Anthropic's S-1 Shows Agent Labor Is a Margin Business, and Half of It Runs Through Someone Else's Cloud

Start with distribution. Latent Space reports that 48% of Anthropic's revenue flows through cloud partners. In plain terms, nearly one dollar in two reaches Anthropic after passing through a hyperscaler's marketplace, contract, or console. The customer signs with their cloud vendor. Anthropic gets paid as the supplier of the capability inside.

Through the lens of Aggregation Theory, this is the uncomfortable part of the filing. Platforms win by owning the demand relationship and commoditizing supply. On that half of the business, the cloud partners own demand. They aggregate enterprise buyers who already have committed spend, security reviews, and procurement paths, and they offer Claude as one supplier among several. Anthropic's leverage on that half rests on product preference, not on owning the customer.

There is a reasonable counterargument, and the numbers support it: the channel is clearly working. A company does not post an 85% margin while being squeezed into commodity pricing by its distributors. The more plausible read is that Claude currently has enough pull that enterprises ask for it by name, and the clouds compete to be the easiest place to buy it. That is a strong supplier position. It is also a position that depends on staying preferred.

The other 52% matters just as much. That is the direct business: API customers, subscriptions, and managed agent products where Anthropic owns the relationship outright. The strategic tension the S-1 makes visible is between these two halves. Push direct too hard and you antagonize the distributors carrying half your revenue. Lean on distributors too long and you become an ingredient brand.

• Direct channel: Anthropic owns pricing, usage data, and the user experience.
• Partner channel: the cloud owns the contract, consolidates billing, and can steer buyers toward alternatives.
• The swing factor: whichever channel grows faster over the next two years tells you who is winning the aggregation fight.

The second headline figure is the 85% margin reported in the filing summary. The headline does not specify which margin line that is, and readers should go to the filing itself before building models on it. But even read conservatively, it is a striking number for a business whose core input is expensive accelerator time.

The consensus assumption for most of the last two years was that frontier model companies would look structurally like utilities: enormous capital costs, thin spreads, pricing pressure from every new open-weight release. An 85% margin is not a utility number. It is a software number. It implies that customers are paying for outcomes (a drafted contract, a resolved ticket, a refactored module) at prices well above the raw cost of the tokens that produce them.

That is the clearest evidence yet that agent labor has found its price. Buyers are not purchasing compute. They are purchasing work, and they are benchmarking it against the cost of the human time it replaces or augments, not against the cost of the GPUs. Whenever the reference price is a salary rather than a server, margins can stay wide for a long time.

The Latent Space writeup frames the efficiency story as "told in two loops," and the shape of the argument is familiar: better models generate more usage, more usage funds better models, and operating leverage compounds as revenue grows faster than cost. Whether that loop holds once competitors match capability is the open question. Margins this wide are an invitation. Every competitor and every open-weight project now has a published target to undercut.

For operators asking about openclaw cost per month or comparing hosted versus self-run setups, this cuts both ways. Wide upstream margins mean there is room for prices to fall if competition intensifies. They also mean the supplier has no urgent reason to cut them while demand outruns supply.

The milestone that will get the most airtime is profitability. Per Latent Space, Q2 delivered roughly $1B in operating profit, "the first profitable quarter in Anthropic history," driven by "significant operating leverage." The same sentence carries the qualifier: that profit is adjusted for about $1B per quarter of stock-based compensation.

Read that carefully. Stock-based compensation is a real cost; it is paid in dilution rather than cash. A billion dollars of adjusted operating profit sitting next to roughly a billion dollars of quarterly SBC means that, on an unadjusted basis, the company is somewhere near breakeven rather than comfortably in the black. Adjusted profitability is a respectable milestone. It is not the same milestone, and the market will price the difference.

The balance sheet changes the risk profile more than the income statement does. The filing, as summarized, shows $134B in cash against a lifetime burn of $10B. That ratio is unusual for any company at IPO, let alone one in a capital-intensive category. It means Anthropic is not going public because it needs the money to survive. It has years of runway at any plausible burn rate. The IPO is about liquidity, currency for acquisitions, and the credibility of public-market reporting.

The $10B lifetime burn figure deserves its own note. If accurate, it suggests the path from founding to adjusted profitability cost less cash than many observers assumed, with much of the infrastructure cost apparently absorbed by partners and investors in ways that did not hit Anthropic's own cash line. That is consistent with the 48% partner share: the clouds that distribute Claude also have every incentive to help fund the capacity that runs it.

• $1B adjusted operating profit in Q2
• ~$1B/quarter stock-based compensation excluded from that figure
• $134B cash on hand
• $10B lifetime burn

The forward pitch in the S-1 is "virtual collaborators" for over a billion knowledge workers by 2028, and Latent Space wryly flags the word choice: "don't say agent!" The avoidance is easy to read as squeamishness about a hyped term. It is more useful to read it as pricing strategy.

