Bret Taylor, chairman of OpenAI and CEO of the AI agent company Sierra, argues that the companies best placed to profit from AI agents are not the startups in San Francisco but the century-old banks, security firms and lenders that hold decades of customer records. Together with his Sierra co-founder Clay Bavor, Taylor makes that case with a simple rule, "rent the intelligence and own the context," while admitting that the AI industry has done a poor job explaining to ordinary communities why they should welcome the data centers that power it.

Old companies are the early adopters

Bavor starts from a candid observation: Silicon Valley lives inside its own bubble, where billboards advertise agentic frameworks and inference technology to an audience of insiders. The more telling signal, he said, is what is happening far from that 25-square-mile stretch of San Francisco.

The customers he pointed to are anything but startups:

  • Santander, the Spanish bank founded in 1857, uses AI agents to manage complex client financial relationships.
  • ADT, whose initials originally stood for American District Telegraph, deploys conversational agents that help homeowners assess and choose a security setup.
  • Rocket Mortgage uses custom agents to walk borrowers through loan qualification questions and affordability calculations. Those agents have helped originate more than $1 billion in mortgages.

Bavor's point is that these deployments are not primarily cost-cutting exercises on the service desk. They sit at customer touchpoints and generate top-line revenue and customer growth. One unnamed retailer with more than $40 billion in annual revenue took a Sierra agent from start to live production in 27 days, which he offered as evidence that large legacy organizations can move faster than their reputation suggests.

Rent the intelligence, own the context

Taylor framed the moment as a platform shift on the scale of the web browser or the smartphone. Those shifts have historically favored startups. He cited Steve Jobs' line that it is more fun to be a pirate than to join the navy, and noted that the strengths an established company built up over decades often flip into liabilities when the platform underneath them changes.

But he sees a structural reason this shift may play out differently. Foundation models, the large general-purpose AI systems that companies license from labs such as OpenAI, are heading toward commodity status. In his view, raw intelligence will be ubiquitous, so it cannot be a durable differentiator for any one business. A hypothetical "GPT-28" will be available to everyone.

What a future model cannot replicate is context: decades or centuries of proprietary customer data, transaction history and brand trust. That is the asset a bank founded in 1857 holds and a startup does not. Hence Sierra's guiding mantra for its customers: rent the intelligence from the labs, and own the context yourself.

Corporate building with a locked vault of customer records below and a rented glowing brain cloud above

▲ Rent the intelligence, own the context

From token panic to paying for outcomes

Cost is the other half of the story. Earlier this year, Bavor said, chief financial officers were anxious about unpredictable developer spending on commercial model APIs, a phenomenon some called "token maxing." Tokens are the units of text that AI models process and that providers meter and bill.

The response has been a shift in how enterprises buy and use models:

Earlier practice Emerging practice
One flagship frontier model for every query Routing each task to the cheapest model that can handle it
Metered billing by token consumption Paying only when the agent delivers a business result
Developers choosing models ad hoc CFOs demanding ROI transparency

Bavor compared using a top-tier frontier model for a trivial request to driving a Ferrari to buy groceries. Instead, enterprises are adopting model routing, through tools such as OpenRouter or internal systems like the one at Ramp, that send lightweight work to smaller or fine-tuned open-weights models (models whose trained parameters are published so companies can run and adapt them) and reserve frontier models for nuanced problems.

Sierra itself built its business on outcome-based pricing. A client pays only when an agent resolves a support ticket, completes a purchase or originates a loan. Sierra handles the applied layer behind the scenes, continuously tuning inference and routing across models so the customer never has to manage model-level volatility.

Personal agents are burning compute on browser emulation

Personal AI agents, which act on a user's behalf, have surged in the past year. Taylor named Muse, Instinct and his own use of OpenAI's Dot as examples of their usefulness. But he described their current architecture as wasteful: to complete a task, an agent typically spins up a virtual machine and emulates a human clicking through a website. That creates enormous compute and inference overhead for jobs that should be simple data transfers.

Sierra has led an open standard called the Personal Agent Protocol to address this. It lets personal agents talk to corporate software directly over structured APIs rather than scraping or navigating visual web pages. Taylor expects that to cut computational load sharply while making consumer transactions through agents smoother and more secure.

