Cathie Wood Says Investors Should Track the Money AI Agents Spend, Not the Chatbots
Key Takeaways
- •ARK research shows AI agents advanced from handling tasks of roughly 12 minutes in early 2025 to more than 180 minutes by early 2026, crediting innovators such as Anthropic.
- •The firm projects AI-driven software spending could reach between $3 trillion and $7 trillion, equivalent to 19% to 56% growth, driven by enterprise and consumer adoption of persistent agents.
- •Wood's September 30, 2026 commentary highlighted OpenAI's headless agents-as-a-service approach, with monetization potentially moving from flat monthly subscriptions to pricing tied to the computing each completed job consumes.
- •During the week of September 21, 2026, ARK purchased approximately $24.7 million of CoreWeave, a cloud provider selling the GPU compute used to build and run AI models.
- •OpenAI's revenue projections point to a run-rate in the tens of billions with Anthropic showing similar growth, while traditional vendors relying on seat-based subscriptions could see that model weakened by headless agents.

ARK Invest CEO Cathie Wood has a new recommendation for investors trying to read the AI market: stop watching the chatbots and start watching the wallets. According to Wood, smart investors should track where autonomous AI agents are spending money.
From Quick Chats to Multi-Hour Shifts
ARK's research reports show that AI agents progressed from handling tasks of roughly 12 minutes in early 2025 to more than 180 minutes by early 2026. The firm credits ongoing innovation, including work from companies such as Anthropic, the maker of the Claude models, with pushing these capabilities forward. The shift marks a move away brief, chat-based exchanges toward complex jobs that run independently for hours at a time. Where a chatbot answers a prompt and waits, an agent is software built to carry out multi-step work on its own.
ARK's spending forecast is where the figures grow substantially. The firm projects that AI-driven software spending could reach between $3 trillion and $7 trillion. ARK's analysts frame that as potential growth of 19% to 56% in AI-driven software expenditures. The firm expects both enterprises and consumers to drive that growth as they adopt persistent agent technology.
The End of the Subscription Era, Maybe
Wood's commentary on September 30, 2026, pointed to a possible change in how AI companies make money. She highlighted OpenAI's move toward what she described as headless agents-as-a-service, an approach aimed at higher per-user inference costs. In this framing, "headless" means the agent works in the background with no chat window and no human in the loop, while inference refers to the computing that runs a model each time it does its job.
ARK suggests monetization could move toward a pay-for-work-completed model, away from traditional subscriptions. Such pricing would tie what a customer pays to the computing each job actually consumes, rather than to a flat monthly fee.
The revenue picture at the leading AI labs is part of why ARK takes this trend seriously. OpenAI's revenue projections point to a run-rate in the tens of billions, supported by expanding agent capabilities, with Anthropic showing similar growth.
Where ARK Is Putting Its Own Money
During the week of September 21, 2026, ARK bought approximately $24.7 million worth of CoreWeave, a cloud provider that sells the GPU compute used to build and run AI models. The firm has long built its brand on high-conviction bets in disruptive technology, and it continues to hold positions in digital assets such as Bitcoin. Its AI agent thesis, however, centers on traditional tech and infrastructure markets rather than tokens or protocols.
What This Means for Investors and the AI Market
If agents really are moving minutes-long tasks to hours-long ones, the companies capturing value could shift. Infrastructure providers like CoreWeave stand to benefit from rising compute demand. Model developers such as OpenAI and Anthropic could gain if per-task pricing lifts revenue per user.
Traditional software vendors face a trickier position. Seat-based subscriptions assume a human sits in each seat. If a headless agent does the work instead, the math behind those subscriptions starts to wobble.
Signals worth watching include whether leading labs begin publishing per-task pricing for agent work, and whether future ARK research updates show agents handling still-longer assignments. The $3 trillion to $7 trillion range is a projection, not a guarantee, and the gap between its low and high ends is itself a sign of uncertainty.