For the past two years, the investment case for artificial intelligence has been remarkably simple. Demand appeared limitless, hyperscalers announced ever larger capital expenditure plans, semiconductor manufacturers struggled to keep pace, and investors concluded that the AI revolution would justify almost any valuation. That assumption is now facing its first real test. One of the most closely watched indicators of AI monetisation, the Silicon Data LLM Token Expenditure Index, has fallen almost 20% from its May peak. The index measures what users actually spend on AI tokens rather than simply tracking published prices, making it one of the few indicators providing a real-time view of demand for AI services. Its recent decline should not be ignored.
The question is not whether artificial intelligence remains transformational. It almost certainly does. The real question is whether the extraordinary investment currently flowing into AI infrastructure will generate returns commensurate with the capital being deployed. That distinction is becoming increasingly important. The index’s decline does not necessarily mean AI has become cheaper. Several explanations are possible. Model providers may be cutting prices to attract customers. Businesses may be shifting towards lower-cost models rather than the most sophisticated ones. Or buyers may simply be becoming more selective about what they are willing to pay for. Each interpretation carries very different implications.
The optimistic view argues that this is simply the natural evolution of a rapidly expanding market. Token prices have already fallen dramatically since 2023, yet overall spending has continued to rise as lower prices have encouraged broader adoption. From this perspective, lower pricing is exactly what should happen during the transition from early adoption to mass deployment. Lower costs stimulate higher usage, ultimately supporting further investment in chips, memory and data centres. That remains entirely possible. However, markets rarely worry about today’s revenues. They worry about tomorrow’s pricing power. If customers are becoming increasingly price-sensitive before AI has even reached full commercial maturity, investors must begin asking whether the industry’s future margins will justify today’s valuations. This is where the numbers become uncomfortable. According to Allianz Research, investment in AI infrastructure is growing roughly 46% faster than the revenues generated by its beneficiaries. During the technology bubble that culminated in 2001, the equivalent gap peaked at around 32%. History never repeats perfectly. But it often rhymes.
None of this suggests that demand for AI hardware is collapsing. Quite the opposite. High-end GPUs and advanced memory remain largely sold out well into 2026, with supply constraints expected to persist for several years. The issue is not whether infrastructure is needed. It is unclear whether the most profitable part of the AI ecosystem is gradually shifting. The market may be moving away from expensive model training towards inference, where models are already trained and simply process user requests. That evolution changes who captures the profits. The winners of the next phase may not be the same as those of the first.
At the same time, regulation is quietly becoming another headwind. Washington is tightening controls over advanced AI technologies, while Europe is imposing increasingly demanding transparency and compliance requirements under its AI Act. These measures do not directly regulate prices, but they increase deployment costs and may encourage businesses to adopt simpler, cheaper models where possible. Competition is intensifying as well. Chinese developers continue to improve rapidly, open-source models are becoming increasingly capable, and large corporate customers are behaving exactly as procurement departments always do: once a technology becomes essential, they negotiate aggressively on price. Technology leadership alone no longer guarantees pricing power.
For investors, this may be the most important lesson emerging from recent data. The AI revolution remains intact. Productivity gains could prove enormous over the coming decade, and demand for computing power is unlikely to disappear. But transformative technologies do not necessarily produce extraordinary shareholder returns for every company involved. Railways transformed the nineteenth century. The internet transformed the twenty-first. Many companies that built those revolutions never generated the profits investors originally expected. Artificial intelligence may follow a similar path. The companies creating the greatest technological breakthroughs may not ultimately be those delivering the highest investment returns. As always, innovation creates value. Capturing that value is a very different challenge.