The Filing That Opened the AI Lab's Books
On October 26, 2026, Anthropic submitted its preliminary S-1 registration statement for a proposed initial public offering, applying to list on the Nasdaq Global Select Market under the ticker symbol ANT. The document ran over 300 pages, and for the first time, the public received a comprehensive accounting of how a pure-play AI laboratory generates and spends money at a scale that few companies in history have reached.
The numbers were striking. Anthropic reported $16.286 billion in revenue for the first six months of 2026, compared to $1.163 billion in the same period of 2025 a fourteen-fold increase. The second quarter alone brought in $11.56 billion, nearly triple the $4.7 billion from the first quarter. That kind of sequential growth, sustained across consecutive quarters, is rare in any industry. In AI, it is nearly unprecedented.
But alongside the revenue, the filing disclosed infrastructure commitments that raised immediate questions about sustainability and margin structure. To understand what Anthropic disclosed and what it means for prospective investors requires tracing a thread that runs through every line item: the cost of compute.
What Compute Means in an AI Context
Before examining the numbers, it helps to define the resource at the center of Anthropic's business model. Compute, in the context of artificial intelligence, refers to the processing capacity required to train and run large language models. Training a frontier model demands thousands of specialized chips working in parallel over weeks or months. Serving that model to millions of paying customers requires ongoing computational resources each time a user prompts Claude. Both phases consume enormous amounts of electricity and hardware capacity.
Anthropic's S-1 describes compute and infrastructure expense as the largest operational cost category. For the six months ended June 30, 2026, that line item totaled $8.945 billion. The company also separately reports cost of revenue $2.380 billion for the same period which covers the compute costs substantially associated with serving paying customers, plus certain personnel costs and allocated overhead.
The two figures describe different accounting categories, not a precise workload-by-workload breakdown. Anthropic manages computing capacity as a shared pool, directing it among research, model development, customer service, and internal use. The S-1 does not give a dollar allocation for each workload, which means outside analysts must estimate how much of the total compute spend goes specifically to customer-facing inference versus frontier training.
"Anthropic reported $8.945 billion in compute and infrastructure expense for the six months ended June 30, 2026. Its $2.380 billion cost of revenue is a separate accounting category, made up substantially of compute costs for serving customers, plus certain personnel costs and allocated overhead."
The Four Providers Holding the Keys
One of the most striking disclosures in the S-1 concerned the distribution of Anthropic's computing infrastructure across major technology partners. The filing reveals that Anthropic has distributed its computing needs evenly over four major providers: Broadcom, SpaceXAI, Google, and AWS. Each holds between $110 billion and $161 billion in commitments, together accounting for approximately four-fifths of the $637 billion total committed compute spend.
This concentration matters for several reasons. First, it means Anthropic's cost structure is deeply intertwined with the capacity decisions and pricing power of a small number of counterparties. Second, it suggests a strategic decision to avoid single-provider dependency while maintaining leverage in negotiations. Third, it raises questions about what happens if one provider experiences supply constraints or if the economics of chip production shift.
The commitment structure itself is revealing. The S-1 discloses $265.6 billion in fixed or minimum guarantees as of June 30, 2026, with a further $245 billion in agreements executed since that date. These are not speculative future orders they represent contractual obligations that persist regardless of Anthropic's actual revenue trajectory. As one analysis noted, the company carries essentially no drawn debt. The $15 billion revolving credit facility is undrawn, and the only outstanding notes are $6.2 billion of non-interest-bearing Amazon convertibles. The financial leverage, in other words, lives in the commitments rather than on the balance sheet.
From $0.40 to $2.08: The Efficiency Metric That Changed
Perhaps the most closely watched figure in the filing was the ratio of revenue to compute spend. Anthropic did not state this ratio directly; analysts calculated it from the quarterly results disclosed in the prospectus. The calculation shows that Anthropic generated $2.08 in revenue for every dollar spent on compute and infrastructure in the second quarter of 2026. Just five quarters earlier, in the first quarter of 2025, that same ratio stood at $0.40.
That fivefold improvement in efficiency tells a particular story about how AI economics evolve as companies scale. Revenue grew 1,368% year over year in Q2 2026, while compute costs rose 300%. The gap between those growth rates revenue accelerating faster than infrastructure spending is the engine behind the margin improvement Anthropic highlighted.
The efficiency gains are not accidental. The S-1 describes how the September 2026 release of Claude Fable 5.1 reduced the cost of typical workloads by approximately 25% compared to Fable 5, with savings reaching up to 45% for complex coding and highly agentic tasks. The simultaneous release of Opus 5.5 cut customers' cost per task by approximately 40% compared to Opus 5. The company noted that when Opus 4.5 launched in November 2025 offering more capability and a lower cost per task than its predecessor revenue growth accelerated from 48% quarter over quarter to 132% the following quarter.
This pattern where efficiency improvements drive adoption, which drives revenue growth, which improves unit economics is what analysts describe as Jevons paradox in the real world. Lower costs do not simply preserve margins; they expand the addressable market in ways that compound the revenue effect.
The Profit Milestone and What Surrounds It
The second quarter of 2026 marked Anthropic's first quarterly operating profit: $974 million, an 8.4% GAAP margin and 19% on an adjusted basis. For a company that posted an $8.1 billion operating loss in 2025 which included a $1.5 billion copyright settlement the swing represents a dramatic inflection point. The loss narrowed every quarter of 2026 before turning positive.
