Key metrics measuring large model token usage expenditures have recently weakened, prompting the market to reassess the demand foundation for this wave of AI commercialization. According to Bloomberg, citing Silo Data, its LLM Token Expenditure Index has declined nearly 20% since peaking in May, after nearly doubling since its launch in December last year.
This metric is commonly used to observe the marginal willingness of businesses and users to pay for AI services. However, a decline in the index does not equate to a broad decrease in AI service prices. Sil token issuer Data believes that this metric is influenced by both price and usage, making it a closer proxy for changes in willingness to pay, and thus cannot be simply interpreted as a single price signal.
Market divergence is centered on demand-side factors.
Currently, there are two main interpretations of this decline in the market.
A more optimistic view holds that since 2023, token prices have declined significantly, lowering the barrier to entry, and overall spending growth may still continue. In this scenario, the index pullback resembles a restructuring of demand rather than a clear overall weakening of demand.
Another, more cautious perspective suggests that users' marginal willingness to pay may be approaching a阶段性 ceiling. Allianz Research notes a significant gap between the current pace of AI investment and actual sales, implying that if demand continues to slow, the risk of downward pressure on valuations will emerge more rapidly.
Computing demand is shifting from training to inference.
Despite fluctuations in paid demand signals, investment in AI infrastructure has not significantly reversed. The report notes that high-end GPUs and high-bandwidth memory remain in tight supply, with the imbalance potentially lasting until 2026, and some projections extending to 2028.
However, market focus is shifting. Previously, capital expenditures were primarily centered on model training, but they are now gradually moving toward inference. This means that compute demand is no longer concentrated solely on the most advanced training chips; the importance of inference-optimized hardware is rising, and the beneficiaries across the supply chain may also shift accordingly.
From this perspective, the current changes do not necessarily indicate that the chip industry is entering a downturn, but the sources of growth are shifting. The previous model, driven solely by high-end training GPUs, is transitioning toward a more diversified hardware demand structure.
Regulatory changes increase commercialization costs.
In addition to demand and hardware infrastructure, regulatory factors are also influencing the pricing and deployment of AI products. Reports note that U.S. regulators have recently requested adjustments to the release timelines and access arrangements for certain models; the EU’s AI Act imposes additional evaluation and transparency requirements on advanced models.
These changes do not directly lower model prices, but they increase deployment and compliance costs. For enterprises, the importance of cost optimization when allocating workloads across different models will continue to rise, potentially further impacting the adoption rate and pricing power of high-end models.