Cathie Wood: AI Demand is Highly Price Elastic
ARK Invest founder Cathie Wood shared the judgment of company researcher Karim Mattar, concluding that the demand for artificial intelligence is highly price elastic. Mattar stated that the unlocking point lies in the simultaneous shift of costs and capabilities to the frontier, with each new efficiency point showing that cheaper intelligence will be used more.
She previously aligned this argument with price and usage in August. At that time, the average cost per million tokens dropped from about $2.07 on May 28 to about $1.02, a decline of over 50%, attributed to OpenAI's price cuts and the entry of open-source models like DeepSeek and Kimi at significantly lower prices. Wood expressed that prices are collapsing while usage is exploding, indicating a huge elasticity of demand for higher productivity and intelligence, with a virtuous cycle still in its infancy.
Third-party trading data reflects the same spending response. Citadel Securities cited Silicon Data, stating that the volume-weighted token price has dropped by about 40% since the end of June; Ramp's spending index covering over 70,000 companies showed that the top 1% of companies' per capita AI spending increased by about 49% month-over-month in July, the top 10% by about 25%, and the median by about 9%. Their conclusion is that total spending continues to rise when unit costs fall, which is the shape of high elasticity demand, and they are beginning to write the Jevons Paradox as the dominant narrative for calculating demand.
ARK's own mid-year review raised the usage curve further. The Big Ideas 2026 mid-review stated that inference tokens on the OpenRouter interface increased from about 70 trillion per week to over 56 trillion, nearly a sevenfold increase in seven months; data center chip spending is estimated to be revised up from about $580 billion to over $800 billion. She also mentioned that inference token usage is expected to increase about 25 times within a year, thereby shifting the narrative of U.S. real GDP growth from an earlier estimate of about 5% to a double-digit range. According to official data, the annualized rate of U.S. real GDP in Q2 2026 was about 1.5%, down from about 2.1% in the previous quarter.
The revenue side is used to demonstrate that elasticity is not limited to free chat. ARK stated that Anthropic's annual recurring revenue expanded from about $9 billion to about $47 billion, surpassing Salesforce, while Cursor grew from about $1 billion to about $4 billion; Shopify's AI-driven traffic increased about eightfold year-over-year, and AI-driven orders increased about thirteenfold; after Amazon integrated Rufus into Alexa, average spending per shopper increased by over 40%. A Stanford choice experiment estimated that U.S. chatbots create about $172 billion in consumer surplus annually, with payments far below perceived value.
The market mechanism is to exchange price cuts for throughput. The buy side is companies converting cheaper tokens into more tasks and longer agent durations; the sell side is closed-source labs and open-source weights simultaneously cutting prices. Funds are flowing from enterprise software and cloud budgets to token bills, then back to chips and power plants. Beneficiaries are cloud and chip vendors who can still increase total spending after halving unit prices, as well as open-source models that undercut closed-source prices; pressured parties are labs that must rely on high unit prices to cover training depreciation, and other electricity uses that are squeezed when power supply is nearly inelastic. The event-driven aspect is ARK framing "efficiency point migration" as a public investment slogan rather than a new product release.
The rebuttal on the power side is equally specific: each token consumes electricity and silicon, and when the supply curve is steep, elastic demand will not eliminate rents but will shift rents from the model layer to the energy and packaging layers. Wood describes the cycle as being in its infancy, while the market marks the same statement as a bet on whether future capital expenditures will keep pace with usage.
ABAB AI Insight
Cathie Wood frames elasticity as a consistent principle at ARK over the years: falling prices must be validated by exploding usage; otherwise, thematic funds are merely paying for a story. In 2021, she described the AI cost curve as a deflationary engine; by August 2026, the average price per million tokens was halved, completing an empirical test, and in September, she summarized it into four words using the efficiency point graph from researchers. She simultaneously pushed the same elasticity to the macro level: token usage increasing 25 times could lead GDP into double digits. The official growth rate of 1.5% in Q2 is juxtaposed with this slogan, effectively requiring the market to fill the gap with unaccounted productivity.
The capital path involves ETFs and venture capital funds simultaneously standing on the usage side. ARK Venture has included SpaceX, Anthropic, and OpenAI in a retail-accessible private equity pool exceeding $1 billion and is discussing moving labs into public ETFs after their IPOs. The money is not betting on a single model premium but on the premise that "cheaper will be used more" will continue to elevate nominal spending on chips, electricity, and agent software. Shopify's order multiples, Alexa's average spending per user, and Anthropic's revenue surpassing Salesforce are seen as evidence that elasticity has already penetrated trading from APIs.
The analogy is not with internet advertising prices but with the Jevons cycle in the history of electricity and cloud computing: as lighting becomes cheaper, total electricity consumption rises; as cloud hosting becomes cheaper, total machine hours increase. Citadel linking July corporate spending with token price reductions is reiterating this curve. Epoch and Chad Jones's macro elasticity estimates provide a contrasting calibration—if substitution elasticity is below 1, even unlimited automation in an economy will only raise GDP by a limited percentage. The industry phase is expanding usage and competing for pricing power: open-source undercuts closed-source prices, while closed-source exchanges capabilities for premiums, with chip vendors collecting physical layer rents in between.
The structural change is a transfer of pricing power. The mechanism is that the pricing unit for intelligence shifts from "model subscription" to "million tokens"; once prices are comparable, the demand curve becomes steep; whoever can continue to push out cost points can turn others' margins into their own usage. High elasticity does not guarantee labs will profit; it only ensures that someone must produce more watts and more silicon. Rents shift from the software layer to the layer with the least elastic supply—electricity, packaging, and advanced processes.
ABAB News · Cognitive Laws
- After prices are halved, usage must prove the story.
- When elasticity is high, profits do not disappear; they just move to the hardest supply.
- Cheaper intelligence will first consume more expensive workflows.