Anthropic’s IPO Bet: Can a Frontier AI Lab Convince Wall Street That AGI Is Worth Trillions?
Anthropic's S-1 IPO filing reveals a company that has flipped from negative 94 percent to positive 85 percent gross margins on inference in two years, while targeting a valuation of $1.5 to $2 trillion. The filing is ultimately a bet that AGI-driven automation of white-collar work will arrive fast enough to justify one of the largest public market debuts in history.
Anthropic Files Its S-1 — And the Numbers Are Staggering
When a company files an S-1 — the regulatory document required before listing on a public stock exchange — it lifts the curtain on finances that were previously invisible to the world. For Anthropic, one of the most closely watched frontier AI labs on the planet, that moment has arrived. The company filed its S-1 to list on the Nasdaq, and analysts are already projecting a valuation somewhere between $1.5 trillion and $2 trillion. If that range holds, Anthropic would immediately rank among the most valuable companies ever to go public.
But a valuation number, however jaw-dropping, is just a headline. The real story lives in the hundreds of pages of financial disclosures inside that document. To unpack what those pages actually reveal, Nilay Patel on Decoder brought in Joey Brookhart, an analyst at SemiAnalysis and co-lead of the firm’s Tokenomics model. The full conversation is available at The Verge, and it is essential listening for anyone trying to understand the true economics of building frontier AI.
What the S-1 Actually Reveals
SemiAnalysis had already attempted to model Anthropic’s finances from the outside before the filing — a task that required piecing together press reports and inference cost estimates. Now, with audited numbers in hand, the picture has snapped into much sharper focus.
According to Brookhart, Anthropic has reached over $70 billion in annualised recurring revenue (ARR) remarkably quickly, and the gross margins, compute margins, and free cash flow metrics are described as exactly what investors would want to see. The filing triggered a rally across semiconductor and neocloud stocks, which is itself a signal of how deeply the broader compute supply chain is tied to the fortunes of labs like Anthropic.
Perhaps the most striking data point is the gross margin trajectory on inference. Two years ago, Anthropic was operating at negative 94 percent gross margins on inference — meaning it was losing money on virtually every token it served. Today, that figure stands at a positive 85 percent. That is an extraordinary reversal, and it did not happen by accident.
The Engine Behind the Margin Improvement: Tokenomics
Brookhart’s Tokenomics framework treats AI businesses the way traditional analysts treat manufacturing businesses — by focusing on unit economics. For AI labs, the fundamental unit is the token, and the key metric is how much revenue a lab can extract per unit of compute power.
As of the time of the S-1 filing, Anthropic was generating approximately $65 million of revenue per inference megawatt. A year earlier, that figure was in the mid-teens. Brookhart projects that with future model generations running on Nvidia Blackwell hardware, that number could reach $100 million to $120 million per megawatt. Meanwhile, Anthropic’s compute costs sit at just over $10 million per megawatt in cost of goods sold — a gap that creates enormous operating leverage.
Token throughput is central to this story. Anthropic runs a diverse chip fleet that spans Trainium chips, TPUs, and GPUs. Token throughput on Trainium chips alone is reportedly up five to ten times year-over-year. Both hardware-level efficiency gains and improvements in how models interact with the underlying silicon are driving this. The implication is that Anthropic can serve more tokens for the same or lower cost with each passing year, compressing the denominator of the unit economics equation.
Where the Compute Actually Goes
One detail that surprises many observers: only about 30 percent of Anthropic’s total compute is devoted to inference — the revenue-generating activity of serving responses to users and enterprise customers. The remaining 70 percent goes toward model training, post-training, experimentation, and research. This is the investment in the next model, and it is where, as Brookhart frames it, the company’s real defensibility is built.
This structure explains the strategic logic of Anthropic’s business. High inference margins fund the training compute that produces better models. Better models attract more customers and higher prices. More revenue funds even larger training runs. The flywheel, if it spins, is self-reinforcing. The S-1 is, in many ways, Anthropic’s argument to Wall Street that the flywheel is already spinning.
The AGI Assumption at the Heart of Everything
None of the financial projections in the S-1 make sense without accepting a set of ambitious assumptions about what AI will become. Anthropic CEO Dario Amodei has spoken openly about a path to $1 trillion in annual revenue. Brookhart is candid about what that requires: the market for AI cannot remain as narrow as it is today.
Right now, Anthropic’s revenue is heavily concentrated in coding applications. It is, in Brookhart’s framing, almost a single-vertical company. The top one percent of its customers are spending $50,000 to $100,000 per employee. Broader enterprise adoption — the Fortune 500 and Global 2000 companies that drive massive IT budgets — has barely begun. Brookhart notes that the major AI labs have not yet experienced a full enterprise Q4 budgeting cycle in any meaningful way, since ARR only began accelerating at the start of the year.
Reaching $1 trillion in revenue requires a fundamental shift: AI must move from augmenting white-collar work to replacing or significantly displacing portions of it. The S-1 implicitly asks investors to believe that this transition is coming — and that Anthropic will be positioned to capture a disproportionate share of the value created.
The Competitive Threat Is Real
Wall Street would be naive to ignore the competitive landscape. In the first half of the year, Anthropic captured an extraordinary 90 percent of the industry’s net new ARR. By the third quarter, OpenAI had rebounded to take what Brookhart estimates was the majority of net new ARR. Meanwhile, the open-source inference market grew from roughly $1 billion in annualised token costs in Q1 to over $10 billion by Q3 — a ten-fold increase in a single year.
The pressure from open-weight models is particularly significant. Many enterprise customers — law firms are cited as one example — are actively choosing to run AI on-premises rather than send sensitive data to a third-party lab. Custom training and post-training on open-weight models is growing among AI-native companies as well. Switching costs in coding, one of Anthropic’s core verticals, are low: enterprises can and do move between providers relatively easily.
Well-funded competitors are not standing still either. DeepMind, Meta’s Superintelligence Labs, and others are investing heavily in training compute. The implication is that the moat Anthropic is building must come from model quality, trust, safety reputation, and enterprise relationships — none of which are permanently defensible.
What This Means for the Broader AI Economy in India and Beyond
For Indian enterprises and technology leaders watching this IPO, the S-1 contains signals worth reading carefully. The codification of AI economics — the move from opaque private valuations to audited public numbers — raises the bar for every company that buys or builds AI. At a projected valuation of $1.5 trillion to $2 trillion (roughly ₹127 lakh crore to ₹170 lakh crore), Anthropic is making a claim about the size of the prize that, if believed, justifies enormous infrastructure investment across the global compute ecosystem.
The vertical adoption story is also instructive. Coding led the first wave. Cybersecurity, biology, and enterprise white-collar automation are positioned as the next frontiers. Indian IT services firms, which have built their model on human labour in precisely those white-collar categories, will need to reckon with the S-1’s implicit argument: that this automation is not speculative, but already underway.
The Bottom Line
Anthropics’s S-1 is not just a company going public. It is a comprehensive argument — made in the language of audited financials and regulatory disclosures — that AGI is a tractable goal, that the economics of pursuing it are sound, and that the company chasing it deserves to be among the most valuable on earth. Whether Wall Street accepts that argument will depend on how quickly the market moves from narrow coding use cases to the broad white-collar transformation that the $1 trillion revenue vision requires.
The numbers have improved dramatically. The margins are real. The flywheel exists. The question Anthropic cannot yet answer — and neither can anyone else — is whether the world it is betting on will arrive fast enough to justify the valuation it is asking investors to assign today.
