AI Computing Power Pricing Drops Over 50%, Tech Stocks and Chinese ADRs Under Simultaneous Pressure
nashnova research
AI token prices have plunged more than 50% from their peaks and enterprise spending has flatlined for two straight months — the 'unlimited compute demand' narrative is cracking, with tech stocks and Chinese equities falling in tandem.
How far have token prices actually fallen?
Silicon Data's LLM Token Expenditure Index has dropped below $1 per million tokens, down 54% from its late-May peak of $2.07.
Ramp's index — built on real enterprise spending data from OpenAI and Anthropic customers — is down 48% from its late-February peak of $1.22.
This means → Two independent data sources point to the same conclusion: the unit price of AI compute is in a halving-scale decline.
Why is the price dropping so fast?
Model switching: Enterprises are shifting compute demand from proprietary frontier models to open-source alternatives. Ramp chief economist Ara Kharazian says "billions of dollars in spending migrate from one provider to another each month."
Vendor-driven price cuts: Silicon Data's Steve Hou notes that each new model generation is "more capable but also cheaper."
Caching effects: The rise of AI agents — programs that autonomously execute multi-step tasks — means long interactions generate large volumes of cached input. Cached tokens cost far less than fresh ones. In plain terms = Three forces are compressing prices at once: users switch to cheaper options, vendors cut voluntarily, and the technology itself is getting more efficient.
Usage is rising, but spending isn't — what's wrong?
Under the Jevons Paradox — lower prices drive more usage, which can sustain or grow total spending — token usage volumes are indeed still growing fast.
But Ramp's data reveals a more troubling signal: enterprise token spending has been flat for two consecutive months, not rising with usage. Kharazian calls it "the first time we've seen usage up and spending flat — and it's held for two months."
This means → Volume growth can no longer offset price declines. The revenue engine for AI model companies may be stalling.
What does this mean for AI IPOs and capital spending?
Token spending maps directly to AI model-company revenue. Flat spending means slowing revenue growth.
That puts real pressure on companies like Anthropic preparing for IPOs: valuations are built on high-growth expectations, and once growth slows, the pricing logic wobbles.
Goldman Sachs projects that the largest US hyperscalers will spend a combined $800 billion in capex this year, with another $1.1 trillion expected by 2027. This reflects a widening gap — capital investment is still accelerating, but the revenue signal is already weakening.
What's happening to tech and Chinese stocks?
In pre-market trading on October 22, tech stocks fell broadly. Neocloud names — newer cloud providers specializing in renting AI compute — led losses: CoreWeave and Nebius were both set to drop more than 5%, with Nvidia, Intel, and Qualcomm also lower.
In China, the CSI 300 index fell into bear-market territory, down 20% from its June peak and declining for ten consecutive weeks.
However, sovereign wealth funds are accelerating their return to China, with 2026 allocations up 190% versus 2025. In plain terms = The broad market is falling, but the longest-duration institutional capital is buying the dip.
What to watch next?
Nvidia CEO Jensen Huang told CNBC that neocloud demand for compute is "off the charts," but the statement failed to lift sentiment.
The key variable: whether compute pricing can stabilize even as usage continues to grow.
This means → If prices keep falling and spending stays flat, the "unlimited compute demand" story will need to be repriced — and that question sits at the heart of whether the entire AI investment cycle can sustain itself.
市场有风险，内容仅供研究参考，不构成投资建议。