Orply.

Anthropic’s IPO Case Depends on AI Agents Expanding Enterprise Demand

Alex KantrowitzAlex KantrowitzThursday, November 19, 202611 min read

Box CEO Aaron Levie argues that the bull case for Anthropic rests less on its recent surge in revenue than on whether frontier AI can spread into business work beyond coding. He expects agents to increase demand for models by making more tasks worth automating, while cheaper models could complement frontier systems rather than displace them. Levie says adoption and model reliability remain uncertain, and the infrastructure spending behind the expansion could become a serious liability if revenue arrives too slowly.

The bull case rests on adoption, not another year of 12× growth

? aaron-levie argues that the case for Anthropic is better understood as a bet on the frontier-AI market than as a prediction about the right price for its shares. He declined to estimate a fair valuation. His claim is narrower: frontier models are already useful to businesses, adoption is still early, and the work they can do is expanding.

The revenue figures discussed on air illustrate both the scale of the opportunity and the danger of extrapolating too quickly. Kantrowitz cited reported quarterly revenue of $1.4 billion in the third quarter of 2025 and $17 billion in the third quarter of 2026. Levie called the more-than-twelvefold rise extraordinary, but cautioned against expecting another twelvefold increase over the next year. A separate chart shown during the discussion put Anthropic’s run-rate revenue at approximately $81 billion as of October 2026; that is a run-rate figure, not the third-quarter revenue total. The more durable question, in Levie’s view, is how much business activity will eventually run through frontier models.

One measure of adoption is the number of customers spending at least $100,000 a year with Anthropic. The figures shown on screen put that number at 950 in September 2025 and 11,900 a year later.

11,900
Anthropic customers spending at least $100,000 a year, September 2026; figure shown in the discussion

Levie sees that breadth as evidence that AI is generating enough utility for a growing number of organizations to commit substantial budgets—not merely a handful of early adopters experimenting with the technology. If the customer base grows further, and more companies move into spending millions or tens of millions, he says, the scale of the resulting business becomes easier to see.

That is a market argument, not a promise about margins. Levie acknowledged open-weight models could push down inference prices and leave frontier providers with difficult questions about profitability. But he sees the economy’s adoption of AI as being in its early stages. The key uncertainty is whether useful models can spread widely enough, and fast enough, to support the investment being made in them.

Coding is the first market; agents could make other work code-adjacent

Coding is currently the most visible business use for AI, and Levie credits it with driving a large share of the rapid growth at Anthropic and OpenAI. He disagrees, though, with the inference that success in coding may remain confined to software development. Once models can use computers, call tools and carry out sequences of actions, he argues, many other jobs begin to resemble coding workflows: a goal is specified, a system works through tasks, checks results and tries alternatives.

That shift could expand model use in cybersecurity, IT operations, sales, legal work, manufacturing and product design. In manufacturing, Levie points not only to factory-floor automation but to the design work that comes before production. Product designers work with plans, renderings and CAD files; AI systems could help generate and revise these materials, and support repeated prototyping. He described an online example of a model producing a home stereo system’s circuit design, physical design and manufacturing plans. The example matters to his argument because it suggests that models may produce substantial token demand before a physical product is made.

Cybersecurity offers a different route to the same conclusion. Agents could review code, infrastructure, connected devices and system logs, looking for risks across an organization. In research-heavy work, agents could divide a goal into parallel investigations, test hypotheses and compare results. Levie also imagines systems that review a person’s medical records in light of new research or prepare a sales representative for a customer meeting by gathering information from across a company’s records.

These are projections, not established deployments at economy-wide scale. Levie said that some companies are already using such systems, but that broad adoption will require better models, suitable software and organizational change. He described that diffusion as a multiyear process. His revenue thesis depends on the possibility that, as these systems become practical across more workflows, the token use associated with them could be far greater than today’s use in coding alone. At the frontier, he said, spending could grow tenfold; across the economy, he suggested, total token use could grow by a hundredfold.

