Anthropic files for an IPO as profitability comes into view
Anthropic quietly submitted IPO paperwork to the U.S. Securities and Exchange Commission in July 2026, according to a lengthy SemiAnalysis report cited by Z Finance. The report’s central claim is straightforward: Anthropic is no longer behaving like a pure research lab with open-ended losses. It is starting to look like a high-margin, high-retention software business built on large-model infrastructure.
SemiAnalysis said Anthropic’s GAAP operating profit, or EBIT, topped $1 billion in the third quarter of 2026, with an operating margin of 6%. Over the same stretch, annual recurring revenue climbed from $9 billion at the end of 2025 to more than $60 billion.
The comparison with OpenAI is sharp. SemiAnalysis said OpenAI was still operating at about a -100% EBIT margin in the second quarter of 2026. In the report’s framing, this is the first time a leading AI lab has shown clean profit figures strong enough to argue that the large-model business model is working in commercial terms.
Why Anthropic would list now
SemiAnalysis said Anthropic is not being pushed toward an IPO by a cash shortage. The company has raised more than $100 billion since inception, has already reached profitability on a non-GAAP basis, and has stable gross margins. The report instead ties the timing to compute needs, enterprise credibility, and hiring pressure.
Compute is at the center of the argument. SemiAnalysis projects that OpenAI and Anthropic together will need more than 100GW of compute by the end of 2030. That would require roughly 90GW of net new supply over five years. The report notes that only 2.5GW was added in all of 2025, with 5GW expected in 2026, while the two companies currently have only a little over 6GW available in total.
That shortage has already been visible to users. The report points to rate limits, outages, and service degradation in the first quarter of 2026. Anthropic CFO Krishna Rao said on the Invest Like the Best podcast in early May that demand at the time was far above available supply.
SemiAnalysis lays out three reasons a public listing matters. First, public equity and debt markets could fund model training, compute leases, and new data center contracts more quickly and at lower cost. Second, audited public financials could ease concerns for large enterprise customers considering API contracts worth hundreds of millions of dollars a year. Third, liquid equity is a powerful recruiting and retention tool in a market where top AI researchers are paid at levels comparable to professional athletes.
Trying to set the valuation anchor before OpenAI
The report treats the planned IPO as a strategic move in Anthropic’s rivalry with OpenAI. In public markets, the first high-growth company in a new category often becomes the valuation anchor for everyone that follows. Revenue growth, margins, retention, and revenue multiples become the default reference points.
SemiAnalysis argues that Anthropic wants to lock in that position while its financial profile compares favorably with OpenAI’s. It highlights Anthropic’s 6% EBIT margin, gross margin above 60%, and 500% net dollar retention. For OpenAI, the report points to an EBIT margin of about -100%, a large free-user burden, and an NDR figure that has not been disclosed.
On that logic, if Anthropic lists at 20x ARR, or roughly a $6 trillion valuation, that multiple could become the market baseline. Any later listing by OpenAI would then be measured against Anthropic on revenue quality, profitability, and the durability of growth.
Claude Code and the shift to usage-driven revenue
SemiAnalysis identifies Claude Code as the key product behind Anthropic’s recent acceleration. The coding assistant, which the report says took off in early 2026, now accounts for more than 7% of all GitHub commits.
The revenue ramp described in the report is unusually steep. Anthropic ended 2025 at $9 billion in ARR, then added $3 billion in January 2026, $7 billion in February, and another $11 billion in March, reaching $30 billion in a single quarter.
That growth, SemiAnalysis said, is tied much more to API usage than to subscriptions. APIs account for 75% to 85% of Anthropic’s revenue mix, while subscriptions make up only 15%. Consumer revenue is about 5% of Anthropic’s ARR, versus roughly 40% for OpenAI.
The consumer business is still meaningful. SemiAnalysis estimates Anthropic has about 55 million to 60 million free monthly active users, with a paid conversion rate of about 9%, above OpenAI’s roughly 6%.
500% retention and the API flywheel
A major part of the report focuses on how API billing changes the economics of AI. Subscription models grow closer to seat count. API revenue scales with actual token consumption, which can rise sharply as customers expand agent-based workflows.
Krishna Rao disclosed a 500% net dollar retention rate, according to the report. SemiAnalysis interprets that to mean customers already on the platform in the first quarter of 2025 were spending five times as much a year later. It says that cohort went from $2 billion in ARR in the first quarter of 2025 to $12 billion in the first quarter of 2026, equal to 40% of Anthropic’s $30 billion ARR at the time.
The report argues that this is far beyond normal SaaS comparisons. Good SaaS companies often post NDR around 120%, with top performers reaching 150%. Anthropic’s figure is driven by how agentic workflows consume tokens. A standard chat exchange may use a few hundred tokens; an autonomous coding task can burn through millions.
Gross margin climbs from -94% to above 60%
SemiAnalysis said Anthropic’s gross margin was still -94% in 2024, meaning the company was losing almost $1 in cost for every $1 of revenue. It now puts gross margin at more than 60%.
The report points to three drivers. New frontier models launch at higher prices than prior generations. Inference efficiency has improved through better accelerator chips, stronger inference stacks, and smarter caching. And compute costs are relatively fixed, so as each unit of compute generates more revenue, incremental gross margin moves closer to 100%.
One metric in the report captures that change. ARR per megawatt of compute was $16 million nine months ago and is expected to reach $60 million later this year. SemiAnalysis also says Anthropic’s API business carries gross margins above 80%. Even after giving up 20% to 30% of revenue to channels such as AWS Bedrock, and even with newly leased compute priced at twice market rates, the report says blended gross margins still have room to rise.
