Anthropic's S-1: $70B ARR, 85% Margins, and a Bet on AGI

Anthropic has filed its S-1 to go public on the Nasdaq at a valuation between $1.5 trillion and $2 trillion, with trading possible as soon as next month — and the document's numbers, dissected this week on The Verge's Decoder podcast by SemiAnalysis analyst Joey Brookhart, describe a frontier AI lab that reached $70 billion-plus in annualized recurring revenue (ARR) while burning only $10 billion of cash.

The interview, recorded days after the filing, offers the first detailed outside read of a document running hundreds of pages. Brookhart, co-lead of SemiAnalysis' Tokenomics practice (a research model of AI unit economics, from hyperscaler spending down to per-megawatt revenue), built a model of Anthropic's finances from the outside in July. The audited numbers, he said, largely validated the bullish case.

What does the S-1 actually say about Anthropic's finances?

"It was an awesome S-1," Brookhart told Decoder host Nilay Patel. "Some people were somewhat disappointed in the September [ARR] numbers, but what we saw here was a company that has efficiently scaled to an amazing level, with a $1 trillion-plus valuation, $70 billion-plus of ARR in a very short amount of time."

The headline figures:

• 85 percent positive gross margins on inference, up from negative 94 percent two years ago.
• 11 percent non-GAAP EBITA margins — a level Brookhart says Anthropic itself did not expect.
• $120 billion-plus in cash and marketable securities on the balance sheet, with roughly $100 billion more expected post-IPO.
• Only $10 billion of total cash burned to reach this scale.
• 900 percent net revenue retention, an extreme figure by software standards.

The widely circulated $80 billion loss in the first half of 2026, which dominated comment sections, is a non-cash accounting charge. "It's a non-cash charge because their valuation has gone up so much and they've issued warrants or convertible notes," Brookhart explained. "If you don't know accounting, it looks bad, but it's because they've grown so much and their valuation has gone up so much."

Where does the efficiency come from?

Anthropic rents nearly all of its compute from SpaceXAI, Google and Amazon rather than owning data centers. Brookhart attributed the margin gains to a diverse chip fleet spanning Trainium, TPU and GPU systems, with token throughput on Trainium chips up 5x to 10x year-over-year.

The core metric is revenue per inference megawatt. Anthropic now generates roughly $65 million per megawatt — mid-teens millions at this time last year — against compute costs of just over $10 million per megawatt. Brookhart and his Tokenomics co-lead Max Kan have published research suggesting Opus 5.5 on Nvidia Blackwell hardware could reach $100 million to $120 million per megawatt.

Only about 30 percent of Anthropic's compute serves inference revenue. The remainder funds training, post-training and experimentation. Brookhart framed that split as the strategic engine: "That's more of the R&D side and where you actually get your returns and defensibility in this business."

Can Anthropic reach $1 trillion in revenue?

CEO Dario Amodei has said he can see a path to $1 trillion in revenue, and the S-1's growth case rests on penetration of white-collar work that remains nearly untouched. Anthropic disclosed it serves 50 percent of the Fortune 500, yet that spending equals roughly 0.1 percent of those customers' total operating expenses — about $8 a month per user excluding cost of goods sold, or sub-$50 including them. The top 5 percent of Claude Code users spend $18,000 a year on average, while non-Claude Code users spend well under $75 a month.

"To get to $1 trillion, we do have to move to some of that white collar work, where it goes from AI augmentation to labor saving or replacement in some areas," Brookhart said, citing Microsoft's expanded buyout offers to employees as evidence of the shift.

Today's market remains narrow. The two leading labs hold about $120 billion of combined ARR, driven disproportionately by the top 1 percent of customers spending $50,000 to $100,000 per employee. Coding dominates. "It's almost a single vertical company," Brookhart said, pointing to cyber — with eight-figure-plus contracts — and bio as the next verticals, and noting Anthropic's beta cyber program gives away 25,000 in credits per month.

Why stay at the frontier?

One S-1 figure frames the competitive reality: 80 percent of revenue in the last three months came from models released in the last six months. A frontier lab must keep shipping frontier models, because in adversarial verticals like cyber, "if an adversary is using a frontier model, you need to be defending with a frontier model," Brookhart said.

The competitive picture shifted this year. Anthropic captured 90 percent of the industry's net new ARR in the first half; OpenAI took the majority in Q3 after releasing ChatGPT 5.6 and Astra. Open-source token inference ARR grew from roughly $1 billion in Q1 to $10 billion-plus in Q3. SemiAnalysis itself moved from 90-plus percent Anthropic usage down to 70 percent after the OpenAI releases, then back above 90 percent with Opus 5.5.

What about the compute commitments and dilution?

Analyses of the S-1 cite $500 to $600 billion in compute commitments over 7 to 10 years — some readings put the figure above $800 billion. At roughly $11 million per megawatt in current spend, Brookhart argued the commitments imply a manageable gigawatt footprint, and at 20 gigawatts future commitments would run "well into the trillions" anyway. "They're not in any dire situations there," he said.

The Anthropic-Amazon dynamic drew scrutiny, since AWS is simultaneously supplier, distribution channel and beneficiary. Some 48 percent of Anthropic's business flows through Bedrock, Foundry and Vertex, where customers pay through existing cloud contracts and the clouds take a rev share — a structure Brookhart said is lifting margins across AWS, Azure and Google Cloud, with AWS posting an exceptionally strong Q2 on the back of Anthropic's Bedrock growth.

On dilution, Brookhart was blunt: all tech companies dilute, and "if you believe in this long-term AGI bull case... you're probably looking at a $10 trillion-plus business. They can outrun dilution at that point."

What happens next — for Anthropic and OpenAI?

Anthropic's weak September stemmed from identifiable churn: Meta, likely over 10 percent of revenue, cut Claude Code seats from 60,000 to 30,000, shifting volume to internal models, and Anthropic cracked down on anti-distillation usage that had represented 5 to 7 percent of revenue. Brookhart's view is that October's post-Opus 5.5 acceleration has restored a Q4 pace of $8 billion to $10 billion in monthly net new ARR, after Q3's drop from Q2's $10 billion to $4 billion.

OpenAI, 60 percent consumer and 40 percent enterprise at the end of Q1 before flipping enterprise-majority, delayed its own IPO. Its enterprise business grew 100 percent quarter-over-quarter in Q3 against 77 percent for the total business, but Brookhart does not expect its financials to match Anthropic's when its S-1 lands, possibly in early 2027 — the Wall Street Journal has reported negative 100 percent EBITA margins there versus Anthropic's positive 11 percent.

The structural risk, Brookhart said, is timing: whether revenue arrives fast enough in 2028 to justify the return on more than $2 trillion of data center capital expenditure, in a market now debating whether a compute glut looms. The roadshow, he expects, will see Anthropic seesawing between the long-term AGI vision — risk factors include the possibility that Claude is conscious — and deliberate short-term expectations management ahead of its first public quarter.