From Inference to Execution: AI’s Next Economy

Authors: Jaya Gupta and Ashu Garg (@ashugarg)

For the past few years, the AI story has been written by the companies supplying intelligence and their infrastructure. The central questions revolved around NVIDIA, frontier labs, inference providers and neoclouds: how large could a model be, who could secure enough GPUs to train and serve it, and who could finance the next cluster. That era is ending.

The next phase will be organized around a different question: who will convert intelligence into useful work, and ultimately control the economics of getting it done?

CoreWeave makes the stakes concrete: its second-quarter revenue more than doubled to $2.575 billion, while its June backlog reached approximately $104 billion. Yet net interest expense consumed $640 million, roughly a quarter of revenue. On October 14, 2026, its shares fell about 16%, and yields on its 2031 bonds rose to nearly 15%. CoreWeave’s revenue is growing while investors demanded much greater compensation for financing its expansion. Demand alone doesn’t answer the question investors are now asking: who captures the value of the work that capacity performs?

What those customers buy next matters as much as how much they spend. Agents generate a growing workload around inference, and applications are gaining more control over how that workload is performed. The intelligence they use and the infrastructure that executes the work are becoming separately purchasable components.

My bet is that power shifts toward the companies deciding which steps need frontier reasoning, which can use smaller models or ordinary software, where the work runs, and which suppliers get paid. An application can buy more computing overall while continually revising its mix of inputs. Its suppliers have financed models, machines and capacity against expectations about those purchases.

Agent work is moving from inference to execution

Ask an assistant to watch flights for two weeks and rebook if the price drops. That one request creates two workloads. Inference interprets the request and judges unfamiliar options. Frontier inference typically runs on GPUs and other accelerators built for the parallel math of neural networks. Execution checks fares, stores results and completes the booking. It runs on CPU-based services: programs, databases and network calls.

Most of the job is execution. A scheduled worker checks prices every hour, throws out unchanged results and calls a model only when an option needs judgment. Between checks, the worker shuts down while the schedule and history persist. A booking still has to be authorized, reach the airline, commit and be recorded. If the confirmation is lost, the system has to establish whether the booking happened before it tries again. Those steps can run without another model call, but execution, storage and recovery still cost money.

This shift is easy to miss in an industry focused on training and serving models. A chat session generates work when someone asks a question. An agent can keep generating work after the conversation ends. Execution tracks jobs: scheduled checks, incoming events, unfinished transactions and failed steps that need another attempt. One instruction can create weeks of demand for CPUs, networks, storage and code hosts. The company assembling the service decides where that work runs and which suppliers get paid.

Look at what the new agents actually require. Meta’s Muse runs on Meta’s own model, but the work happens on a dedicated cloud virtual machine that keeps going after the app closes. A separate process gates what it sends to the internet, Stripe issues one-time cards for its purchases, and on the Mac it needs the operating system’s permission to reach local files and mail. One assistant combines model serving, a cloud computer, network controls, payments and local permissions. Each introduces a cost, a dependency or another company’s authority over the work.

The other launches show how those resources stay busy between conversations. OpenAI’s Dots monitors customer feedback, builds and tests fixes, and prepares pull requests. xAI’s Grok Bot reruns workflows on a schedule from a shared cloud computer with persistent browser sessions. Microsoft’s Copilot Cowork runs scheduled tasks and responds to email or Teams triggers. A single user can launch several assignments that keep running without another prompt.

That work lands on ordinary infrastructure first. Daily requests from AI agents on Cloudflare’s network grew more than 1,700% in a year. GitHub’s Git activity more than doubled, to 473 billion events a month, and pushes rose almost fivefold. The shape of the load changed too: agents commit after nearly every action, thousands of them merge onto a single branch, and each push sets off thousands of fetches a minute. GitHub’s architecture couldn’t absorb that, so it is rebuilding its storage layer.

The chip makers have noticed. In late August, NVIDIA began shipping Vera, its first processor built for agents, to Anthropic, OpenAI and Oracle. The executive who delivered it said, “we are seeing the demand for CPUs skyrocket.” Intel describes the CPU-to-GPU mix moving from about one CPU per eight GPUs in training toward one-to-one in agentic deployments, and it has moved production from consumer chips to server parts. By the end of September, reports said AMD’s 2027 server CPU capacity was booked.

So have the buyers. Anthropic agreed to buy about $11.6 billion of capacity from Akamai over initial seven-year terms. Akamai ties the deal to growing CPU workloads and expects about $5.5 billion of related capital spending. A lab that sells the complete service becomes a major buyer of the infrastructure around its models.

Execution is also being sold in pieces. Temporal records completed steps and coordinates retries. Sandbox providers such as E2B bill for the CPU and memory a sandbox uses while it runs. The execution bill is growing, and so is the list of suppliers competing for it!

Who assembles the agent controls the mix

The companies assembling agents decide what each step requires, and their choices are multiplying in three directions: which intelligence to use, where the work runs and how each operation is routed. A single assignment can use several models, run across different environments and take different routes into the same applications. Whoever assembles those pieces decides which suppliers participate and what gets bought from each.

