Anthropic Filed for IPO With Estimated $1 Billion Quarterly Profit While OpenAI Still Burns Cash, SemiAnalysis Model Shows
Confidential S-1 filed June 1; independent financial model estimates put Anthropic on path to $6 trillion valuation
Anthropic confidentially filed for an initial public offering on June 1, 2026, setting up what could be the first public listing from a frontier AI lab of its scale. Because the filing is confidential, no official financials have been disclosed. But SemiAnalysis's Tokenomics team, which builds bottom-up financial models of AI labs by SKU, tier, and customer type, has published detailed estimates of Anthropic's business that paint a picture of a company far ahead of OpenAI on the two metrics that matter most: growth and profitability.
The headline number: SemiAnalysis estimates Anthropic generated over $1 billion in GAAP EBIT in the third quarter of 2026, a 6% margin, while OpenAI remains at roughly negative 100% EBIT margins. The firm notes that a recent Wall Street Journal article on Anthropic's financials corroborated the accuracy of its modeling approach, lending credibility to the projections that follow.
ARR Exploded From $9 Billion to Over $60 Billion in Two Quarters
The most striking data point in the report is the speed of Anthropic's revenue acceleration. The company ended 2025 with $9 billion of annualized recurring revenue. By the end of the first quarter of 2026, that figure had reached $30 billion, adding $3 billion in January, $7 billion in February, and $11 billion in March alone. SemiAnalysis now estimates ARR has surpassed $60 billion, driven almost entirely by Claude Code, which the firm says now accounts for more than 7% of total GitHub commits.
Underpinning that growth is a net dollar retention rate of 500%, a figure Anthropic CFO Krishna Rao disclosed on the Invest Like the Best podcast in early May. SemiAnalysis unpacks what that means in dollar terms: of the $30 billion in ARR at the end of the first quarter, $12 billion came from customers who represented just $2 billion of ARR a year earlier. The remaining $18 billion came from new customers still in the early stages of ramping spend, suggesting the growth engine is far from exhausted.
API-First Model Gives Anthropic a Structural Margin Edge Over OpenAI
The report's central thesis is that Anthropic's business model is simply better constructed than OpenAI's. Usage-based API revenue makes up 75-85% of Anthropic's total ARR, with subscriptions accounting for just 15%. OpenAI is the mirror image: more than 65% of its revenue was subscription-based as of the first quarter of 2026, and consumer subscriptions alone made up an estimated 40% of OpenAI's ARR versus just 5% at Anthropic.
This distinction matters enormously for margins. SemiAnalysis estimates OpenAI supports over 900 million free users who cost roughly $0.70 per month each to serve, a drag the firm says knocks 20-30 percentage points off gross margins. "If both OpenAI and Anthropic had $100 billion of ARR, OpenAI would have $25 billion less in gross profit," the report states. That gap compounds over time, since gross profit is what funds the next round of model training.
Blended gross margins for Anthropic have moved from negative 94% in 2024 to the mid-60% range today, a shift the firm attributes to rising ARR per megawatt of compute, which it estimates will reach $60 million by later this year, up from just $16 million nine months ago. "These compute costs are largely fixed per unit of compute, so when you can either get more tokens through a given unit or higher prices on the tokens you run through a given unit of compute, incremental margins approach 100%," the report explains.
Compute Scarcity Is the Real Constraint, Not Demand
SemiAnalysis argues Anthropic's profitability is partly involuntary. "We assume Anthropic likely does not want to be profitable, but compute constraints limit what they can reinvest into training and new compute deals," the report states. During the first quarter of 2026, users experienced rate limits, downtime, and throttling as demand outstripped available infrastructure.
The scale of the looming compute gap is significant. SemiAnalysis forecasts more than 100 gigawatts of combined compute demand from OpenAI and Anthropic by the end of 2030, requiring over 90 gigawatts of net additions through that period, compared with just 2.5 gigawatts added in 2025 and 5 gigawatts in 2026. Today, the two labs combined have access to just over 6 gigawatts. That imbalance is central to why the firm believes Anthropic needs to tap public markets now, alongside hyperscalers like Alphabet, which raised $84.75 billion in equity in June after not issuing equity since 2006, and Meta, which is reportedly preparing its own raise.
