More Severe Than the Dot-Com Bubble: Token Consumption Plunges by 20%, AI Investment and Sales Discrepancy Reaches 46%

According to Farsight Beating monitoring, the Silicon Data LLM Token Consumption Index, which tracks user actual computational power expenditure, has dropped by nearly 20% from its May peak. This sudden halt to the high-growth trend has sent a critical warning to investors: large model providers may be losing pricing power in front of cost-sensitive clients, raising doubts in the market about the ultimate return on investment of the multi-billion-dollar AI capital expenditure.

A deepening divide between bullish and bearish camps has followed. The bearish faction points out that according to Allianz Research data, the growth gap between AI investment and revenue has reached 46%, surpassing the imbalance level of 32% seen during the 2001 telecom bubble burst. The bullish camp, on the other hand, rebuts that while the Token average price has plummeted by 90% since 2023, total expenditure has nearly doubled, implying that the index's decline is merely a structural digestion after price cuts to stimulate consumption. Furthermore, the long-term investment return during the inference phase is expected to be much more optimistic than during the training phase.

The strengthening of policy regulations is translating into implicit compliance costs for corporate users. Washington has imposed more stringent policy reviews on the distribution and cross-border access of cutting-edge models (such as scrutiny over OpenAI releases and geoeconomic export controls on Anthropic models). In addition, the EU's "Artificial Intelligence Act" has placed strict compliance requirements on top-tier models, burdening leading platforms with heavy policy baggage. To mitigate geopolitical and compliance risks, Chief Financial Officers (CFOs) now have more rational reasons to proactively divert workloads to middleweight and lightweight models that do not carry regulatory burdens.

A subtle shift is also apparent in the hardware chip sector. Despite top-tier GPU and High-Bandwidth Memory (HBM) orders being fully booked until 2026 with substantial supply-demand equilibrium not expected until 2028, market procurement leaders have begun transitioning from training chips to inference-optimized hardware, reshuffling the winner's structure.