There are far more ways to supply the thinking. OpenRouter alone offers more than 400 models and handles more than 10 trillion tokens a day. Small decision models answer narrow questions without writing a long reply; Cloudflare’s Clef-flash picks an option or answers yes or no for $0.038 per million input tokens. Some inference stays on the device: Apple’s new Siri chooses between the phone and Apple’s cloud, and Windows is adding routing to local models. And some inference only has to happen once. Google’s new Gemini agent can generate a BigQuery report, save the query and rerun it later with no token charges.

The ways to carry out the work are multiplying just as fast. Take a refund. One service confirms which customer the agent is acting for and whether it may send the money. Another submits the transaction, through a direct API or a signed-in browser. A third records the result so the job survives a crash. The worker running all of it can sit on a laptop, in a cloud sandbox or at the edge, and each piece can come from a different company. OpenAI’s Agents API launched with nine named sandbox partners, so a developer can buy the harness from one company, the model from another and the computer from a third.

Then there is the layer that chooses among all of it. Follow one support request. A gateway checks whether it has been answered before; Cloudflare, Portkey and LiteLLM can serve repeats from a cache. If not, a small model sorts it: a password reset goes to a cheap model, a disputed charge to a frontier one. If that provider fails, the gateway switches to another.

Multiply these choices together and** the companies assembling agents become always-on negotiators**. Much enterprise software is still bought on multi-year contracts, but agent workloads are increasingly bought by the request. Assemblers test combinations of models, tools and runtimes against completion rate, speed and cost, and send more work to whatever wins. Swapping a runtime takes integration work; once routes are wired up, switching models takes a configuration change. The assembler can re-decide the mix on every run, so every supplier is competing for the next request.

Each choice cuts into lab revenue differently: routing to a smaller model still buys inference, a prompt cache cheapens the repeated part of a request, and reusing an answer or running saved code removes the request entirely.

Outcome pricing rewards this continuous adjustment. Sierra charges for results such as resolved conversations or retained customers. Delivering that result consumes inference, execution capacity, storage and recovery services. At a fixed price per successful outcome, completion rates, retry amplification and machine utilization determine how much revenue becomes margin.

Every layer wants to control the route

The next platform war is over who gets to turn everyone else into a supplier. Each layer wants to use its existing advantage to organize the whole service, sell it to the customer and buy the remaining capabilities as inputs.

Browsers and connectors are competing for the route into applications. Chrome’s DevTools MCP lets an agent use the user’s signed-in browser, with permission, while a direct connector performs the same action without loading the page.

Devices and clouds are competing to hold the assignment. A device-centered agent keeps local context and permissions and buys inference from the cloud. A cloud agent keeps the task’s history, schedules the next run and treats the device as an endpoint. What matters is which system remembers what’s unfinished when the conversation ends.

Labs and infrastructure platforms are competing for the operating layer itself. Anthropic’s Managed Agents keeps the session history and the work queue even when customers bring their own sandbox workers. Cloudflare comes from the other side, offering execution environments and persistent browser sessions that developers build on with whatever model they choose.

Business applications are defending the transactions underneath. Salesforce exposes actions to Claude but enforces its own permissions and rules. The assistant can coordinate the work, but Salesforce decides which changes its records accept. Owning the conversation doesn’t carry authority over the systems where the work happens.

I expect cybersecurity to become one of the largest enterprise uses of inference (maybe even 30% next year). Continuous code review, exploit validation and remediation testing create recurring model work as software changes. Security incumbents already have customer relationships, telemetry and access to the systems being protected. Palo Alto’s Unit 42 assembles multiple models behind a proprietary harness for continuous testing; its AI gateway enforces enterprise policy and routes requests across permitted models and tools. A substantial new pool of token demand could emerge inside security contracts incumbents already own, with the labs competing to supply it along with a plethora of cheap post-trained open weight models.

These contests decide who writes the purchasing rules and gets paid to coordinate the work. Warp charges platform credits per agent-hour for every cloud run, even when customers bring their own inference and workers, including runs with Claude Code or Codex. The coordinating service can keep earning its fee while the suppliers underneath it change.

Capital is committed to a mix that is already changing

Applications and assemblers can change what they buy long before their suppliers have earned back what they spent.

Nortel and Lucent are the warning. They supplied the equipment for a network that would become indispensable. The internet kept growing. Traffic kept rising. But carriers cut equipment orders and used the capacity they had already installed. The technology succeeded while the purchases expected to pay for its buildout collapsed.

Today NVIDIA supplies the equipment, and neoclouds like CoreWeave finance and operate the capacity. Their customers can expand the work they complete while changing the models, the number of inference calls and the infrastructure each job requires. CoreWeave itself warns that customers may redesign their systems to need fewer of its services and decline to renew or expand their purchases. Existing contracts support committed revenue, however, the exposure sits in what comes next: renewal prices, uncommitted capacity and the next cluster someone has to finance. That is the danger: customers can keep growing while the spending pattern that financed their suppliers comes apart.

The bet

My bet is that the next phase of AI breaks the link between owning the most GPUs and capturing the most value. Thousands of applications and assemblers will buy a changing mix of frontier inference, smaller models, open weight models, reusable software and execution services, and routing will spread those purchases across a growing field of suppliers.

The AI boom can succeed while parts of the investment thesis financing it fail.