How Much Revenue Can Anthropic Generate for Every $1 Spent on Computing Power?

TL;DR
· The revenue generated for every $1 spent on computing power and infrastructure costs is projected to rise from $0.40 in Q1 2025 to $2.08 in Q2 2026, before falling back to $1.72 in Q3. Of the approximately $9.9 billion spent on computing power that quarter, about $5 billion came from higher-cost short-term computing.
· Quarterly revenue increased approximately 45 times over six quarters, reaching $17 billion in Q3, but the quarter-on-quarter growth rate dropped from 144% to 47%, with the quarterly revenue increment also decreasing from $6.8 billion to $5.4 billion.
· About 80% of revenue is billed based on usage. Claude Code allows the same group of employees to delegate more tasks to the model, and clients can also shift some calls to lower-priced models.
· Models released in the past six months contributed to over 80% of revenue in Q3. Anthropic needs to maintain its premium with new versions while recouping training investments before older versions are phased out.
· The revised prospectus discloses approximately $900.9 billion in contractual obligations, almost all related to computing power and infrastructure. Chiltern estimates that the minimum annual payments for computing contracts from 2027 to 2030 will be around $77 billion to $82 billion.
· Anthropic's IPO pricing range is set at $70 to $78 per share, with $74 corresponding to an approximate fully diluted equity valuation of $1.33 trillion. Chiltern's segment valuation is $69 per share, with a benchmark discounted cash flow valuation of $98, and a return of computing revenue ratio above 1.8 times is listed as one of the conditions for increasing positions.

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

This valuation raises a specific question. How much revenue can Anthropic generate for every $1 spent on computing power? Revenue in Q3 grew 47% quarter-on-quarter to $17 billion, but the revenue per dollar of computing power fell from $2.08 to $1.72, and the adjusted operating profit margin also declined. Most clients pay based on actual usage, yet Anthropic signed computing and leasing contracts years in advance. Whether the paid usage can keep pace with the new capacity delivery will determine whether the next batch of computing will enhance profits or increase costs.

Chiltern is a heavy-paying client of Claude and uses Claude to assist in reading the prospectus, building valuation models, and drafting research. The institution is willing to initially build a half position within the pricing range and lists the recovery of revenue per dollar of computing to above $1.80 as one of the conditions for increasing positions.

How Much Revenue Can Be Generated for Every $1 Spent on Computing Power?

Chiltern divides the quarterly revenue by the corresponding computing and infrastructure costs, with the denominator including training and inference expenses. In Q3 2026, $17 billion in revenue divided by approximately $9.9 billion in computing costs equals about $1.72. The company also needs to pay channel fees, personnel, and other operating expenses from this revenue.

In Q1 2025, Anthropic generated only $0.40 in revenue for every $1 spent on computing power. By Q2 2026, this figure rose to $2.08. The effects of scaling are also reflected in the cost ratio. The costs of computing and infrastructure fell from 160% of total revenue in 2025 to 57% of revenue in the first nine months of 2026.

In Q3, this ratio fell back to 1.72 times, and the company’s adjusted operating profit margin also dropped from 19% to 15%. Of the approximately $9.9 billion spent on computing that quarter, about $5 billion came from higher-cost short-term computing. Meanwhile, the quarterly revenue increment decreased from $6.8 billion to $5.4 billion. Whether subsequent long-term computing deliveries can lower unit costs and whether the new capacity can be fully utilized by paid demand will determine if the ratio can recover.

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

Claude Code Brings More Usage, Lower Prices Pressure Revenue Per Call

Anthropic's revenue for 2023 is projected to be $41 million, increasing to $4.6 billion in 2025 and reaching $33.2 billion in the first nine months of 2026. Claude Code is a significant driver of this growth. It can execute multi-step tasks within codebases. After integrating access permissions, project context, and engineering workflows, the same group of employees can delegate more work to the model. According to the company, the annualized revenue run rate for Claude Code increased from approximately $500 million in September 2025 to over $2.5 billion in February 2026. The run rate is an annualized level derived from recent revenue.

About 80% of Anthropic's revenue is charged based on usage, with enterprises and developers paying based on the tokens (units of processing text and other content) read and generated by the model, while the remaining approximately 20% mainly comes from subscriptions. Even with little change in the number of seats, an increase in tasks can lead to more revenue. This structure also allows revenue to fluctuate with model pricing and customer purchasing decisions. As of the end of September, remaining performance obligations amounted to $13.8 billion, with approximately three-quarters expected to be recognized within a year. The revenue share of the top five clients dropped from 36% in Q1 to 21% in Q3, and the migration of major clients will still affect quarterly performance.