The data center backlash and a messaging failure

Following the midterm elections, public and legislative pushback against data center construction has grown, driven by grid capacity and land use concerns. Bavor acknowledged the debate as a natural and necessary part of democratic governance, but argued that the underlying demand is unconstrained. Every company, in his analogy, would happily upgrade its staff-level engineers to principal engineers or multiply its high-quality customer touchpoints if it could, and that appetite for cognitive compute is what keeps infrastructure expanding.

He laid out the local trade-off plainly: on one side, manufacturing, training and service jobs plus tax revenue; on the other, noise and rising electricity bills. Ideas that sounded far-fetched 24 months ago, such as orbital solar arrays and space-based data centers, are now worth exploring given the pace of demand and advancing launch capability from Elon Musk's rocketry efforts.

Taylor was more self-critical about the industry. He said the AI sector has largely failed to convey a compelling value proposition to communities in the United States, and that residents do not yet see a direct personal dividend such as a lower tax bill. Until benefits become tangible, he expects projects to keep meeting resistance. He also described the historical pattern known as Jevons paradox: whenever energy gets cheaper and cleaner, total demand rises to absorb the new supply rather than leveling off. He expects cheaper compute to behave the same way, unlocking unforeseen uses in science and industry that in turn require even more compute. His prescription is more science and engineering, paired with proactive community deals that deliver mutual local benefit.

There is also a geopolitical angle: if domestic grid delays and political friction persist, compute could migrate to regions such as Malaysia or the Middle East. Taylor noted that AI is a global phenomenon and that every nation will pursue its own power, compute and technological sovereignty according to its national security priorities.

Town hall where residents weigh a data center model against houses, with coins balanced against a power pylon

▲ Local trade-offs in data center approval

Scarcity as an accelerator

Bavor also sees resource caps pushing the United States toward more creative architectures, recalling that Google co-founder Sergey Brin often cited scarcity as the true catalyst for invention. DeepSeek, the Chinese lab that demonstrated high capability under severe compute restrictions, and distilled models, smaller systems trained to replicate the behavior of larger ones, stand in his view as proof that constraint forces efficiency. Intense demand for frontier intelligence, he said, is pushing engineers to extract more from every token and every GPU, from model design to data center layout to silicon.

A memory hierarchy for intelligence

On pricing pressure from low-cost overseas models, Taylor praised OpenAI's Luna model family, which he described as remarkable and, for many enterprise use cases, cheaper to run than self-hosting open-weight alternatives. He offered two analogies for where the market is going:

  • The carrier phone store: at Verizon or T-Mobile, a buyer can pay about $1,000 for a flagship phone or get an older model free with a plan. Both are valid choices for different needs.
  • The storage hierarchy: software mixes spinning disks, flash storage and RAM according to latency and cost targets rather than putting everything in the fastest tier.

Intelligence, he argued, will follow the same logic. Within two to three years, he expects deployments to rely on an array of models mapped to specific cost, speed and capability requirements rather than a single massive model.

A warning for IPO investors

With potential public offerings from Anthropic and OpenAI on the horizon, Bavor's advice for public market investors was humility. Over the past three years, coding agents, DeepSeek and OpenAI's o1 reasoning models each arrived faster than forecasts anticipated. Aggregate demand for intelligence appears limitless, he said, but its technical form, commercial packaging and geographic distribution remain highly volatile. Only a tiny cohort of labs operates at the frontier, and anyone claiming certainty about AI's exact commercial or technical path deserves deep skepticism. He suggested an impressionistic, long-term exposure rather than reliance on rigid predictions. None of this constitutes investment advice; it is the perspective of two executives whose company sells AI agents.

What this means and what to do

The throughline is that the value of AI agents is shifting from the model to the context around it. For established companies, that is encouraging: the customer records, transaction history and trust they already hold are the scarce ingredient. For the industry, the message is harder: compute demand may be boundless, but the public that hosts the hardware has not been shown why it should care.

Practical takeaways:

  1. Treat your proprietary customer data and CRM history as the core asset of any agent deployment, and rent model intelligence rather than trying to build it.
  2. Stop sending routine queries to frontier models. Use routing to match each task to the cheapest adequate model, including distilled or fine-tuned open-weights options.
  3. Budget AI on outcomes such as resolved tickets, completed sales or originated loans rather than token volume.
  4. If you build consumer-facing systems, expose structured APIs so personal agents can transact without emulating a browser; the Personal Agent Protocol is one standard to consider.
  5. If you build infrastructure, engage host communities early with concrete offsets such as job training, utility relief or tax benefits, because abstract national benefits have not persuaded them.
  6. If you invest, assume the next three years will surprise you as much as the last three did.