One number now commands more attention from some investors than almost any earnings report. It has fallen about 20% since May, and people cannot agree whether that is healthy or the start of something bad.
The number is the Silicon Data LLM Token Expenditure Index. To see why it matters, you need two ideas.
The unit that runs the whole industry
AI systems measure their work in tokens. A token is a small chunk of text, roughly a few characters. Every question you ask and every answer you get is counted in tokens.
Companies like OpenAI, Anthropic and Google mostly charge by the token. So token spending is the closest thing the industry has to a meter running on the wall.
The index measures willingness to pay, not price
Here is the part people get wrong.
The index is not the sticker price of any one model, and it is not the total volume of tokens used. It is an expenditure-weighted average of what the whole market pays per million tokens, whatever model that money goes to.
Silicon Data, which built it, describes it as a proxy for the market's marginal willingness to pay. Bloomberg reported that the firm has told people to stop reading it as a straight price tag.
That distinction explains the wild ride. List prices for tokens have collapsed more than 90% since 2023. Yet the index nearly doubled from its December start, because users kept moving up to pricier, more capable frontier models. Now they appear to be drifting back toward cheaper ones.
Why a falling line unsettles Wall Street
The whole AI build-out rests on one assumption. That customers will keep paying more for better models, which justifies the next round of spending on chips and data centres.
That spending is enormous. The capex boom already tops $700 billion, and the industry is heading toward roughly $1 trillion in 2027.
If willingness to pay is peaking, the revenue that funds the next order of GPUs and memory starts to look thin. That is why some investors call this index the cleanest read on the entire trade.
Two ways to read the same chart
There is a calm interpretation and an anxious one.
The calm read: cheaper tokens have widened the market. Total spend has roughly doubled since last year, so the dip is digestion, not decline. "The net use of AI delivers a positive return on investment for companies, at least over the long term," said David Miller, senior portfolio manager at Catalyst Funds.
The anxious read: this is the moment customers stop trading up, as costs bite. Allianz Research notes a growth gap of nearly 46% between AI investment and sales. In the 2001 telecom bust, that divergence was 32%.
Washington and Brussels enter the picture
Regulation now adds to the pressure.
This week the US government removed foreign-access restrictions on Anthropic's Fable 5 model, days after regulators asked OpenAI to stagger an upcoming release. In Europe, the AI Act imposes mandatory evaluations and strict transparency on frontier systems.
None of these caps prices. But they load extra compliance duties onto the top platforms that cheaper, still-useful models avoid, giving customers one more reason to move down-market.
The hardware clock keeps ticking
Underneath the pricing drama sits a physical constraint.
Top-end GPUs and high-bandwidth memory are sold out through 2026. Little relief is expected before 2028.
So the industry keeps ordering capacity years ahead, on the strength of a demand signal that has started to wobble.
That is the tension in a single chart. The march toward $1 trillion in spending is funded by a belief about pricing power. The gauge that tracks that belief has turned down.
