TL;DR

The income generated from every 1 dollar invested in computing power and infrastructure costs rises from 0.40 dollars in Q1 2025 to 2.08 dollars in Q2 2026, before falling back to 1.72 dollars in Q3. Of the approximately 9.9 billion dollars spent on computing power that quarter, about 5 billion dollars came from the more costly short-term contracts.

Single-quarter income grows to about 45 times over six quarters, reaching 17 billion dollars in the third quarter, but the quarter-on-quarter growth rate slows from 144% to 47%, and the incremental income also decreases from 6.8 billion dollars to 5.4 billion dollars.

About 80% of revenue is billed based on usage. Claude Code enables the same group of employees to delegate more tasks to the model, and clients can also assign some calls to models with lower prices.

Models released in the past six months contributed to over 80% of revenue in the third quarter. Anthropic needs to maintain premiums with new versions while recouping training investments before older versions are phased out.

The revised prospectus reveals approximately 900.9 billion dollars in contractual obligations, almost all related to computing power and infrastructure. Chiltern estimates the minimum annual payments for computing power contracts from 2027 to 2030 will be about 77 to 82 billion dollars.

Anthropic's IPO pricing range is set at 70 to 78 dollars per share, with 74 dollars corresponding to an approximate fully diluted equity valuation of 1.33 trillion dollars. Chiltern's segment valuation results in approximately 69 dollars per share, with a benchmark discounted cash flow valuation of 98 dollars, and a return to a computing revenue multiple above 1.8 is one of the conditions for increasing positions.

Editor's Note: On November 10, Claude developer Anthropic updated its prospectus intending to list on NASDAQ at 70 to 78 dollars per share. At 74 dollars, Chiltern Capital’s fully diluted equity valuation is approximately 1.33 trillion dollars. During the same period, Google and NVIDIA plan to subscribe 10 billion dollars each, and seven strategic investors collectively subscribe for 26 billion dollars. Companies providing funding to Anthropic are also selling chips, computing power, and cloud services to it.

This valuation raises a specific question. How much income can Anthropic generate for every 1 dollar spent on computing power? In Q3, revenue increased by 47% quarter-on-quarter to 17 billion dollars, but income corresponding to each dollar of computing power fell from 2.08 to 1.72 dollars, with adjusted operating margins also declining. Most clients pay based on actual usage, yet Anthropic signed computing and leasing contracts years in advance. Whether the paid usage can catch up with new capacity delivery will determine if the next batch of computing power will improve profits or inflate costs.

Chiltern is a heavy paying client of Claude, also using Claude to assist in reading the prospectus, building valuation models, and drafting research. The organization is willing to initially build a half position within the pricing range and considers a rebound of revenue to above 1.8 dollars per dollar of computing power spent as one of the conditions for increasing positions.

The following is the original text:

How much income can be generated for every 1 dollar spent on computing power?

Chiltern calculates the single-quarter income divided by the corresponding computing power and infrastructure costs, where the denominator includes training and inference expenditures. In Q3 2026, 17 billion dollars of revenue divided by approximately 9.9 billion dollars of computing costs equals roughly 1.72 dollars. The company also needs to pay channel fees, personnel, and other operating expenses from this income.

In Q1 2025, Anthropic only generated 0.40 dollars in revenue for every dollar spent on computing power. By Q2 2026, this figure rose to 2.08 dollars. The effect of scale expansion is also reflected in the cost ratio. The costs of computing power and infrastructure fell from 160% of total revenue in 2025 to 57% for the first nine months of 2026.

In Q3, this ratio fell back to 1.72 times, and the company’s adjusted operating margin also decreased from 19% to 15%. Of the approximately 9.9 billion dollars in computing expenses that quarter, about 5 billion dollars came from costlier short-term contracts. Meanwhile, the quarterly revenue increment fell from 6.8 billion dollars to 5.4 billion dollars. Whether subsequent long-term computing deliveries can lower unit costs and whether new capacity can meet paid demand will determine if the ratio can recover.

