Original author: Xu Chao
Source: Wall Street Journal
Recent analysis by research firm SemiAnalysis reveals that Anthropic is reshaping the commercialization of AI with profitability and growth rates far surpassing its competitors. With a high-margin, API-centric business model, Anthropic has become the leader in the B2B AI market.
According to a deep report by SemiAnalysis, Anthropic is projected to achieve $1 billion in GAAP EBITDA in the third quarter of 2026, representing a margin of approximately 6%. Meanwhile, its annual recurring revenue (ARR) has surged from $90 billion at the end of 2025 to over $600 billion currently. The firm predicts that if Anthropic maintains its pace of approximately $15 billion in net new ARR (NNARR) per month, its ARR by the end of 2027 could reach $3 trillion, corresponding to a $60 trillion enterprise value, making it the most valuable company in the world.
Anthropic secretly filed its IPO application on June 1. SemiAnalysis believes there is strategic urgency to go public now—Alphabet has completed an $84.75 billion equity financing, and Meta is rumored to be planning a multibillion-dollar fundraising round, as the capital markets window narrows. The report notes that Anthropic’s superior financial metrics and business model mean it should go public before OpenAI to seize the initiative in the capital competition.
Claude Code ignites the B2B market, with ARR growing more than threefold in a single quarter
Anthropic's inflection point stems from the explosive adoption of Claude Code. According to SemiAnalysis data, Claude Code now accounts for over 7% of all code commits on GitHub, directly driving the company's ARR to surge from $3 billion in January to $11 billion in March.
In terms of revenue structure, Anthropic and OpenAI show significant divergence. Anthropic derives approximately 75% to 85% of its ARR from usage-based API services, with consumer subscriptions accounting for only about 5% of total ARR. In contrast, OpenAI still generated over 65% of its revenue from subscription models in Q1 2026, with consumer ARR accounting for approximately 40%.
SemiAnalysis points out that the core advantage of the API model is the absence of an upper revenue limit per user—revenue and token consumption continue to grow as the same customer adopts more agentic workflows, enabling expansion without acquiring new customers. In May of this year, Anthropic’s CFO, Krishna Rao, disclosed on a podcast that the company’s net revenue retention rate (NRR) reached an impressive 500%, meaning that customers who contributed $30 billion in annual recurring revenue (ARR) in Q1 had only generated $2 billion from the same group a year earlier.
The gross margin advantage creates a compounding flywheel, with OpenAI showing a clear gap.
Differences in business models are directly reflected in gross margins. SemiAnalysis estimates that Anthropic’s current consolidated gross margin has risen to the mid-60% range, compared to -94% in 2024. The gross margin for its API business exceeds 80%.
The primary driver behind the significant improvement in gross margin is the enhancement of inference efficiency. Measured by ARR per megawatt of computing power, Anthropic’s metric is projected to reach $60 million by the end of this year, up from $16 million nine months ago. Since the cost of inference computing power is largely fixed, marginal profit margins approach 100% when the number of tokens processed per unit of computing power or the pricing per token increases.
According to the report, if both Anthropic and OpenAI reach $100 billion in ARR, OpenAI’s gross profit will be approximately $25 billion lower than Anthropic’s due to the cost of supporting over 900 million free users (estimated by SemiAnalysis at about $0.70 per user per month). This gap will directly impact both companies’ ability to reinvest in training next-generation models.
SemiAnalysis introduces "pre-tax, pre-interest profit" (EBTIT) as a core metric to measure a lab's capacity for reinvestment, with Anthropic achieving a 36% EBTIT margin in Q2 2026. The report forecasts that Anthropic's cumulative EBTIT before 2028 will be $250 billion higher than OpenAI's.
Beyond programming, cybersecurity could become the next growth engine.
SemiAnalysis estimates that over 65% of current lab ARR comes from programming-related use cases, with programming tool startups such as Cursor, Cognition, Loveable, and Replit collectively contributing approximately $6 billion in ARR. Meta is Anthropic’s largest single customer, but its share remains between 3% and 5%.
The report suggests that cybersecurity will be the next explosive vertical after programming, and anticipates that the release of Fable’s new model will further enhance token pricing and expand use cases, driving monthly NNARR to surpass the current $10 billion per month level in the second half of 2026. Verticals such as healthcare, finance, and biotechnology are also identified as potential major TAM expansion opportunities.
The "Token-as-a-Service" (TaaS) model, sold indirectly through hyperscale cloud platforms such as AWS Bedrock and Azure Foundry, is growing rapidly on distribution channels and now accounts for 15% to 20% of Anthropic’s ARR, up from 5% to 10% just one quarter ago. SemiAnalysis believes that paying 20% to 30% of revenue to hyperscale cloud platforms remains economically rational in terms of enterprise customer reach efficiency and compliance convenience.
Hashrate bottleneck is the biggest variable; IPO provides a funding channel.
The core constraint on Anthropic's growth prospects comes from compute supply.
SemiAnalysis predicts that by 2030, the combined unconstrained compute demand from Anthropic and OpenAI will exceed 100 gigawatts (GW), while net new compute additions in 2025 and 2026 are projected to be only 2.5 GW and 5 GW respectively, with the combined available compute currently exceeding 6 GW.
This supply-demand gap gives the IPO clear strategic significance. The report states that the funds raised will primarily be used to bridge the growing gap in computing power demand between inference operations and new model training, while locking in computing resources at more favorable financing costs. The report also mentions that Meta is considering leasing computing power to external parties (source: market rumors as of July 1, 2026), and anticipates that Anthropic will procure additional computing power from such trusted suppliers.
SemiAnalysis also lists key risk factors, including: rumors of OpenAI’s pricing cuts, competitive pressure from Google DeepMind and Meta in programming models, potential government regulatory restrictions on the release of frontier models, and the dilutive effect on overall gross margin due to the rising share of TaaS revenue. The report explicitly states that if regulatory frameworks hinder model releases and narrow the capability gap between open-source and frontier proprietary models, Anthropic’s business moat would be fundamentally undermined.