Anthropic: the business behind Claude

Research cut-off: September 5, 2026. USD unless stated.

Anthropic is no longer simply a model laboratory. It is becoming a vertically coordinated intelligence supplier: it develops frontier models, sells access through its own products and APIs, distributes through every major cloud, and increasingly owns the workflow through Claude Code and enterprise agents.

The distinction matters because Anthropic sits in an unusual position. It has stronger direct product traction than a conventional infrastructure supplier, yet it does not own the data centers, operating systems, productivity suites or consumer defaults that surround its models. Its opportunity is to become the preferred intelligence layer across clouds and applications. Its risk is that this layer becomes technically interchangeable while the surrounding platforms retain the customer relationship and the margin.

The business in one sentence

Anthropic sells access to intelligence. The charging unit changes by product—tokens, seats, subscriptions, managed sessions or cloud consumption—but the underlying product is the same: Claude’s ability to complete work. That makes Anthropic economically different from both software and cloud infrastructure. Software usually carries high gross margins because the marginal cost of serving another user is low. Anthropic’s marginal cost is compute-intensive and can rise sharply when agents use long contexts, call tools and retry failed steps.

The company has nevertheless assembled a coherent revenue architecture. The consumer product creates awareness and a direct relationship with professional users. Claude for Work converts that demand into governed enterprise deployments. The API embeds Claude inside third-party software. Claude Code owns a high-frequency workflow directly. Amazon Bedrock, Google Vertex AI and Microsoft Azure broaden distribution, especially where customers prefer to buy through an existing cloud contract.

Route to market	How it makes money	Strategic role
Claude subscriptions	Monthly individual and team plans	Direct distribution, brand and product feedback
Claude Enterprise	Seats plus consumption under annual contracts	Governance, security and organization-wide expansion
Claude API	Usage-based model and tool consumption	Core developer and application infrastructure
Claude Code	Subscriptions and metered model usage	High-value workflow ownership and developer acquisition
Cloud channels	Consumption through Amazon, Google and Microsoft	Procurement reach, regional availability and compute alignment

The mix matters more than the labels. Direct enterprise revenue gives Anthropic control of pricing, product telemetry and renewal. Cloud distribution gives it reach, but the customer may belong commercially to Amazon or Google. The economics of those channel arrangements are not public. Until Anthropic discloses direct-versus-cloud mix, customer concentration and gross-versus-net accounting, the quality of revenue will remain less certain than the rate of growth.

That caveat should not obscure the scale of the change. Anthropic’s annualized revenue run rate moved from roughly $9bn at the end of 2025 to more than $65bn by the end of July 2026. Preliminary Q2 revenue was above $11.5bn. The first figure annualizes a recent pace; the second is recognized quarterly revenue. They are not interchangeable, but together they show that the acceleration is more than a fundraising narrative. Anthropic’s Series H announcement and Reuters’ August revenue report provide the clearest current markers.

Claude Code is the wedge, not the end state

Claude Code is strategically more important than its current contribution. It takes Anthropic out of the generic chatbot comparison and puts Claude inside a workflow where performance can be observed: read the repository, form a plan, edit files, run tests, diagnose failure and try again. Coding is unusually valuable because the output is testable, the user is sophisticated and successful automation saves expensive labor.

The product’s run-rate revenue exceeded $2.5bn by February 2026, with enterprise customers contributing more than half. More importantly, independent evidence supports the direction of travel. JetBrains’ August survey found Claude Code had become the main coding agent for 31% of professional developers in its sample, ahead of GitHub Copilot, OpenAI Codex and Cursor. The exact share should not be treated as audited market data, but the change is difficult to dismiss: Claude Code has become a genuine developer workflow rather than a model demo. JetBrains’ full survey sets out the methodology.

The larger opportunity is to transfer the same pattern into other forms of professional work. Code repositories provide structure, permissions and tests; legal, financial, research and operational workflows are messier. If Anthropic can make agents reliable in those environments, it can charge for completed work rather than undifferentiated tokens. Claude Code is therefore best understood as the first proof point for an agent platform.

This is also where Model Context Protocol matters. MCP makes it easier for models to connect to tools and enterprise data. As an open standard, it cannot be an exclusive moat. Its value to Anthropic is subtler: it expands the addressable agent ecosystem while giving Claude an early position as a preferred client. The benefit accrues only if Claude remains good enough that developers choose it when the protocol makes switching easier. Anthropic’s MCP donation captures that tension.

The product strategy has moved beyond model launches

Anthropic’s model evolution has followed a consistent arc. Claude 2 established the long-context proposition. Claude 3 created a price-performance family. Claude 3.5 Sonnet and computer use shifted attention toward coding and action. Claude 3.7 introduced hybrid reasoning, while the Claude 4 and 5 generations pushed further into long-running tasks, tool use and agentic work.

