Written by: Trend Research

On June 1st, Anthropic secretly submitted its S-1 to the SEC.

Five weeks later, the well-known research institution SemiAnalysis published a lengthy report, using its own model to piece together the complete financial statements of the company from the ground up based on product lines, pricing tiers, and customer types. Wall Street refers to this kind of work as shadow accounting, and the conclusion can be summed up in one sentence: Anthropic could become the first company in human history to reach a market value of $6 trillion.

What does this number mean? It is more expensive than any publicly traded company on Earth, including Nvidia. SemiAnalysis is not known for sensationalism; the organization started by breaking down chip supply chains and data center costs. So where does its conclusion come from?

An extremely steep curve

First, let's look at the three core numbers in the report.

Annual recurring revenue (ARR): $9 billion by the end of 2025, currently over $60 billion, nearly seven times in half a year.

Monthly net new ARR: $3 billion in January of this year, jumping to $11 billion in March.

Profit: By the third quarter of 2026, GAAP basis earnings before interest and taxes (EBIT) will exceed $1 billion.

For reference, during the pandemic, Zoom, which set a record for growth in the SaaS industry, took about four years to reach $1 billion ARR. Anthropic took 11 months, and then stacked several billion on top over the following 17 months.

The steeply rising engine has a name: Claude Code.

This AI programming tool will launch mid-2025, and the report estimates that it currently accounts for over 7% of all code submissions on GitHub. In other words, for every 14 lines of code submitted by programmers worldwide, 1 line comes from it.

The data on customer retention is even more staggering; the report cites the metrics revealed by Anthropic's CFO, indicating a net revenue retention rate (NRR) of 500%: the batch of customers contributing $30 billion in ARR this year previously contributed only $2 billion a year ago. Existing customers grow on their own, which is the most valuable kind of growth in the software industry.

From money-burning machine to money-printing machine

There are many fast-growing AI companies, but the real highlight of this report lies in the profit statement.

SemiAnalysis estimates that Anthropic's overall gross margin will rebound from -94% in 2024 to around 65% currently, with an API business gross margin exceeding 80%. In two years, this machine has transformed from the most aggressive money-burning furnace in the industry to a money-printing machine with gross margins comparable to mature software companies.

The secret lies in inference efficiency.

Measured by ARR per megawatt of computing power, this figure was $16 million nine months ago; it will reach $60 million later this year. The same amount of electricity is generating nearly four times the revenue.

In November last year, when Opus 4.5 was released, it slashed the price to one-third of its predecessor, and the market initially thought it was a price war; SemiAnalysis provided the opposite judgment at the time: after the price drop, Anthropic's profit margin on this product line was actually higher because the production cost per token dropped faster.

This scene has historical parallels.

In April 2015, Amazon first disclosed AWS's financial figures separately, and Wall Street suddenly discovered that this money-burning machine contained a high-margin business, rewriting the valuation logic for Amazon. The day Anthropic's S-1 goes public could very well mark the AWS moment for the AI industry.

OpenAI has become the opposite in the mirror

The sharpest part of the report is a set of mirror comparisons using OpenAI as a reference.

In Anthropic’s revenue structure, 75% to 85% comes from usage-based billing APIs, while consumer subscriptions only account for 5%; in contrast, OpenAI has it reversed, with subscriptions being the majority, backed by hundreds of millions of free users.

Free users act as a moat in the narrative, but form a bottomless pit in the cost structure: the report hypothesizes that if both companies reach $100 billion in ARR, OpenAI's gross profit will be about $25 billion less than Anthropic due to the need to support free users. This difference directly determines who has more ammunition to train the next generation of models. By 2028, the report estimates that the cumulative EBIT gap between the two will widen to $250 billion.

The report used a provocative phrase: Anthropic is capable of making OpenAI dance to its own tune. Going public first is the first step, transferring the pressure of public financial disclosure and market scrutiny intact to the competitor, with rumors of OpenAI's IPO being postponed to 2027.

The real motive behind going public: buying electricity and computing power

Why now? The report calculated a computing power forecast.

By 2030, the combined unconstrained computing power demand from Anthropic and OpenAI will exceed 100 gigawatts. The reality is that the industry's net new computing power will only be 2.5 gigawatts and 5 gigawatts in 2025 and 2026 respectively, and the combined available computing power both companies currently have is just over 6 gigawatts. There is a 94-gigawatt gap between demand and supply, which can only be filled with one material: money, and cheap money.

The window is closing. Alphabet has completed $84.75 billion in equity financing, and Meta is reportedly planning for a $100 billion-level fundraising; no matter how big the appetite of the capital market for AI infrastructure, it cannot withstand several trillion-dollar giants extracting profits in succession. Whoever goes public first secures the ammunition. This IPO is more like a margin for computing power futures than a milestone.

Cracks in the shadow accounting

Having praised the report, it's time to discuss the parts it didn't mention.

The endpoint value of $6 trillion is based on an extremely thin assumption: that monthly net new ARR stabilizes at $15 billion without decline, pushing all the way to a $300 billion ARR by the end of 2027. The acceleration from $3 billion to $11 billion indeed occurred, but drawing a two-year straight line from a three-month slope is one of the oldest tricks in sell-side research. Any quarterly slowdown would shrink the $6 trillion to $4 trillion or $3 trillion, changing the nature of the story.

The accounting itself has a hidden flaw.

Forecasting agency FutureSearch cites data from Sacra, indicating that Anthropic records revenue by counting the end customer spending on cloud platform resale at full value, while platform splits are counted as expenses. If required to report on a net basis during IPO due diligence, the book ARR could potentially shrink by 20% to 40% in one go.

There is also a stance issue. SemiAnalysis disclosed that its annual expenditure on Claude tokens once reached $10.95 million; it is one of Anthropic's most intensive paying customers, using Claude to research Claude's parent company. While the model can be extremely detailed, the data can’t determine the light that shines where.

Regulatory risks are no longer hypothetical. In June, a U.S. government export control order caused Anthropic's recently launched flagship models Fable 5 and Mythos 5 to be taken offline globally. The report placed regulatory risk at the end of the risk warning section, but in reality, it just fired a shot last month.

There will be a day when the secret offering document is unveiled, likely in the fall or winter of this year.