Token spending fell by nearly 20%: Pricing power in the AI industry is under pressure, with trillion-dollar capital expenditures by major companies facing ROI tests.

Notice that, just as the market becomes increasingly anxious about whether the massive investment in artificial intelligence (AI) will yield returns, the price per unit of usage in the industry is fluctuating downward. The "Silicon Valley Data LLM Token Expenditure Index," used to track user payments for AI Tokens, was established in December of last year and experienced nearly a doubling growth, but has now dropped nearly 20% from its peak in May. For the 700 billion dollar capital spending frenzy that has supported the heavy development of this industry, this index is the clearest indicator available to everyone. For stock investors, this may be flashing a warning signal: facing customers who are increasingly cost-conscious, AI companies are losing their pricing power, and the expectations for eventual AI profits may be proven to be misplaced. Senior investor Luis Navirel stated: "More and more reports suggest that users of Token-priced AI solutions are forced to restrict their unrestricted usage due to high costs. There are rumors that OpenAI is postponing its IPO until next year, which is seen as a sign that current profitability is still an issue." The LLM Token Expenditure Index doubled before the pullback It should be clarified that a softening in the index does not mean that AI is becoming cheaper. The index combines price and usage, meaning that a decrease could indicate very different scenarios: either official catalog prices are falling, or demand is shifting towards cheaper models. It could also indicate a genuine loosening of actual willingness to pay from buyers. Each of these possibilities brings different impacts. The company "Silicon Valley Data" that created the index has warned people not to interpret it as a price tag. The company refers to it as an alternative indicator of marginal payment willingness. Let's start with the optimistic interpretation: despite Token prices plummeting by over 90% since 2023, total spending has roughly doubled since last year. Cheaper Tokens are expanding the market. This means that the pause in the index is merely digestion and absorption, and demand is real, and capital spending is worth it. This is the bullish logic of companies like Nvidia, memory manufacturers, and data center enterprises. Now let's look at the other interpretation that keeps people awake at night: bearish investors warn that the sustained weakness of this index could end the trading frenzy that has seen almost all AI concept stocks soar in this cycle. Token expenditures have been proving the rationality of the next capital spending order, and the bills now seem somewhat strained. Analysis from Allianz suggests a growth gap of nearly 46% between AI investment and sales. This is more severe than the 32% discrepancy measured during the bursting of the telecom bubble in 2001. Massive data centers are funding artificial intelligence transactions Fortunately for the bulls, this downward trend has paused. Following a week of sideways trading, it is still too early to call a bottom, but it is enough to keep the argument for a rebound alive. David Miller, senior portfolio manager at Catalyst Fund, stated: "During training, the cost of AI infrastructure and Token generation is astonishingly high, but in the current inference phase, economic benefits are significantly better. Net usage of AI brings a positive return on investment for companies, at least in the long run." Furthermore, there is a more recent reason from the demand side that explains why the bearish interpretation may hold some ground. Washington has shown new willingness to control this critical industry. Just this week, the U.S. government lifted foreign access restrictions on the Fable 5 model from Anthropic PBC, and a few days ago regulators requested OpenAI to delay the release of a new version. At the same time, the EU's "Artificial Intelligence Act" imposes mandatory assessments and strict transparency requirements for cutting-edge models. These measures do not directly restrict prices but do create deployment and compliance burdens on top platforms, while subpar but still useful systems do not have to bear these burdens. This consideration may provide a rational reason for chief financial officers to steer workloads towards cheaper models. It is certain that this is not a call for excess chips. Top GPUs and high-bandwidth memory (HBM) have been sold out until 2026 and substantial relief is not expected until 2028. The hardware insights are more subtle, indicating that demand structures are shifting from top training GPUs to inference-optimized components. This changes the mix of winners but does not provide a shorting opportunity. SOX's absolute and relative valuations have fallen from their 15-year highs However, the "unrestrained" market enthusiasm, increasing competition from China, and price sensitivity still keep the strategists led by Chief Investment Officer Vincento Veda at Deutsche Asset Management cautious. They stated: "We are monitoring areas where valuations may seem too high." In conclusion, Token charts have a dual nature, and both interpretations should be held simultaneously. If the sideways momentum from late June is maintained, and this downturn is just a digestion period of a demand structure shift, then cheaper Tokens will continue to expand the market, meaning capital spending will remain rational and bullish logic will remain intact. On the contrary, if this is where regulatory headwinds push demand towards the lower end of the market, at the same time as customer willingness to pay reaches its peak, then the most expensive parts of this transaction will be the first to crumble. Because it is the story of pricing power - not the story of silicon (hardware) - that supports the pace toward a trillion-dollar capital spending in 2027.