locks in customers, raises switching costs and harvests margins for years. In frontier AI, continuous multi-billion-dollar investment is required simply to keep one's place.

Open-weight models and local AI

A further force makes the competitive picture even more demanding: open-weight models. When a developer publishes model parameters openly, anyone can download and run the model free of charge. According to OECD benchmark data, the best open-weight models have tracked the best closed models with a lag of only around six months of development time. Publishing capable models free of charge suppresses willingness to pay across the market and undercuts the business models of developers who charge for access.

A structural factor that receives remarkably little attention is local AI. While training a serious model requires a data centre, inference does not. In principle it can run anywhere with sufficient processing power, including a laptop, a phone or a car. Since roughly 90% of AI data-centre usage today is inference, even a partial shift in that direction would mean that future data-centre demand looks quite different from what the current buildout is designed for.

Key takeaways

Taken together, the picture is this: frontier models are expensive to build, they depreciate within months, and they face growing competition not just from each other but from free, open alternatives. We find it hard to see where durable excess margins in AI model development are supposed to come from. Layered on top sits the prospect of local inference, which questions the demand assumptions behind the data-centre buildout itself. 

None of this implies that AI will fail to create enormous value. It almost certainly will. The open question is who captures it. History suggests that transformative technologies and profitable business models do not always coincide, and for investors underwriting today's AI expansion, that distinction deserves more attention than it currently receives.