More severe than the internet bubble: Token consumption plunges 20%, the gap between AI investment and sales growth reaches 46%

According to Dongcha Beating monitoring, the Silicon Data LLM Token Consumption Index, which tracks users’ actual computing power expenditure, has fallen nearly 20% from its May high. This abrupt stop in a period of high growth momentum sends a key warning to investors: large model vendors may be losing pricing power in front of cost-sensitive customers, and it also raises doubts in the market about the ultimate return on investment from hundreds of billions of dollars in AI capital expenditures.

Disagreements between bullish and bearish camps have since intensified. The bear camp points out that Allianz Research data shows the growth gap between AI investment and sales has reached 46%, with the degree of imbalance exceeding the 32% seen when the telecom bubble burst in 2001. The bull camp counters that although the average token price has plunged by 90% since 2023, total spending has still grown by nearly one time, meaning the index’s decline is merely the structural digestion of lower pricing driving consumption, and that the return on investment during the inference phase is, over the long run, far more optimistic than during the training phase.

Strengthening policy and regulation is turning into implicit compliance costs for enterprise users. Washington has imposed more stringent policy reviews on the distribution of frontier models and cross-border access (such as release reviews for OpenAI and geopolitical export controls on Anthropic models), together with the EU AI Act’s strict compliance requirements for top-tier models, which places a heavy policy burden on leading platforms. To avoid geopolitical and compliance risks, corporate CFOs now have more rational reasons to proactively divert workloads toward mid- and smaller lightweight models that carry no regulatory burden.

Subtle changes are also emerging in the hardware chip sector. Although orders for top-tier GPUs and high-bandwidth memory HBM are already fully booked through 2026, and any meaningful easing of supply-and-demand is not expected until 2028, the market’s main purchasing force has begun shifting from training chips to inference-optimized hardware, and the winners’ structure is being reshuffled.