According to Chiltern's breakdown of the cost structure, for every 100 dollars in revenue, about 15 dollars are used for inference, that is, running the models and processing customer requests, while about 40 dollars are used for training and research computing power. The more expensive and rapidly iterating the next-generation models are, the more the company needs to recoup training investments in a shorter time. Of the approximately 104 billion dollars in net losses over the first three quarters, about 94 billion dollars came from non-cash fair value changes of financing instruments like convertible bonds, while operating cash flows for the same period were approximately 6.6 billion dollars positive. To assess the profitability of daily operations, training expenditures, channel fees, and equity incentives also need to be considered.

Claude Code brings more usage, price drops reduce income per call

Anthropic's income for 2023 is projected to be 41 million dollars, growing to 4.6 billion dollars in 2025, and reaching 33.2 billion dollars in the first nine months of 2026. Claude Code is a significant driver of this growth. It can perform multi-step tasks within codebases. With access to permissions, project context, and engineering processes, the same group of employees can delegate more work to the model. According to the company, the annualized revenue run rate for Claude Code rose from about 500 million dollars in September 2025 to over 2.5 billion dollars in February 2026. The run rate is an annualized level projected from recent revenues.

Approximately 80% of Anthropic's revenue is based on usage charges, where businesses and developers pay according to the tokens (units of processing text and other content) read and generated by the model, while the remaining around 20% primarily comes from subscriptions. Even if the number of seats changes little, more and longer tasks contribute to higher revenues. This structure also allows revenues to fluctuate with model pricing and client purchasing decisions. As of the end of September, remaining performance obligations totaled 13.8 billion dollars, with about three-quarters expected to be recognized within a year. The revenue share from the top five clients fell from 36% in Q1 to 21% in Q3, indicating continued impacts on quarterly performance from key clients' transitions.

Opus 4.5, released in November 2025, is about 67% cheaper than its predecessor, and the quarter-over-quarter revenue growth rates in the following two quarters reached 132% and 144%, respectively. Chiltern views this as evidence that price reductions are stimulating demand. By Q3 2026, the company estimates that lowered prices could reduce quarterly income by about 600 million dollars. New usage must be sufficient to offset the price decline and cover the increased inference costs for these requests in order to improve income per dollar of computing power.

Channels will also change how this ratio's income is recognized. In the first nine months of 2026, approximately 44% of the revenue was sold through cloud marketplaces of AWS, Google, and Microsoft. Anthropic recognizes related revenues on a gross basis, with channel fees accounted for separately. Chiltern estimates that if revenues were recognized on a net basis, the past 12 months' revenue would be about 15% lower, with EBITDA unchanged. When comparing models companies using income multiples, how channel accounting is recognized will also impact results.

New models contribute 80% of revenue, tech premiums depend on continuous iterations

The prospectus shows that over 80% of revenue in Q3 2026 comes from model versions released in the past six months. According to Chiltern's compiled benchmarks, Opus 5.5 leads in comprehensive ability metrics on Artificial Analysis and Epoch. In tests for Artificial Analysis, the cost per task for Opus 5.5 is about 5.98 dollars, while GPT-6.1 Sol is around 0.72 dollars, a difference of six scoring points. How much businesses are willing to pay for this price difference will affect whether Anthropic can recoup the next round of training investments.

Open weight models now account for about 70% of token usage in OpenRouter and Vercel gateways, while closed-source models still account for 73% and 77% of expenditures in the two gateways, respectively. Anthropic's share of spending on Vercel has decreased from 64% in August to 44% in September. Vercel's clients tend to be startups, so the representativeness of monthly data is limited. If cheaper models continue to narrow the capability gap, Claude will need to maintain premiums by handling higher-value tasks.

