ARK Invest CEO Cathie Wood suggests that investors tracking the artificial intelligence market should focus on where autonomous AI agents spend money rather than solely observing consumer chatbots. According to ARK reports, the operational capacity of AI agents expanded significantly, moving from tasks lasting approximately 12 minutes in early 2025 to more than 180 minutes by early 2026. ARK attributes this progress to ongoing industry innovation, including contributions from companies like Anthropic, enabling agents to run independently on complex jobs for hours.
Projections and Revenue Models
ARK projects that AI-driven software spending could reach between $3 trillion and $7 trillion, representing a potential growth range of 19% to 56% in AI-driven software expenditures. Both enterprises and consumers are expected to drive this growth through the adoption of persistent agent technology.
On September 30, 2026, Wood pointed to a potential shift in how AI companies generate revenue, highlighting OpenAI's move toward headless agents-as-a-service designed for higher per-user inference costs. ARK suggests monetization may transition away from traditional subscriptions toward a pay-for-work-completed model. Leading labs are already seeing substantial revenue figures, with OpenAI's projections indicating a run-rate in the tens of billions, backed by expanding agent capabilities and similar growth at Anthropic.
Investment Actions and Market Implications
Reflecting its focus on disruptive technology, ARK purchased approximately $24.7 million worth of CoreWeave shares during the week of September 21, 2026. While ARK maintains positions in digital assets like Bitcoin, its AI agent thesis centers on traditional technology and infrastructure markets.
As agents transition from short interactions to hours-long tasks, the distribution of value within the market may change. Infrastructure providers like CoreWeave could benefit from increased compute demand, and model developers may see higher revenue per user under per-task pricing. Conversely, traditional software vendors relying on seat-based subscriptions face uncertainty, as the traditional model assumes a human operator for each purchased seat.
