- The Silicon Data LLM Token Expenditure Index, which tracks what users pay per AI token and serves as the market’s best real-time gauge of AI monetization health, has fallen nearly 20% from a May 2026 peak after nearly doubling since its December 2025 launch; the index blends prices and usage, so the decline could reflect falling list prices, demand shifting to cheaper models, or genuine softening in enterprise willingness to pay — each with very different implications for the $700 billion-plus capex boom underwriting AI infrastructure.
- The bearish read is that sustained index weakness could unravel the trade that drove the AI cohort’s massive rally: token spending justifies capex orders, and Allianz Research estimates a 46% growth gap between AI investment and AI sales — worse than the 32% divergence at the peak of the 2001 telecom bust; veteran investor Louis Navellier notes growing reports of enterprises rationing AI token use due to high costs, and chatter that OpenAI is pushing back its IPO to next year signals profitability remains elusive at current price levels.
- The benign read is that cheaper tokens are expanding total addressable market: while token prices have collapsed more than 90% since 2023, total AI spend has roughly doubled since last year, and the index’s downward trend has flattened over the past week — not enough to call a bottom, but enough to keep the bull case (digestion, not deterioration) alive; hardware supply remains firmly tight, with top-end GPUs and high-bandwidth memory sold out through 2026 and no meaningful relief expected until 2028.
- Regulatory headwinds add a new demand-side risk that could rationally push enterprise buyers down-market: the EU AI Act imposes mandatory evaluations and transparency requirements on frontier models, the US government only recently lifted restrictions on Anthropic’s Fable 5 model after a 2.5-week shutdown, and regulators asked OpenAI to stagger its next model release — compliance burdens on top-tier platforms that cheaper, lower-capability models don’t carry give CFOs a structural reason to route workloads away from the most expensive AI services.
What Happened?
The Silicon Data LLM Token Expenditure Index — Wall Street’s most-watched real-time proxy for AI monetization — fell nearly 20% from its May 2026 high, reversing a gain of nearly 100% since the index launched in December 2025. The index tracks what enterprise customers actually pay per token across major AI platforms, making it the cleanest available signal for whether the economics underlying the AI capex boom are holding up. Bloomberg reports the decline has paused over the past week, but the trajectory has unsettled investors who have been relying on sustained token-spend growth to justify continued triple-digit capex spending by hyperscalers. The index is not a simple price gauge — it blends prices and usage volumes — which makes interpretation difficult: the same reading can reflect falling list prices, mix-shift toward cheaper models like DeepSeek, or actual demand softening. Silicon Data itself has warned investors to stop reading the index as a price tag, describing it as a proxy for marginal willingness to pay.
Why It Matters?
The entire AI investment thesis ultimately rests on a pricing-power story, not a silicon story. Nvidia, memory makers, and data-center operators can sell every chip they produce through 2028 — but the justification for the next capex order depends on whether the companies buying that compute are generating returns large enough to keep spending. Token spend is the closest approximation of that return signal. A 20% pullback from peak in the index, combined with growing reports of enterprises rationing AI usage due to cost, raises the question of whether the industry is approaching a willingness-to-pay ceiling just as the capex commitment is scaling toward $1 trillion annually. Allianz Research’s comparison to the 2001 telecom bust — where infrastructure investment outpaced revenue generation by a similar margin — is the most pointed warning in the market right now, even if timing any bust remains difficult.
What’s Next?
The next several weeks of token index data are critical: if the late-June flattening holds and the index begins recovering, the bull case gains credibility and the capex cycle stays intact. If the index resumes declining through July and August, the market will be forced to confront harder questions about AI monetization timelines. Watch also for OpenAI’s IPO timeline — if it slips to 2027 as rumored, it would be the strongest public signal yet that the AI majors don’t believe current revenue trajectories support public-market valuations. The demand-mix shift toward inference-optimized chips over training GPUs is already underway; that changes the winner set within semis but doesn’t hand you a clean short. The real risk is a compression of the premium investors are paying for pricing power that hasn’t yet materialized at scale.
Source: Bloomberg