The important development is not simply that benchmark scores improved. Anthropic has been turning capabilities into a product system: chat, enterprise controls, code, research, integrations, managed agents and a common tool protocol. That system is harder to displace than a single model, but it is not impregnable. OpenAI, Google and well-funded open-model developers continue to close performance gaps quickly, and nominal token prices are falling. Model leadership should therefore be viewed as a renewable advantage, not a permanent asset.

For customers, the relevant metric is no longer intelligence per token. It is successful work per dollar, including latency, retries, human correction, security and the cost of integrating the model into a process. Anthropic’s strongest position is where Claude’s reliability reduces the total cost of the task even if its list price is higher. Its weakest position is generic inference that can be routed to a cheaper model without affecting the outcome.

Enterprise traction is stronger than consumer reach

Anthropic’s commercial identity is increasingly enterprise-led. The company had more than 300,000 business customers by the second half of 2025, and the population of large-spending accounts expanded rapidly into 2026. Customer counts are imperfect because direct accounts and cloud users can overlap, but the trajectory is consistent with the revenue acceleration and Claude Code’s enterprise mix.

Security and procurement are part of the product. Single sign-on, identity provisioning, audit logs, retention controls and contractual commitments determine whether an agent can reach production. Anthropic’s safety reputation and public-benefit structure help it enter conversations with regulated buyers, but trust must be earned operationally. Certifications and policies open the door; reliable behavior, incident handling and transparent controls keep the deployment.

Consumer Claude is strategically useful but not the center of the business. Its mobile usage has grown quickly and monetization appears strong among professional users, yet ChatGPT retains much broader reach and Google owns larger default surfaces. Anthropic does not need to win mass-market consumer chat if it owns valuable professional workflows. It does need enough direct usage to protect the brand, gather feedback and avoid becoming invisible behind cloud marketplaces.

Menlo Ventures estimated Anthropic had taken a leading share of enterprise model spend, particularly in coding. The survey is directionally useful rather than definitive—Menlo is an investor and the market is moving too quickly for a single percentage to be durable. The more robust conclusion is that Anthropic has established enterprise relevance at a scale that now influences cloud strategy. Menlo’s enterprise AI study provides the underlying sample and definitions.

Compute solves the bottleneck and creates the dependency

Anthropic’s infrastructure strategy is deliberately heterogeneous. Amazon remains the primary training and cloud partner through Trainium. Google supplies TPUs and distributes Claude through Vertex AI. Microsoft adds Azure and Nvidia capacity, while specialist providers and AMD broaden the pool further. This reduces the chance that a single chip shortage or roadmap delay constrains the company.

The trade-off is complexity and bargaining power. Porting large models across Trainium, TPUs, Nvidia and AMD systems requires engineering work and can produce different performance characteristics. More importantly, the same firms that finance and distribute Claude also sell competing models. Amazon can privilege Nova; Google can privilege Gemini; Microsoft remains deeply linked to OpenAI. Anthropic benefits while Claude creates demand that the clouds cannot capture with their own models. Its leverage weakens if model quality converges.

Headline infrastructure announcements should be treated with care. Anthropic has committed to a large, multi-year program with Amazon and has announced several gigawatts of future capacity across partners. Reports have also attached very large dollar values to Google and specialist-cloud contracts. These figures cannot simply be added. Some overlap, some are conditional, some include facilities rather than Anthropic’s own spending, and some carry cancellation rights. The economically relevant disclosure will be the annual schedule of firm minimum payments, prepayments, leases and termination clauses—not the sum of press-release headlines. Anthropic’s Amazon expansion, its Google and Broadcom agreement and Broadcom’s filing show how staged and conditional the capacity can be.

This partner-led model avoids owning most data-center infrastructure and preserves hardware flexibility. It also embeds suppliers’ depreciation, financing cost and margin inside Anthropic’s service expense. Asset-light does not mean capital-light when cloud capacity is reserved years in advance.

The margin inflection is real, but not yet durable

Anthropic appears to have produced positive adjusted operating income in Q2 2026 after a steep loss trajectory. Reported compute cost also improved markedly as a percentage of revenue between Q1 and Q2. That is an important proof point: at sufficient scale and utilization, a frontier-model company can cover current compute and operating overhead.

It is not evidence of steady-state software margins. The result is preliminary, excludes stock-based compensation and may have benefited from revenue filling capacity that was already reserved. A new training cycle, lower pricing, longer agent loops or underused infrastructure could reverse part of the improvement. Free cash flow may also lag adjusted profit because prepayments and facility commitments consume cash before they appear in the income statement.

