Takeaways
-
Token spend is the toll booth of the AI trade. It gives investors a better read on whether end users are still willing to pay for the infrastructure being built.
-
A weaker token index is not automatically bearish. Lower prices can widen adoption, but persistent weakness would raise questions about monetization and customer willingness to pay.
-
The likely next phase is not necessarily a chip glut. It may be a rotation from premium training demand toward cheaper, more practical inference workloads.
-
With valuations already elevated, the market does not need an AI bust to correct. It only needs the pricing-power story to lose momentum.
AI’s Toll Booth Starts Charging Less
The AI trade has been running on a familiar feedback loop: more data-centre spending, more chip demand, more earnings optimism and another reason to push the infrastructure winners higher.
But one of the cleaner signals inside that loop is beginning to soften.
As we have pointed out on numerous occasions over the past 3 weeks, the LLM Token Expenditure Index, which tracks what users are effectively paying for AI usage, has fallen almost 20% from its May peak after nearly doubling since December. That does not mean AI demand has collapsed. But it does suggest that customers may be getting more selective about what they are prepared to pay for.
Think of token spending as the toll booth on the AI superhighway. Chipmakers sell the shovels, hyperscalers build the road and software companies promise the traffic. But token spend tells you whether companies are still willing to pay to use the road.
The problem is that a weaker index can mean several things. List prices may be falling. Customers may be shifting toward cheaper models. Or buyers may simply be discovering that unlimited AI usage looks a lot less attractive once the monthly bill arrives.
The bullish read is still credible. Token costs have fallen sharply since 2023, yet total spending has continued to rise. Cheaper models can expand the market, pulling AI from an expensive corporate experiment into a broader commercial utility. Under that scenario, the recent dip is merely digestion after a strong run, not a warning that demand is disappearing.
That keeps the capex story alive. The mix may shift from giant training clusters toward practical inference use, but the industry still needs compute, memory, networking, power and data centres.
The less comfortable interpretation is that the market is starting to see the gap between investment and monetization more clearly.
The industry has spent aggressively ahead of proven returns. That is normal in a major technology buildout, but markets eventually stop rewarding capex for its own sake. At some point, they ask whether the traffic can pay for the road.
That is why the token chart matters. It is not a clean price index. It is closer to a measure of marginal willingness to pay. And if that willingness is fading just as spending heads toward the next trillion-dollar capex milestone, the most expensive part of the AI trade becomes vulnerable.
The latest flattening offers bulls some room. One steady week does not make a bottom, but it stops the bearish story from becoming a straight line. If token spending stabilizes, the market can still argue that this is a healthy mix shift: lower prices, wider adoption and a larger eventual market.
Still, the road is getting more crowded. Cheaper models are improving. Competition is intensifying. Regulatory costs around frontier systems are rising. Financial chiefs now have a rational reason to direct workloads toward systems that are cheaper, good enough and easier to deploy.
That does not create a chip glut. Top-end GPUs and high-bandwidth memory remain tight. But it could change the winning mix. Training is the racehorse: expensive, glamorous and capital-intensive. Inference is the workhorse: steadier, cheaper and closer to real commercial use.
That distinction matters because valuations are already rich. The trade does not need a collapse in AI demand to wobble. It only needs the market to realise that revenue and pricing power may not rise as quickly as the capex curve.
The bull case remains intact if cheaper tokens broaden adoption and total usage keeps climbing. But if customers are rationing premium AI workloads and shifting down-market, this becomes less of a silicon story and more of a pricing-power story.
And pricing power is the one signal the AI trade cannot afford to lose while the spending bill is still climbing.
