AI Token Spending Index Drops Nearly 20%, Putting Pressure on Big Tech's Capex ROI
Miles Bennett
The Silicon Valley Data LLM Token Spending Index has fallen nearly 20% from its May peak, just as the question of whether hyperscalers' cumulative $700 billion-plus in capex will ever pay off grows louder — the chart is quietly undermining the pricing-power narrative behind a trillion-dollar investment wave.
What does this index actually measure?
Created last December by Silicon Valley Data, the index blends price and usage volume. It nearly doubled before falling roughly 20% from its May high.
The creators explicitly warn against reading it as a price tag — it is a "proxy for marginal willingness to pay." This means → it tracks not "how much a token costs" but "how much more buyers are willing to spend on AI."
The decline could point to three very different scenarios: list-price cuts, demand shifting to cheaper models, or genuine contraction in willingness to pay. Each carries different implications for the AI supply chain.
Why does the bull case still hold?
Token prices have fallen over 90% since 2023, yet total spending roughly doubled over the past year. In plain terms = the product got cheaper, but far more people bought it, so the total bill grew.
Under this logic the index pause is just a structural-demand digestion period, and the bull case for Nvidia, memory-chip makers, and data-center operators remains intact.
David Miller, senior portfolio manager at Catalyst Funds, said: "At the current inference stage, the economics are considerably better — net AI usage is delivering positive ROI for companies, at least over the longer term."
What deeper risk does the bear side see?
Allianz Research data shows a nearly 46% growth gap between AI investment and AI-linked revenue — wider than the 32% measured during the 2001 telecom-bubble bust. This means → spending is outrunning revenue far faster than it did in the last comparable bubble.
Veteran investor Louis Navellier notes that more users are capping AI usage because costs are too high. Rumors that OpenAI is pushing its IPO to next year are seen as a sign that profitability remains unresolved.
In plain terms = if even the leading AI company hasn't found a sustainable profit model, what underpins the downstream spend?
What is happening on the regulatory and hardware fronts?
The U.S. government this week lifted some offshore-access restrictions on Anthropic's Fable 5 model, while regulators asked OpenAI to stagger an upcoming release. The EU AI Act imposes mandatory assessment and transparency requirements on frontier models.
None of these directly cap prices, but they add deployment and compliance burdens on top-tier platforms. This reflects a practical effect: CFOs now have a rational basis to steer workloads toward cheaper models.
Hardware signals are equally nuanced: top-tier GPUs and HBM — high-bandwidth memory, an ultra-fast memory designed specifically for AI chips — are sold out through 2026, with meaningful relief not expected until 2028. But demand is shifting from top-tier training GPUs to inference-optimized components, changing the mix of beneficiaries rather than pointing to a simple short.
What should investors watch next to judge direction?
After a week of flat trading, whether the index has bottomed is unresolved. Strategists led by DWS chief investment officer Vincenzo Vedda said: "We are monitoring areas where valuations may look stretched."
The real test: if the late-June flat trend holds and the dip proves to be structural-demand digestion, the capex-justification narrative survives.
But if willingness to pay has peaked and regulation keeps pushing demand downmarket — This means → what underpins the trillion-dollar capex pace is a pricing-power story, not a hardware story — and that story is being quietly challenged by this chart.
Token spending pullback — healthy digestion or demand peak?
BULL
Volume-price split is healthy
Prices fell 90%+ yet total spend doubled — cheaper tokens expanded the market.
Inference-stage economics are better
Fund manager cites positive ROI at the inference stage, unlike training-phase cash burn.
Hardware still undersupplied
Top GPUs and HBM sold out through 2026 — demand signals remain strong.
BEAR
Investment-revenue gap exceeds telecom bubble
Allianz data shows a 46% gap, above the 32% in the 2001 telecom bust.
Users forced to cut back
High costs push enterprises to cap token usage; OpenAI IPO reportedly delayed.
Regulation steers demand downmarket
Compliance burdens give CFOs reason to shift to cheaper models, pressuring top-tier pricing power.
In plain terms = both sides have hard data — the real disagreement is not whether hardware sells out, but whether buyers will keep paying premium prices for top-tier models. The late-June index trajectory is the near-term test.
Content is for reference only, not financial advice.