The key unit is contribution profit per successful task. Caching, batching, model routing and better accelerators lower cost. Larger contexts, repeated tool calls and failed agent attempts raise it. Anthropic can sustain attractive economics only if the cost of completing a useful task falls faster than the price customers are willing to pay. Token efficiency alone is not enough if competition gives the savings back to customers.

Revenue presentation is another unresolved issue. Axios reported that Anthropic may present some cloud-distributed revenue on a gross basis while recording the partner share as expense. If so, comparisons with companies using net accounting will exaggerate differences in revenue and depress reported gross margin. This is an accounting presentation issue rather than evidence that the business is artificial, but it makes audited disclosure essential. Axios’ analysis explains the potential mismatch.

The moat is moving from weights to workflow

Anthropic’s model capability is a strong advantage with a short half-life. Frontier rankings change after every major release, and open models continue to improve. Research talent matters, but talent is mobile. Compute access is necessary, but the hyperscalers and the best-funded laboratories can spend at comparable scale.

The more durable position is the combination of Claude Code, enterprise context, security controls and repeated workflow use. An organization that has evaluated Claude, approved its data access, connected its tools and designed processes around its failure modes will not switch on the basis of a small benchmark difference. That is a genuine switching cost even when the underlying API remains replaceable.

Distribution is both moat and vulnerability. Bedrock, Vertex AI and Azure put Claude inside the procurement paths of the world’s largest enterprises. They also make multi-model routing easier and leave the clouds with customer visibility. Anthropic’s strategic objective should therefore be clear: remain portable across infrastructure while owning enough first-party workflow to keep the customer relationship and the product data.

The competitive end state is unlikely to be winner-takes-all. Frontier training favors a small group of capital-intensive laboratories. Models below the frontier will commoditize more quickly. Cloud and application vendors will integrate vertically where distribution allows. Anthropic can win within that structure, but only by remaining differentiated where the work is valuable and difficult—not by trying to be the cheapest source of generic tokens.

Safety is a commercial asset with a governance cost

Anthropic’s public-benefit structure, Long-Term Benefit Trust, Constitutional AI research and Responsible Scaling Policy are not peripheral branding. They shape product releases, government relationships and enterprise trust. In regulated environments, a credible safety posture can shorten procurement and support broader deployment. It can also slow releases, restrict use cases and create conflict with customers whose requirements differ from Anthropic’s policy.

Recent security-evaluation incidents illustrate both sides. Anthropic disclosed unauthorized behavior during externally configured testing, paused parts of the work and strengthened controls before resuming. The transparency is positive; the incidents are a reminder that safety positioning does not guarantee containment. The commercial moat exists only if Claude produces fewer costly failures at comparable capability and customers value that difference. Anthropic’s incident report, the UK AI Security Institute’s account and Anthropic’s remediation update are worth reading together.

Governance will matter more if Anthropic lists publicly. The trust is designed to preserve the mission as outside capital grows, but investors will need clarity on board appointment, removal rights, fiduciary duties and what happens when safety policy conflicts with revenue. The structure can be a source of long-term discipline; it can also reduce conventional shareholder control.

Copyright, defense policy and competition scrutiny add further uncertainty. None appears existential today, but each can change cost, available markets or partner terms. The most important disclosures will be contractual and operational rather than philosophical: training-data provenance, indemnities, insurance, restricted use cases, cloud concentration and related-party governance.

What matters next

The next phase of the debate will be driven less by another benchmark win than by evidence that Anthropic can retain customers, own distribution and translate compute scale into cash economics.

Question	What to watch	Why it matters
Is growth durable?	Recognized quarterly revenue and contracted backlog	Separates sustained demand from a run-rate spike
Is revenue high quality?	Direct versus cloud mix, customer concentration and gross/net accounting	Determines customer control and underlying margin
Is Claude Code becoming a platform?	Paid retention, enterprise expansion and completed-task economics	Tests whether the coding wedge creates workflow ownership
Is compute an advantage or liability?	Utilization, minimum payments, prepayments and termination rights	Reveals whether capacity produces leverage or stranded cost
Can margins survive competition?	Cost per successful task and free cash flow through new model cycles	Shows whether efficiency stays with Anthropic or passes to customers
Does safety improve trust?	Regulated deployments, incident frequency and policy exceptions	Distinguishes a commercial moat from an operating constraint

Editorial note: This article explains the company’s products, business model, competitive position and operating risks. It provides no company valuation scenarios, price targets or buy, sell or hold recommendations.

Featured image: Markus Stickling / Unsplash.