Opus 4.5, launched in November 2025, is about 67% cheaper than its predecessor, with the revenue growth rates for the following two quarters reaching 132% and 144%, respectively. Chiltern views this as evidence that price reductions are releasing demand. By Q3 2026, the company estimates that price reductions will decrease revenue for that quarter by approximately $600 million. New usage must be sufficient to offset the decline in unit price and cover the increased inference costs for these requests to improve the revenue output per dollar of computing power.

Channels will also change the revenue recognition approach for this ratio. In the first nine months of 2026, approximately 44% of revenue was sold through cloud marketplaces like AWS, Google, and Microsoft. Anthropic recognizes related revenue on a gross basis, with channel fees accounted for separately. Chiltern estimates that if revenue were recognized on a net basis, the revenue for the past 12 months would be approximately 15% lower, with EBITDA remaining unchanged. When comparing model companies using revenue multiples, the accounting approach for channels will also affect the results.

New Models Contribute 80% of Revenue, Technical Premium Relies on Continuous Iteration

The prospectus shows that over 80% of revenue in Q3 2026 came from models released in the past six months. According to benchmark tests compiled by Chiltern, Opus 5.5 leads in comprehensive capability metrics in Artificial Analysis and Epoch. In the tests of Artificial Analysis, the cost per task for Opus 5.5 is approximately $5.98, while GPT-6.1 Sol is about $0.72, with a difference of 6 scoring points. How much enterprises are willing to pay for this price difference relates to Anthropic's ability to recoup the next round of training investments.

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

Security reviews, export restrictions, and corporate governance will also affect the pace of new model releases. The prospectus discloses that the long-term benefit trust can appoint four of the seven directors, with the founding entity holding a fixed 50.1% voting power. Chiltern believes that delays in releases or restrictions on the scope could alter the revenue growth trajectory. The institution views coworking, industry tools, and international markets as potential sources of growth, with the benchmark scenario primarily supported by existing sales.

Long-term Contracts Lock in Payments, Demand Determines Capacity Utilization

The revised prospectus discloses total contractual obligations of approximately $900.9 billion, almost all related to computing power and infrastructure, including approximately $112.3 billion in commitments signed after the end of September. Within the 12 months following the end of September, approximately $70.7 billion is expected to be paid. Chiltern estimates that by the end of 2030, cumulative maturities will amount to approximately $452 billion. The prospectus discloses that approximately $186.2 billion of the commitments are conditional or allow for termination before initiation.

Excluding cancellable portions, Chiltern estimates that the minimum annual payments from 2027 to 2030 will be around $77 billion to $82 billion. When demand does not keep up, idle computing power will still incur contractual expenses. In Chiltern's model, if the revenue generated per dollar of computing power continues to be below approximately $1.50, Anthropic will incur losses under adjusted metrics. The institution's benchmark scenario expects that as computing power is concentrated in deliveries, this ratio may first drop to 1.51 times in Q4, before recovering to 1.87 times in 2028.

Funding and computing power often come from the same group of companies. Google participates in funding while leasing TPU computing power to Anthropic. Amazon is both an investor and a supplier of Trainium computing power. Broadcom has also agreed to provide up to $42 billion in note financing, with funds limited to paying related rents. Chiltern expects that after the IPO, Anthropic will have approximately $216 billion in cash at year-end. Its stress test assumes revenue remains at an annualized $86 billion calculated for Q4, showing that by the end of 2031, there will still be approximately $147 billion in liquidity. Under this set of assumptions, a slowdown in growth will first pressure profits and valuations.

Buying at $74 Requires a Recovery in Computing Efficiency

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

Chiltern's benchmark scenario expects revenue of $133 billion in 2027, reaching $194 billion in 2028, with the adjusted operating profit margin rising to 25%. Based on an exit at the end of 2028, the benchmark scenario suggests approximately $120 per share, an optimistic scenario of about $279 per share, a price war scenario of about $34 per share, and a scenario of losing technological leadership pressure of about $18 per share. The institution assigns weights of 45%, 20%, 25%, and 10% to the four scenarios, resulting in a probability-weighted exit price of approximately $120 and a 35% probability of loss.

Chiltern is willing to initially build a half position at $70 to $78 per share, accounting for approximately 0.8% of the portfolio's net value. Its benchmark expectation for Q4 revenue is $21.5 billion. Revenue exceeding $22.5 billion, with revenue per dollar of computing recovering to above $1.80, or a stock price dropping below approximately $66, could support increasing positions. If it rises to about $92 on the first day, the expected annualized return from its model drops to 13%, below the set threshold of 15%. The first financial report after listing will test whether Claude's usage can bring nearly $2 in revenue for every dollar of computing power after long-term computing gradually replaces high-cost short-term capacity.