Reinvestment Advantage Could Widen the Gap With OpenAI
The report introduces "Earnings Before Training, Interest, and Taxes," or EBTIT, as what it calls "the new default metric for lab reinvestment." Anthropic posted a 36% EBTIT margin in the second quarter of 2026. On a dollar basis, SemiAnalysis estimates Anthropic will have $160 billion of capital after cost of goods sold to reinvest in 2027, versus $92 billion at OpenAI, and projects a cumulative EBTIT advantage of $250 billion through 2028.
The firm frames this as compounding: every dollar of reinvestment advantage widens the gap in model capability, which in turn extends the runway before open-source or rival labs catch up, preserving pricing power at the frontier. Anthropic is estimated to be directing over 60% of its current compute toward training and internal use, with a long-range target of roughly 25% of revenue spent on training compute, implying a 48/52 split between training and inference compute by 2030.
Coding Dominates Revenue, but Concentration Risk Is Modest
More than 65% of lab-wide ARR today comes from coding use cases, with wrapper companies like Cursor, Cognition, Loveable, and Replit adding a combined $6 billion of ARR as of the second quarter of 2026. Despite the coding concentration, customer concentration appears manageable: Meta is believed to be Anthropic's largest single customer, yet represents only 3-5% of total revenue, spending in the low nine figures. Wrapper companies collectively account for just 10-15% of total lab ARR, suggesting the growth is broad-based rather than dependent on a handful of accounts.
The report also pushes back on "tokenbudgeting" fears sparked by high-profile cases like Coinbase's workforce reduction, arguing that episode reflected a cyclical downturn in Coinbase's own end market rather than disappointing AI ROI. Anthropic's help pages reportedly show the average Claude Code enterprise user spends just $150 to $250 per month, with 90% of users spending less than $30 per day. Meanwhile, the number of customers spending over $100,000 annually has grown sevenfold in the past year, and customers spending over $1 million annually are up roughly 42 times over two years.
Distribution Shift Toward Hyperscaler Marketplaces
Anthropic is increasingly selling through token-as-a-service channels such as AWS Bedrock, Azure Foundry, and Google's Gemini Agent Enterprise Platform, with indirect channels now representing 15-20% of ARR, up from just 5-10% one quarter earlier. The overall TaaS market was estimated at $28 billion of ARR as of the second quarter, with 85% share held by the three major hyperscalers. SemiAnalysis frames the revenue share paid to these platforms, typically 20-30%, as a worthwhile trade given the difficulty and cost of direct enterprise sales, calling it preferable "versus paying a BDR, enterprise sales rep, sales engineer, customer success team to manage the land, expand, and retention."
Risks: Price Wars, Regulation, and a Crowded Frontier
The report does not shy away from risks to the thesis. OpenAI has reportedly considered cutting token pricing to reclaim share, and Google DeepMind and Meta Superintelligence could turn coding into a four-horse race, which SemiAnalysis acknowledges "will obviously" pressure token pricing and gross margins if it materializes. Regulatory risk is also flagged directly: government-imposed delays to frontier model releases, referenced in the report as "Fable style" delays, could erode the capability gap that underpins Anthropic's pricing power, particularly given competition from well-funded hyperscaler labs and Chinese labs the firm says are "distilling" Anthropic's models.
On the upside, the firm points to Anthropic's next major model release, referred to internally as Fable, as a likely accelerant for cybersecurity and other new verticals beyond coding, projecting monthly net new ARR could rise above the current $10 billion pace in the back half of 2026.
A $6 Trillion Valuation Case, With OpenAI Under Pressure to Respond
SemiAnalysis's base case values Anthropic at 20 times projected 2027 ending ARR of $300 billion, implying roughly $15 billion of monthly net new ARR next year and an enterprise value of $6 trillion, which would make it the world's largest company. The firm's overarching argument is that Anthropic should move to IPO before OpenAI specifically because its financials are stronger, arguing this forces OpenAI, which has reportedly pushed its own listing to 2027, to open its books and raise capital from a position of relative weakness. "Anthropic has the ability to truly make OpenAI dance," the report states, though it also credits OpenAI's improvement, noting enterprise API checks since the release of GPT-5.5 and Codex have been "overwhelmingly positive" and that OpenAI's incremental revenue is increasingly B2B and API-driven, the same formula that has powered Anthropic's outperformance.