The underlying change, in Levie’s account, is that the amount of digital work a person can initiate is no longer tied as tightly to that person’s own typing, clicking or design time. A worker could set an agent a goal and budget—for example, to investigate a research question—and let it run. Instead of output being limited mainly by the number of people and hours available, more work could be limited by how much compute an organization is willing to apply.

More capable agents could make previously uneconomic work worth doing

The economic case for agents is not simply that one model completes a task faster. Levie’s broader claim is that lower-cost, more capable systems can make it worthwhile to perform work that companies previously could not afford to do at all. An organization might decide to spend heavily on a security review if the expected improvement in security justifies the cost. In research, it might direct a swarm of agents at a question that would otherwise require much more human time.

That logic also underpins Levie’s answer to the concern that cheaper models will displace frontier providers. He expects different models to handle different parts of a workflow. A frontier model might plan a task, make judgments, choose tools and supervise the work. A less expensive model could then process large volumes of documents or classify data. The example he gave was an AI-driven audit: a frontier model could direct the work while cheaper models extract and assess information across many records.

Levie illustrated the economics with a hypothetical. A task that might cost $100,000 using a more expensive approach could cost $25,000 when frontier and cheaper models are combined. If the value of completing the task is $35,000, the cheaper workflow could make the work commercially viable for the first time. The figures are his example, not reported costs or observed customer spending. His point is that lower prices can expand the set of tasks organizations choose to automate, rather than simply reducing what they spend on models they already use.

He called this a “rising tide lifting all boats” dynamic. Cheaper models could widen the set of tasks that organizations automate, while frontier models remain valuable for orchestration and judgment. The claim is important to the IPO thesis because it offers a way for frontier providers to benefit even as the cost of routine inference falls. It is also conditional: the models must be good enough to carry out useful work, and organizations must find that work worth paying for.

RL and task-specific training could help close the gap in areas where general models still fall short. Levie’s example was a podcast-editing problem raised by Kantrowitz: a system could not reliably distinguish a breath from a spoken “yeah,” so it might delete the wrong sound. Levie said that task may receive less training attention than higher-value problems such as enterprise cybersecurity. A specialized task could improve when the commercial incentive to train for it becomes large enough; but he warned against assuming that progress on one task automatically predicts progress on every other one.

Reliability remains a constraint, even as model behavior changes

Kantrowitz’s editing example brings the reliability question down to a concrete failure: a model can produce useful work in some settings and still make a basic distinction incorrectly in another. Levie agreed that reliability matters, but used the case to make a point about incentives. A niche task with limited commercial value may not attract much training data or investment. A task with greater economic value—such as finding security risks across an enterprise—may attract dedicated training and products built around it.

He does not argue that every model improvement will generalize neatly. Asked whether such problems ultimately come down to tool use, Levie said capabilities can sometimes transfer unexpectedly: training aimed at one kind of software task may unlock related abilities. He also suggested that a sufficiently capable model might eventually infer how to solve a particular audio problem without being trained specifically to remove breaths. But he presented that as a possibility, not a guarantee.

His account of recent models emphasizes a perceived change in how they handle open-ended work. Levie said that, compared with a year earlier, he now uses frontier models to build product prototypes and synthesize research from companies outside Box’s own industry. In one example, he asks a model to study how roughly 20 companies have approached a problem, then draw on broader patterns to design a new enterprise experience. He says the resulting work can combine ideas a product manager would struggle to gather and compare within the time available.

That experience helps explain his confidence in agents, but it does not resolve the reliability problem. The discussion’s tension is that more autonomous systems make more work possible, while also requiring dependable judgment and execution. Levie’s bet is that models, tools and task-specific training will improve enough for more of those workflows to become worthwhile. The audio-editing failure is a reminder that capability is uneven, and that adoption will depend on the details of each task.