Safety reviews, export restrictions, and corporate governance can also influence the pace of new model releases. The prospectus reveals that the long-term interests trust can appoint four of seven directors, and the founder entity holds a fixed 50.1% voting power. Chiltern argues that delaying releases or limiting the scope of open availability could impact revenue growth trajectories. The organization views coworking spaces, industry tools, and international markets as potential sources of growth, with the base scenario primarily supported by existing sales.

Long-term contracts lock in payments; demand determines if capacity can be fully utilized

The revised prospectus discloses total contractual obligations of about 900.9 billion dollars, almost all related to computing power and infrastructure, including approximately 112.3 billion dollars in commitments signed after September. Over the next 12 months from the end of September, approximately 70.7 billion dollars is expected to be payable. Chiltern estimates that by the end of 2030, cumulative maturities will be about 452 billion dollars. The prospectus reveals that about 186.2 billion dollars of commitments come with conditions or allow for termination before initiation.

Excluding cancellable portions, Chiltern estimates minimum annual payments from 2027 to 2030 will be about 77 to 82 billion dollars. When demand does not keep pace, idle computing power will still incur contractual costs. In Chiltern's model, if revenue per dollar of computing remains below approximately 1.50 dollars, Anthropic will incur losses on an adjusted basis. The institution’s baseline scenario forecasts that as computing power is increasingly delivered, this ratio may drop to 1.51 times in the fourth quarter before recovering to 1.87 times in 2028.

Funds and computing power often come from the same set of companies. Google is contributing funding while renting TPU computing to Anthropic. Amazon is both an investor and a supplier of Trainium computing power. Broadcom has also agreed to provide up to 42 billion dollars in note financing, with funds limited to covering related rents. Chiltern estimates that after the IPO, Anthropic’s year-end cash will be approximately 216 billion dollars. Its stress test assumptions include revenue stagnation at an annualized 86 billion dollars for Q4, showing there will still be about 147 billion dollars in liquidity by the end of 2031. Under this set of assumptions, a growth slowdown would first depress profits and valuations.

Buying at 74 dollars requires a rebound in computing efficiency

At 74 dollars, Chiltern estimates Anthropic's enterprise value to be about 8.4 times its projected revenue for 2027, while the median of 16 comparable publicly traded companies is 6.6 times. Comparing usage-based business with AI computing companies, and subscription business with high-growth software companies, the segment valuation comes to about 69 dollars per share. The benchmark discounted cash flow valuation stands at 98 dollars. Deriving from benchmark profit margins, 74 dollars corresponds to approximately 406 billion dollars in revenue by 2035, indicating a compound annual growth rate of about 25% from 2026 to 2035.

Chiltern’s base scenario anticipates revenue of 133 billion dollars in 2027, reaching 194 billion dollars in 2028, with adjusted operating margins rising to 25%. Based on a 2028 exit, the base scenario leads to about 120 dollars per share, while optimistic and price war scenarios yield about 279 dollars and 34 dollars, respectively. The scenario where technological leadership is lost projects a severe downside to about 18 dollars. The institution assigns weights of 45%, 20%, 25%, and 10% to the four scenarios, leading to an approximately 120 dollars probability-weighted exit price and a 35% probability of loss.

Chiltern is willing to initially build a half-plan position at 70 to 78 dollars per share, representing about 0.8% of the portfolio's net value. Its baseline expectation for Q4 revenue is 21.5 billion dollars. Revenues exceeding 22.5 billion dollars, rising income to above 1.80 dollars per dollar of computing power, or a stock price dropping below approximately 66 dollars could all support increasing positions. If the price rises to around 92 dollars on the first day, the expected annualized return from its model drops to 13%, below the set 15% benchmark. The first earnings report post-listing will need to assess whether Claude's usage can bring in nearly two dollars of income for each dollar of computing power as long-term contracts gradually replace higher-priced short-term capacities.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到[email protected]，本平台相关工作人员将会进行核查。