ARK Invest chief executive Cathie Wood has adjusted a rule of thumb she has used for years. Investors, she long argued, should watch where developers concentrate their time if they want an early read on technology.
Speaking on a panel at Robinhood’s HOOD Summit in Houston at the end of September 2026, she suggested the more useful trail may now be the spending of autonomous software.
As agents move from answering questions to carrying out tasks, markets will increasingly need to follow where those agents send money.
The change in emphasis follows a change in capability.
Work cited by ARK describes agents handling jobs that once lasted only minutes and that now run for hours without constant human direction.
Once software can research options, select a service, and complete a purchase, the network it pays through becomes part of the investment question.
Public blockchains and stablecoins are being positioned as rails for machine settlement of data, compute, and other digital goods.
Banks, card networks, and large technology platforms are building competing systems so that agent activity does not leave their infrastructure.
The contest is less about which assistant sounds most fluent and more about which pipes capture repeated machine payments.
That activity is already reshaping financial services.
Generative and agentic tools have moved out of isolated pilots and into payments, compliance, fraud review, and client workflows at banks, brokers, and market utilities. Internal assistants support engineering and operations teams.
Monitoring systems mix older statistical models with newer language tools to scan transactions and documents.
Some firms are testing agents that can start limited actions under human oversight, including on brokerage platforms that let users connect outside models to dedicated accounts.
Surveys of financial professionals show heavier daily use than in many other industries, and executives expect wider embedding in risk, credit, and customer processes over the next few years.
Estimates of banking value and of AI-linked software spending run into the hundreds of billions, even while fully autonomous payment volume remains small next to traditional rails.Warnings to slow development have not stopped the build-out.
Researchers and some policymakers continue to flag labor displacement, security weaknesses, energy demand, and concentration of compute.
Companies have still signed multi-year chip and power agreements, treating capacity as the constraint on product plans.
Trading, document review, and payment initiation are changing in practice while the regulatory argument remains open.Into 2027, three firms sit at the center of that capacity race.
OpenAI has assembled multi-gigawatt commitments that mix custom accelerators with supply deals from major chip designers, aimed at keeping model and agent products scaling.
Anthropic has paired fast revenue growth and large private rounds with long compute reservations, including substantial allocations of advanced processors. NVIDIA remains the main supplier of the training and inference hardware both rely on.
Its data-center results and forward guidance reflect sustained orders even as customers test alternative chips and as power limits and export rules add friction.
Overlapping stakes, cloud pledges, and hardware deliveries scheduled into 2027 mean that progress or delay at any one of them will shape how quickly agent-driven commerce and fintech automation expand.
None of this clears the practical obstacles: energy supply, capital intensity, model reliability, and whether agent payments settle on open networks or closed platforms.
Wood’s point is narrower. If agents are already choosing services and paying for them, that spending may show real adoption more clearly than product announcements alone.