Cheaper inference may help the business, but the investment still has to pay back

The program discussed a measure of Anthropic’s compute efficiency: in the first quarter of 2025, the company was said to generate 40 cents of in-period revenue for each dollar spent on compute and infrastructure; by the third quarter of 2026, the figure was $1.72. Levie called that shift significant. Training frontier models is expensive, he said, but a model that is difficult to replicate can be monetized against strong demand. In his interpretation, inference has become more profitable as it represents a larger share of compute spending.

That does not settle whether the business can sustain attractive economics. Levie acknowledged that frontier providers could be commoditized and that open-weight competition could put pressure on prices. His answer was that a frontier provider may still hold value as the reasoning and orchestration layer, even if cheaper systems do more of the high-volume processing. He expects organizations to spend across both kinds of models as they deploy AI in more workflows.

Kantrowitz put the more immediate financial concern plainly: companies have made enormous commitments to infrastructure, while adoption across industries takes time. The bills may arrive before customers are ready. He cited criticism that Anthropic’s future spending commitments are large, and raised a slowdown in reported revenue growth—from 147% between the first and second quarters to 47% between the second and third—as a reason not to assume the earlier pace will continue.

Levie’s response to the commitments was conditional. If frontier models can support hundreds of billions of dollars in annual revenue, the spending may be justified; if that growth does not materialize, the same commitments could become a serious problem. He also disputed the implication that each new area of capability necessarily requires another enormous pre-training run. Post-training, he said, can make progress on specific domains with millions of dollars in data and compute, and he sees potential applications in cybersecurity, biology, legal work, manufacturing, product design and CAD.

He argued that these are not necessarily small markets just because a particular task sounds specialized. The relevant question, in his view, is the total economic activity in each field and how much of it AI could improve or capture. As an example, he pointed to the substantial existing business around cybersecurity and argued that agentic systems could make the category larger by adding new software capabilities and augmenting security teams. His estimates of how much value frontier labs could eventually capture are possibilities, not reported outcomes.

On timing, Levie accepted that the downside scenario is real: AI adoption could arrive too slowly, funding could tighten, and some companies could fail before the investment pays off. Kantrowitz pointed to reports that OpenAI had sought a larger funding round than it had initially raised, though Levie cautioned that the round was still open and said he did not know what could be concluded from that. Levie’s own evidence for optimism is what he sees at Box and among customers: demand continues to rise, and companies deploying inference still run into compute limits. He sees slow diffusion less as evidence of weak utility than as a large opportunity ahead.

Box Architect tests whether agents can help create the workflows they must make useful

Box’s own products provide Levie with a practical test of the agent thesis. The company has offered agents that can search and process documents, answer questions, and be inserted into workflows. A loan-processing workflow, for example, could use an agent to summarize a loan document for an officer making a decision.

The newer step Box announced is called Architect. Rather than simply placing an agent inside a workflow designed by a person, the product is intended to let an agent design the workflow itself. A customer could ask it to create a contract-management system. The system would draw on Box’s existing workflow engine, no-code application builder, file system and security features, using those underlying tools rather than generating code that the customer must host and maintain. Box says Architect is intended eventually to run and optimize the workflows it creates.

Levie said the idea became more plausible as models improved at computer and tool use and began to show domain knowledge in areas such as legal and banking work. Box’s aim is to combine that knowledge with its own enterprise tools so that an agent can assemble a process around an organization’s content. The announcement is a product direction, not evidence that autonomous workflow design is already reliable across every enterprise process.

For Levie, the connection to adoption is direct: an agent that can help build a workflow could make it easier for an organization to put AI into recurring work, rather than limiting use to individual questions or summaries. But the product also makes the adoption condition concrete. Customers must be able to configure and trust systems that handle consequential processes. Whether agents can do that reliably—and how quickly organizations are willing to put them to work—remains central to the broader business case.

The frontier, in your inbox tomorrow at 08:00.

Sign up free. Pick the industry Briefs you want. Tomorrow morning, they land. No credit card.

Sign up free