Anthropic's Core Concern: Revenue Can Be Withdrawn at Any Time, but Computing Power Costs Are Locked in for up to 20 Years

The latest research report from Zheshang Securities points out that, beneath the impressive data of Anthropic's exponential revenue growth, there lies a structural risk: The company's core constraint is not the continued losses per se, but the decades-long duration mismatch between the revenue side and the cost side, whereby customers can cut spending at any time, while computing power commitments have been locked in for up to 20 years.

On November 9, Zheshang Securities analyst Feng Cuiting released a research report stating that Anthropic's revenue is experiencing usage-driven exponential expansion. In the first three quarters of this year, revenue totaled $33.242 billion, an approximate year-over-year increase of 1,206%, with the operating profit margin turning positive for the first time (8.4%) in Q2, only to return to a loss (-3.3%) in Q3.

Zheshang Securities believes this volatility itself is a concrete manifestation of the aforementioned structural gap: Revenue is still mainly immediate usage-based, with Remaining Performance Obligation (RPO) covering only about 5.1 months of annualized revenue, while irrevocable commitments on the cost side have been locked in at $895.4 billion, extending up to 2045.

The research report also notes that Anthropic plans to list on Nasdaq in mid-November, and its S-1 provides the first audited public benchmark for frontier model companies. Its finalized pricing will offer domestic companies such as Zhipu and MiniMax the opportunity for valuation recovery.

Anthropic's revenue curve is showing a rare pattern of accelerated expansion.

According to the Anthropic S-1 data cited in the Zheshang Securities report, the company's revenue grew from $41 million in 2023 to $386 million in 2024 and $4.585 billion in 2025, representing year-over-year growth rates of 841% and 1,088%, respectively.

Acceleration picked up further entering 2026: Second quarter revenue reached $1.1555 billion, a quarter-on-quarter increase of 144.2%; third quarter was $1.6956 billion; aggregate for the first three quarters was $3.3242 billion, a year-over-year increase of about 1,206%, already exceeding 7.2 times the entire 2025 total revenue.

Revenue structure is primarily usage-based billing. In the first half of 2026, usage-based billing accounted for $1.3064 billion (80.2%), while subscription revenue was $322.2 million (19.8%).

The research report points out that the two are not substitutes, but additive layers. Subscription users naturally transition to incremental usage-based billing once quotas are exhausted, so the subscription proportion fluctuates narrowly each year rather than declining unidirectionally.

The core driver of revenue growth is usage. The report notes that Opus 5.5 lowered per-task customer costs by about 40% compared to the previous generation Opus 5, while Fable 5.1 reduced typical workload costs by about 25%. The coexistence of price decline and revenue acceleration indicates that usage expansion has far outpaced the rate of price decline.

However, the degree of contractualization on the revenue side remains limited. As of September 30, 2026, Anthropic's Remaining Performance Obligations (RPO) stood at only $13.8 billion, about 75% of which will be recognized within the next 12 months, covering just about 5.1 months of annualized revenue. Deferred revenue at the same time was $1.971 billion, equating to barely 0.7 months of annualized revenue.

The report thus concludes that revenue remains mostly on an immediate usage basis, with customers able to adjust spending at any time, and contract rigidity extremely limited.

In stark contrast to the high flexibility on the revenue side, costs are locked in by irrevocable multi-year commitments.

The report details eight major contracts disclosed in Anthropic's S-1/A, with a nominal size totaling about $657 billion. Including new procurement and lease agreements signed up to and after September 30, 2026, total committed scale reaches approximately $895.4 billion, with about $447.5 billion occurring before the end of 2030, and the remainder mainly concentrated between 2031 and 2045.

The report classifies these commitments into four tiers based on rigidity. The most rigid, truly binding, are four contracts that are either non-cancellable or can only be cancelled by the counterparty in case of significant breach: Broadcom $161.2 billion (finance lease with cross-default clauses), Google $114.5 billion, Amazon $110 billion, Microsoft $44.9 billion, totaling about $430.6 billion, accounting for 66% of the nominal total.

Accounting treatment further amplifies the rigidity of these constraints. The report notes that much of the computing power investment does not appear as assets on the balance sheet but is directly expensed against current profits. In the first three quarters of 2026, computing power and infrastructure costs reached $18.808 billion, 56.6% of revenue, while new fixed asset acquisition was only $925 million, less than 3% of revenue.

In other words, every $100 in revenue corresponds to about $57 in current period computing power costs, and these costs "increase as committed computing power comes online for use, not according to current period demand"—the cost pace is decoupled from actual customer usage.

The report estimates that, if computing power commitments are evenly spread across contract terms, full annualized computing expenditure would be about $11.5–12.1 billion, which is 4.6–4.8 times the annualized computing expenses for the first three quarters of 2026. If the 56.6% cost ratio of the first three quarters is maintained, annualized revenue needed to support such computing expenditure would be about $20.85 billion, approximately 4.7 times the current run-rate level.

The report explicitly offers "a different understanding": The main constraint Anthropic faces is not the losses themselves, but the duration gap between revocable revenue (RPO covers about 3 months) and irrevocable cost (with contracts as long as 20 years).

This gap is dramatically asymmetric in numbers:

• On the revenue side, the S-1 shows more enterprise clients are signing non-revocable contracts with minimum spend commitments, but overall contractualization remains low—the months of RPO coverage, by any measure, do not exceed 5.1 months.
• On the cost side, just the four most rigid contracts (Broadcom, Google, Amazon, Microsoft) total $430.6 billion, all with pay-as-agreed or make-whole mechanisms, offering very limited protection in customer default scenarios.

The report also notes that procurement models are shifting from "renting computing power" to "renting equipment", further aggravating the time-sequencing mismatch between payment and recognition.

Previously, purchases were mainly from cloud service providers (CSPs) with payments as computing power became available; now, more contracts involve direct chip leasing, hosted by third-party dedicated data centers—such as Broadcom’s five-year equipment lease. Equipment leases are treated as finance leases, with payment beginning once equipment is "ready for service". If a data center is delayed, payments may even start before computing power is available.

On the financing side, while supporting facilities are in place, all come with strict conditions. The S-1 lists Google $20 billion, Amazon $15 billion, Broadcom $42 billion and other note financing facilities, totaling $87 billion, none of which had been drawn as of end-June 2026, as drawdown is contingent on achieving specific revenue targets or computing power delivery milestones.

Therefore, the report argues that due to excessively stringent conditions, these facilities cannot serve as readily available funding assurance.

The report states that the strategic significance of Anthropic's S-1 lies in providing the first audited, comprehensive public benchmark for a frontier model company. With the S-1 now public and the company planning to list on Nasdaq in mid-November, finalized pricing will offer companies such as Zhipu and MiniMax the first overseas benchmarks for comparison, presenting opportunities for sector valuation recovery.

The report also notes, from the intelligent index dimension, that the capability gaps between frontier models have now narrowed to the scale of individual quarters.

As of early November 2026, the top seven models ranked by Artificial Analysis's intelligence index all fell within the 55 to 60 score range, with only a 5-point spread between first and last. The window of generational leadership has shortened to the period of one quarter, and a single model or reasoning grade no longer constitutes pricing power.