Anthropic Deep Dive: The Enterprise AI Hegemon and the Trillion-Dollar Compute Moat
The Business Model
Anthropic operates fundamentally as an infrastructure utility rather than a consumer application company. The core business model relies on selling metered access to its Claude large language models, monetizing through pay-per-token inference. As of mid-2026, approximately 80 percent of the company's annualized recurring revenue is derived from business-to-business channels, specifically the Claude API and direct enterprise contracts, rather than consumer subscriptions. This structural divergence from its primary competitor, OpenAI, which relies heavily on consumer subscription revenue, has allowed Anthropic to align its unit economics with high-margin, usage-based pricing. By embedding its models into the workflows of developers and large enterprises, Anthropic captures value at the infrastructure layer, charging $10 per million input tokens and $50 per million output tokens for its frontier systems.
The product suite is anchored by the Claude 3 and the newly released 5-series models, including the public-facing Claude Fable 5 and the restricted, highly sensitive Claude Mythos 5. A critical revenue driver has been Claude Code, an agentic coding product that scaled from zero to over $2.5 billion in annualized billings by early 2026. This shift toward agentic workflows consumes significantly more compute than traditional chat functions, driving a massive increase in token volume. Consequently, even as blended token pricing has faced downward pressure across the industry, Anthropic's consumption curve has accelerated exponentially, pushing the company's annualized revenue run-rate past $30 billion by April 2026 and approaching $47 billion by May 2026.
Market Share and Competitive Dynamics
The enterprise artificial intelligence market has consolidated into a strict oligopoly, but the internal hierarchy has violently shifted. By the second quarter of 2026, Anthropic overtook OpenAI in United States enterprise artificial intelligence spending, capturing a 34.4 percent market share compared to OpenAI's 32.3 percent. In the broader $37 billion enterprise foundation model market, Anthropic now commands a 40 percent share, leaving OpenAI at 27 percent and Google at 21 percent. This dominance is driven by a massive base of over 300,000 business customers, with more than 1,000 enterprises spending over $1 million annually. Key end-customers include global systems integrators like Deloitte and Accenture, which have built dedicated Claude Centers of Excellence to train tens of thousands of professionals on the platform.
The competitive landscape is defined by Anthropic winning head-to-head enterprise deployments based on superior performance in complex reasoning and long-context coding tasks. Even Microsoft, the primary backer of OpenAI, integrated Claude Cowork technology into its Microsoft 365 Copilot suite in early 2026, effectively validating Anthropic's enterprise supremacy. While Google leverages ecosystem bundling with Gemini to maintain its footprint, and Meta provides open-weight alternatives through its Llama family for highly customized, on-premise deployments, Anthropic has secured the high-end professional market. This is reflected in average revenue per user metrics, where Anthropic generates more than triple the revenue per user of its closest enterprise competitors.
The Compute Supply Chain
Anthropic's hypergrowth has transformed it into one of the largest consumers of semiconductor compute in the global economy. The company's primary suppliers are Amazon Web Services and Google Cloud, which provide the essential silicon required to train and serve frontier models. In early 2026, facing severe compute constraints that temporarily degraded model performance and forced usage throttling, Anthropic executed a masterstroke of supply chain engineering by locking in 10 gigawatts of compute capacity. This includes a $100 billion, ten-year commitment to Amazon Web Services, securing up to 5 gigawatts of capacity utilizing Amazon's custom Trainium and Graviton silicon.
Simultaneously, Anthropic signed a multi-gigawatt agreement with Google and Broadcom for next-generation Tensor Processing Units coming online in 2027, alongside reported access to SpaceX's Colossus clusters. By diversifying its hardware dependencies across Amazon, Google, and custom silicon providers, Anthropic mitigates vendor lock-in risk while ensuring it has the sheer physical infrastructure required to support its growth. This dynamic creates a fascinating economic loop: hyperscalers are capturing immediate, high-margin infrastructure revenue from Anthropic, while Anthropic leverages their balance sheets to build an insurmountable compute moat against smaller artificial intelligence laboratories.
Competitive Advantages
The primary competitive moat for Anthropic is its proprietary Constitutional AI framework, which embeds safety and interpretability directly at the model level. In the highly regulated enterprise sector, this safety-first architecture is not merely a public relations talking point; it is a strict procurement prerequisite. Enterprises deploy Claude because it demonstrates superior hallucination resistance and explicit non-training on customer data, reducing the legal and operational risks of production-scale artificial intelligence. This advantage was institutionalized in April 2026 with the launch of Project Glasswing, a coalition including Apple, JPMorgan, and CrowdStrike, which utilizes Anthropic's restricted Mythos model to proactively identify and patch critical software vulnerabilities across global infrastructure.