An "agent" invites comparison to software: per-seat licenses, API calls, automation scripts. A "collaborator" invites comparison to a colleague. If the reference price for your product is a fraction of a knowledge worker's salary, the vocabulary that anchors buyers to headcount rather than to tooling is the vocabulary that protects an 85% margin. The billion-worker target sizes the market the same way: by people whose time can be bought, not by developers who integrate APIs.

OpenAI's positioning is converging on the same buyer. Simon Willison observes that in the GPT-6.1 Terra announcement, OpenAI emphasized visual reasoning and computer use, pitching a model that "excels at the communication and computer-based tasks that fill the workday, from drafting proposals and interpreting technical drawings to filling out forms and editing videos." Different vocabulary, same thesis: the prize is the ordinary workday of the ordinary office worker.

When two leading labs aim their flagship messaging at identical work (proposals, forms, documents), the competitive dynamics move away from raw capability and toward distribution and trust. Both can do the tasks. The question becomes who is already inside the customer's workflow, procurement system, and cloud bill. That loops straight back to the 48%.

There is also a governance implication the collaborator framing glosses over. A colleague who fills out forms and edits files on your machine sits further along the Autonomy Spectrum than a chat assistant. Selling at that point on the spectrum, to a billion workers, means the failure modes scale with the pitch.

If model suppliers are earning software margins and distributors are earning a channel cut, the obvious question is where the squeeze goes. The Harness Hypothesis suggests an answer: value in AI accrues to the harness that connects the model to the world, and that is precisely the layer suppliers have the most incentive to commoditize.

The classic move is Commoditize Your Complement. A model vendor wants the harness (the runtime, the tool connectors, the browser control, the sandbox) to be cheap and abundant, so that every dollar of customer value is captured at the model layer. The ecosystem evidence from this week points the same direction. The open harness layer is maturing fast, and maturing means cheaper.

• OpenClaw's 2026.9.35 release is a gateway-only extended-stable release, which the project calls "our current equivalent to LTS." Its maintenance audit covered 3,595 main commits and applied 78 fixes from 77 numbered PRs. Long-term support branches are what open projects ship when enterprises start depending on them, which places OpenClaw in the hardening-to-enterprise stretch of the Molt Cycle.
• Pydantic AI's v2.62.0 lets agents change thinking effort mid-conversation, including through Anthropic's per-message output settings. For users, that means a harness can dial reasoning (and therefore spend) up or down task by task rather than paying premium rates for every turn.
• Willison's evaluation of webframe looks at a small tool that runs a web page "like a program" using the system's existing browser, so it "doesn't need to bundle its own browser engine." Lightweight scraping and sandboxing primitives are becoming free utilities.
• Browser-use's 0.13.14 release adds a protected publishing workflow requiring a second reviewer, the kind of supply-chain discipline that signals a project expects to be embedded in other people's production systems.

On a Wardley Map, these components are sliding from custom-built toward commodity. That is good news for users, who get stable, cheap, interoperable plumbing. It is less good news for anyone hoping to build a durable margin business purely on top of the harness. The S-1 suggests the margin is upstream, and the per-message cost controls in tools like Pydantic AI suggest harness builders will compete by helping users spend less of it.

The filing implicitly poses a question that applies to every layer: who owns the user? Anthropic owns it for half its revenue. Cloud partners own it for the other half. Harness projects like OpenClaw own the daily experience for power users, while having no claim on the billing relationship at all.

The consumer side offers a cautionary note about what direct ownership actually entails. In a quote Willison highlights, 404 Media's Jason Koebler writes that because big tech platforms "make it impossible for their users to contact a human," many people "have turned to ChatGPT," which is increasingly telling them that their routine problems are "journalistically notable" and that 404 Media can help. That is what happens when an AI system becomes the default front door: it inherits every unmet need of the platform behind it, including ones it cannot handle well.

That is the hidden cost of owning the relationship. It brings margin and data, and it also brings responsibility for outcomes. The 48% partner share is, in part, a way of letting someone else carry the support burden, the procurement friction, and the contractual liability for half the customer base. For a company selling "virtual collaborators" into a billion desks, that is not a small convenience.

For power users and operators, the practical takeaway is to map your own dependencies against this structure. If you buy Claude through a cloud marketplace, your pricing and terms are set by two parties with different incentives. If you buy direct, you have one counterparty with more pricing power. If you run an open harness, you have the most portability, and your leverage is the ability to route work to whichever model is cheapest for each task.

The S-1 did not settle who wins the aggregation fight. It settled that the prize is real: agent labor, sold at software margins, to a market sized in knowledge workers rather than developers. Every layer of the stack will now reprice itself around that fact.