It contrasts that with OpenAI, whose more than 950 million free users cost about $0.70 per user per month to serve. SemiAnalysis says that creates a 20% to 30% drag on total gross margin. If both companies were to reach $100 billion in ARR, OpenAI’s gross profit would be $25 billion lower than Anthropic’s, according to the report.
EBTIT emerges as a new AI metric
SemiAnalysis introduces a new metric, EBTIT, or earnings before training, interest, and taxes. The goal is to separate the economics of inference from the heavy spending required to train new models.
By that measure, Anthropic posted a 36% EBTIT margin in the second quarter of 2026, the report said. SemiAnalysis argues that training spend is so large that it can obscure the real profitability of inference. Anthropic is still putting more than 60% of revenue into training and new-model R&D, but the report treats that spending as closer to long-cycle capital investment than to ordinary operating expense.
It adds that if training spend falls to 40% to 50% of revenue over the next few years, GAAP operating margins could move toward EBTIT levels. SemiAnalysis expects EBTIT to become common valuation language for AI labs over the next two to three years.
Anthropic’s B2B API path versus OpenAI’s consumer model
The report frames the Anthropic-OpenAI split as a difference in business design, not just product strategy. Anthropic is built around B2B APIs and usage-based pricing. OpenAI, by contrast, has relied much more heavily on consumer subscriptions and a large free tier.
SemiAnalysis says OpenAI is trying to shift. The release of the 5.5 model and Codex has helped reaccelerate its API business, and B2B APIs have become the main source of monthly net new ARR. But the consumer side still weighs on the company’s economics.
That difference shows up in reinvestment capacity. SemiAnalysis estimates that in 2027 Anthropic will have $160 billion available for reinvestment after cost of goods sold, against $92 billion for OpenAI. That gap of nearly $70 billion a year could be used to train stronger models, secure more compute contracts, and widen the technical gap.
The report also points to TaaS, or Token-as-a-Service, as an important distribution channel. Anthropic now gets about 15% to 20% of ARR from indirect sales through AWS Bedrock, Azure Foundry, and Google Vertex, up from 5% to 10% a quarter earlier. SemiAnalysis estimates the TaaS market reached $28 billion in ARR in the second quarter of 2026, with the three major cloud providers holding 85% share.
Those channels take 20% to 30% of revenue, but the report says they also provide scale, built-in compliance frameworks, and easier access to Fortune 500 customers.
Risks include budget pressure, open source competition, and regulation
SemiAnalysis does not present the story as risk-free. One issue is token budgeting. Companies including Coinbase have publicly discussed reviewing the return on AI spending. Still, the report says this tightening is concentrated among businesses that expanded too aggressively early on. Anthropic’s own data shows Claude Code enterprise users spend $150 to $250 a month on average, while 90% of users spend less than $30 a day.
Another risk is open source competition. If Google DeepMind, Meta SuperIntelligence, or other model builders release stronger coding models, token pricing could come under pressure. SemiAnalysis nonetheless says Anthropic’s net new ARR is unlikely to turn negative even in a four-way race.
The biggest systemic risk in the report is regulation. If the U.S. government restricts frontier model releases on safety grounds, Anthropic’s commercial lead could erode quickly. SemiAnalysis also flags cybersecurity as a possible next growth vertical, saying the Mythos/Fable family could accelerate ARR in the second half of the year even faster than Claude Code did. It adds that the number of customers paying more than $100,000 annually rose 7x over the past year, while customers paying more than $1 million annually rose about 42x over the past two years.
How the report gets to $6 trillion
SemiAnalysis uses a base case of 20x ARR on $300 billion of ARR at the end of 2027, implying a $6 trillion enterprise value. The model assumes monthly net new ARR rises from more than $10 billion today to $15 billion, supported by the Fable model launch, customer ramp-up, and new verticals such as cybersecurity.
The report lays out four valuation cases:
- Base case: $300 billion ARR at the end of 2027, valued at 20x ARR, for a $6 trillion enterprise value.
- Bull case: $400 billion ARR at 25x ARR, for a $10 trillion valuation.
- Conservative case: $200 billion ARR at 15x ARR, for a $3 trillion valuation.
- Bear case: $150 billion ARR at 10x ARR, for a $1.5 trillion valuation.
SemiAnalysis says the three variables that matter most are net new ARR growth, margin expansion, and how competition develops. Of those, net new ARR is the key input because it determines total ARR, the largest driver in the valuation framework.
IPO as a turning point in AI financing
The report closes by widening the frame beyond Anthropic itself. It argues that the AI industry is moving from a funding model led by venture capital and strategic investors toward one led by public markets. Private capital helped build the sector, but the scale of future compute demand is too large to be funded that way alone.
SemiAnalysis says public markets offer two tools private markets cannot match at the same scale: large follow-on equity issuance and long-term debt financing. Because data centers, power systems, and chip deployments look more like infrastructure than ordinary software expenses, they are better suited to debt structures with long asset lives.
The report also notes that Alphabet recently completed an $84.75 billion equity raise and says Meta is expected to follow. Over the next two to three years, it sees the Anthropic, OpenAI, Google, Microsoft, Meta, and Amazon ecosystem needing trillions of dollars in annual funding to build enough compute.
In that reading, Anthropic’s IPO is not just a listing. It is an early marker of AI moving from a research phase into an industrial capital phase, with profitability and access to public financing becoming central to who can keep scaling.