Furthermore, Anthropic benefits from a compounding enterprise flywheel. Deep integration into developer environments via Claude Code creates high switching costs, as engineering teams rebuild their entire software development lifecycles around Anthropic's agentic tools. The resulting usage data refines the model's capabilities in complex, multi-step reasoning, further distancing Anthropic from open-source alternatives that lack the capital to subsidize continuous reinforcement learning at scale. This structural advantage is reflected in the company's unit economics, where inference gross margins expanded into the mid-60 percent range by mid-2026, allowing the company to achieve operating profitability.
Industry Dynamics
The foundational model industry is currently navigating a precarious transition from capability expansion to regulatory friction. The most acute threat to Anthropic is geopolitical and regulatory intervention. This risk materialized abruptly in June 2026, when the United States Department of Commerce forced a global suspension of Claude Fable 5 and Mythos 5 just days after their launch, citing national security concerns regarding the models' dual-use capabilities in cybersecurity and biological research. Although access was restored by July 2026 following the implementation of upgraded safety classifiers, the incident exposed the fragility of artificial intelligence distribution. A single government directive can instantly sever a company's flagship product from the global market.
A secondary threat is the inherent volatility of the artificial intelligence infrastructure buildout. The capital expenditure required to maintain frontier status is astronomical, with industry-wide compute scaling at a multiple of three every year. If the diffusion of artificial intelligence into the broader economy lags this multi-trillion-dollar infrastructure investment, the entire sector faces a severe duration mismatch between compute liabilities and realized enterprise value. Anthropic is effectively betting that its enterprise customers will continue to absorb agentic capabilities at an exponential rate, justifying the billions committed to future data centers.
The Threat of New Entrants
While the capital requirements to train general-purpose frontier models have created a formidable barrier to entry, a new cohort of well-capitalized startups poses a credible threat to the established oligopoly. In early 2026, top researchers departing from Google DeepMind, Meta, and OpenAI secured massive funding rounds to pursue disruptive, non-transformer architectures. Notable entrants include Ineffable Intelligence, which raised a $1.1 billion seed round, and AMI Labs, which secured $1 billion to develop continuous real-world learning systems.
These new entrants are not attempting to match Anthropic's brute-force compute scale on traditional large language models. Instead, they are focusing on specialized vertical models, novel agentic frameworks, and systems that bypass the limitations of static training data. If these alternative architectures prove significantly more compute-efficient or capable of true continuous learning without catastrophic forgetting, they could rapidly commoditize the current generation of models and erode Anthropic's pricing power in the developer ecosystem.
Management Track Record
Chief Executive Officer Dario Amodei has orchestrated one of the most aggressive and successful corporate ascents in modern economic history. Under his leadership, Anthropic transitioned from a research-oriented safety laboratory into a ruthless commercial execution machine, scaling revenue from zero to tens of billions in under three years. Amodei has demonstrated exceptional strategic pragmatism, balancing his publicized warnings about existential artificial intelligence risks with the aggressive monetization of enterprise coding agents. His January 2026 publication, The Adolescence of Technology, masterfully positioned Anthropic as the only responsible steward of artificial intelligence, a narrative that resonated deeply with risk-averse institutional buyers.
His ability to negotiate unprecedented compute agreements with Amazon and Google, effectively forcing hyperscalers to subsidize their own competitor, highlights a sophisticated understanding of platform leverage. Furthermore, management has maintained strict discipline over unit economics, steering the company toward operating profitability while competitors continue to burn capital on low-margin consumer subscriptions. The successful navigation of the June 2026 export-control crisis further cements the executive team's capability to manage complex governmental relations without permanently derailing commercial momentum.
The Scorecard
Anthropic has successfully transitioned from an artificial intelligence research laboratory into the foundational infrastructure layer of the modern digital economy. By aggressively targeting high-margin enterprise and developer workflows, the company has achieved unprecedented revenue scale, surpassing $30 billion in annualized recurring revenue and overtaking its primary competitors in corporate market share. The strategic lock-in of 10 gigawatts of compute capacity across multiple cloud providers ensures that Anthropic possesses the physical resources necessary to maintain its technological frontier, while its Constitutional AI framework provides a durable moat in regulated industries.
However, the path forward is not without severe structural risks. The brief but chilling suspension of its most advanced models by federal regulators underscores the reality that Anthropic's product roadmap is now subject to national security vetoes. Furthermore, the sheer scale of its infrastructure commitments requires enterprise adoption to compound flawlessly over the next decade. Despite these risks, the company's operational discipline, profitability, and dominant position in the developer ecosystem make it the most compelling asset in the artificial intelligence sector today.