[More serious than the Internet foam: Token consumption dropped by 20%, and the gap between AI investment and sales growth reached 46%] According to Beating monitoring, the Silicon Data LLM Token consumption index, which tracks users' actual computing power expenditures, has fallen nearly 20% from its May high. The sudden halt of this high growth trend has issued a key warning to investors: large model manufacturers may be losing pricing power in the face of cost sensitive customers, and has also cast doubt on the final investment return of billions of dollars in AI capital expenditures in the market. The divergence between the long and short camps has intensified. The bear market faction pointed out that Allianz research data showed that the growth gap between AI investment and sales reached 46%, and the imbalance degree has exceeded 32% when the telecom foam burst in 2001. Bull market advocates argue that although the average price of tokens has plummeted by 90% since 2023, total expenditures have still nearly doubled, indicating that the index's decline is only a structural digestion after price reductions promote consumption, and the investment return rate in the inference stage is far more optimistic in the long run than in the training stage. The strengthening of policy regulation is translating into implicit compliance costs for enterprise users. Washington has imposed stronger policy scrutiny on the distribution and cross-border access of cutting-edge models (such as the release review of OpenAI and geo export controls on Anthropic models), coupled with strict compliance requirements for top-level models under the EU's Artificial Intelligence Act, putting a heavy policy burden on top platforms. In order to avoid geopolitical and compliance risks, corporate financial managers (CFOs) have more rational reasons to actively redirect workloads to medium to lightweight models without regulatory burdens. The field of hardware chips also shows subtle changes. Although orders for top tier GPUs and high bandwidth memory HBMs have been fully booked until 2026, and substantial supply and demand easing will not occur until 2028, the main market procurement force has begun to shift from training chips to inference optimization hardware, and the winner structure is being reshuffled. [Original link]