Yet investors reading the S-1 encounter a paradox in the same section. The net loss for Q2 was still $66.3 billion, driven by fair-value charges on convertible instruments. This is not a measure of operational performance but rather an accounting consequence of revaluing financing instruments as the company's implied valuation changed. The $80.748 billion net loss for the first half of 2026 is real accounting, but as one detailed analysis observed, it is a poor standalone description of the underlying business most of that loss stems from revaluing financing instruments rather than paying to run Claude.
Understanding this distinction matters for anyone evaluating Anthropic's financials. The operating income figure reflects what the business earns from customers minus what it costs to serve them. The net loss reflects the paper consequences of equity and convertible valuations. They tell different stories about the same company.
The Commitment Gap: Fixed Costs Meets Flexible Demand
The central question that analysts have drawn from the S-1 is what happens when committed compute meets flexible demand. Anthropic is selling services that customers can often scale down an enterprise that uses Claude for internal tools may reduce usage during a hiring freeze or budget tightening. The infrastructure contracts that power those services, however, can last for years.
This mismatch between flexible customer spending and committed computing costs represents the structural risk at the heart of Anthropic's IPO case. The company proposes to sell public investors an economic interest while keeping decisive shareholder voting power with its founders and majority board-election rights with its Long-Term Benefit Trust. Public shareholders would buy economics with limited control.
Exceptional demand can make that arrangement enormously productive. Slower demand, cheaper competing models, or an expensive safety intervention can expose its rigidity. The filing supplies evidence for both outcomes.
On the demand side, the customer metrics are remarkable. Anthropic sells Claude to more than 11,900 customers who spend over $100,000 per year. Net Revenue Retention stands at approximately 900% as of September 30, 2026. No single customer exceeded 10% of revenue for the nine months to September 30, 2026, and the top five customers together represented about 25%. That concentration has fallen meaningfully from 2025, when customers spending over $100,000 accounted for 76% of revenue a sign that usage is broadening across the customer base.
On the commitment side, the $637 billion in total compute obligations with two-thirds described as non-cancelable represents a fixed cost structure that does not contract when revenue does. Violeta Todorova, senior research analyst at Leverage Shares, told CNBC that it is unclear how much it costs to generate each additional dollar of revenue as computing demand increases. That question is central to any margin analysis.
How Anthropic's Compute Costs Compare to Industry Peers
Direct comparisons between AI laboratories are complicated by differences in disclosure practices, business models, and accounting treatments. OpenAI's reported figures, where available, show the company grew approximately 253% in 2025 meaningful growth, but less than a quarter of Anthropic's 11.9-fold expansion. The derived gross margin for Anthropic stands at approximately 40%, though different accounting treatments and cost allocations make precise peer comparisons difficult.
What is clear from the available data is that both companies operate at a scale where compute costs represent the dominant variable expense. The efficiency ratio revenue per dollar of compute spent provides one lens for comparison. Anthropic's path from $0.40 to $2.08 over five quarters suggests that as AI companies move from research-focused operations to scaled customer-serving infrastructure, the economics can improve dramatically. Whether that trajectory continues as frontier model training costs rise and competition intensifies remains an open question.
What This Means for IPO Investors
For prospective public investors, the Anthropic S-1 presents a company of remarkable growth speed and improving unit economics, set against a commitment structure that locks in substantial fixed costs for years. The bull case is straightforward: a business that grew 11.9-fold on a 40% derived gross margin and has just begun earning an operating profit is worth a very large number. The case against is equally simple: $637 billion of compute commitments against uncertain revenue trajectories in a market where model prices are falling and competition is intensifying.
The governance structure adds another layer of complexity. The filing reveals that CEO Dario Amodei and his six co-founders each own over 200 million shares, which could be worth billions if Anthropic hits the over $1 trillion valuation it is expected to achieve. Under the proposed structure, public shareholders would buy economics with limited control over the decisions that shape those economics.
For readers evaluating whether to participate in the offering, the key variable is not the historical growth rate but the trajectory of the revenue-to-compute ratio. If that metric continues improving if Anthropic can continue extracting more revenue from each dollar of infrastructure spend the fixed commitments become easier to service. If efficiency plateaus or if revenue growth slows as the company reaches market saturation, the commitment structure becomes a weight on returns.
The Road Ahead: Efficiency or Compression?
Anthropic's S-1 also reveals a company actively managing its cost structure through model efficiency. The September 2026 releases of Claude Fable 5.1 and Opus 5.5 both featuring substantial cost-per-task reductions suggest a deliberate strategy to pass savings to customers while improving its own unit economics. When lower costs drive higher adoption, which drives higher revenue, the cycle reinforces itself.
The third quarter of 2026 brought a moderating signal. Revenue growth slowed to approximately 47%, following approximately 144% growth in the second quarter. Whether this represents normalization after explosive growth or the beginning of a deceleration phase is not yet clear from the available data.
What is clear is that Anthropic has moved from a research organization funded by venture capital to a scaled enterprise with real customers, real revenue, and real margin a trajectory that few AI companies have completed. The question the S-1 poses is whether the next chapter is written in the language of compounding efficiency or structural compression. The $637 billion in commitments suggests the company is betting on the former.
Where to Read Further
For readers who want to explore Anthropic's S-1 filing directly, the SEC registration statement provides the full financial statements and expense definitions. The analysis by Tomasz Tunguz on Anthropic S-1 Analysis: How 7 Key Metrics Stack Up offers a detailed breakdown of the seven metrics that matter most in evaluating the offering. For a perspective on the compute commitment structure and governance implications, Visionary Talks' analysis of profits, compute bills, and control traces the filing from its headline loss to Note 8 on conditional financing. And for those interested in the efficiency metrics underlying the business model, Crypto Briefing's calculation of revenue per compute dollar shows how the ratio has evolved across quarters.