TSMC reported a 77 per cent increase in second-quarter earnings, yet its shares and the broader chip sector fell. The AI investment story has relied on frontier model-makers such as Anthropic and Open AI providing products well below cost, subsidising compute capacity for users while investors absorbed the losses. Most value flowed to the chip industry, the only segment reaping immediate returns.

That dynamic is shifting. Microsoft-owned GitHub moved to usage-based pricing in April, marking the "end of the token subsidy," argues Juan Correa at BCA Research. Companies are rationing tokens, driving adoption of cheaper, open-source Chinese models. Torsten Sløk at Apollo sees Jevons paradox at work: falling unit costs may spur total spending.

Yet commoditisation looms. Tyler Frawley at RBC Wealth Management warns AI could become like electricity — value accruing to application builders, not model producers. Morgan Stanley estimates hyperscaler capex will hit $1.2tn next year, while free cash flows collapse. Goldman Sachs sees peak adoption taking 15 years. Meta is already selling excess compute, hinting at malinvestment. Hyperscaler stocks have flatlined despite rising capex guidance, suggesting the AI trade's two pillars — profitability and near-term mass demand — are weakening.