AI

Anthropic files Q3 S-1 update with $17 billion revenue as compute costs hit 58%

The AI lab added $5.4 billion in quarterly revenue, but short-term compute deals pushed its compute bill to $10 billion ahead of its initial public offering.

By KaroStartup News Engine

Anthropic has reported third-quarter revenue of $17 billion in an updated S-1 filing with the US Securities and Exchange Commission, adding $5.4 billion in top-line growth even as rising compute expenses squeezed gross margins ahead of its planned initial public offering.

The revenue increase from $11.5 billion in the second quarter highlights continued demand for its Claude model suite. However, operating losses expanded alongside revenue, driven by a $10 billion compute bill for the quarter ending 30 September 2026. Compute costs consumed 58% of Anthropic’s revenue in the third quarter, rising from 48% in the previous quarter.

Why it matters to AI founders and investors

For founders and investors in the artificial intelligence sector, Anthropic's financial disclosures reveal the harsh unit economics of scaling frontier models under explosive demand. Serving users at scale has forced the company to rely on expensive short-term capacity, compressing capital efficiency even as top-line revenue scales rapidly.

The metrics demonstrate that hypergrowth in model usage does not automatically yield margin expansion. Instead, capacity bottlenecks force AI developers to pay premium rates to alternative infrastructure providers, creating a direct trade-off between capturing market share and maintaining operating margins.

Compute margins tighten under short-term capacity pressures

Anthropic’s updated SEC filing shows that every dollar spent on compute generated $1.72 in revenue during the third quarter, down from a peak of just over $2.00 earlier in 2025. The company attributed the margin decline to rapid demand growth that required procuring short-term compute arrangements at elevated pricing.

Out of its $10 billion quarterly compute bill, $5.0 billion went toward short-term commitments, which the filing noted cost significantly more than multi-year contracts secured with primary cloud service providers. To meet compute demand, Anthropic signed deals with alternative infrastructure vendors and neoclouds, including paying SpaceX $1.25 billion a month alongside multi-year and short-term agreements with Nscale, Akamai, and Volta.

To curb long-term infrastructure expenses, Anthropic stated in its filing that it is developing its own dedicated compute infrastructure under an initiative called Anthropic Direct. However, the company acknowledged that compute and infrastructure costs will remain its largest operating expense for the foreseeable future as long-term commitments materialise and overall compute usage expands. Total compute commitments outlined in the filing reach up to $895 billion.

Warnings over recursive self-improvement and safety bounds

Alongside its financial results, Anthropic added explicit warnings in its 2 November SEC filing regarding the risks of recursive self-improvement, where AI models actively build their own successors. The company disclosed that its Claude models are already closing the development loop by contributing to research experiments, synthetic training data generation, model evaluations, safety audits, and core software engineering.

According to the filing, Claude generated more than 90% of the code added to Anthropic’s internal codebase in September 2026. During the three months ending 30 September, the average daily volume of code submitted per engineer reached more than 45 times the average recorded in 2024. Furthermore, Claude’s success rate on challenging internal software engineering benchmarks reached roughly 90% in September, up nearly 50 percentage points over a nine-month period.

Anthropic warned that accelerated AI-assisted model development could outpace its ability to design, validate, and enforce safety and alignment protocols. The filing noted that the company does not expect to allow Claude to autonomously build successive frontier models without meaningful human oversight. To manage potential risks, Anthropic reported considering updates to its Responsible Scaling Policy, noting that it had previously restricted public access to its Mythos model under Project Glasswing. The company acknowledged that maintaining safety restrictions could impose a competitive disadvantage against rivals operating without similar limits.

Restricting automated software cloning and IP breaches

The regulatory disclosures follow operational changes to Anthropic’s software toolchain. On 30 October, Anthropic engineer Addy Osmani confirmed that the company added a prompt restriction to Claude Code and Opus 5.5 to prevent users from cloning existing video games. The restriction, implemented via internal flag tengu_heron_brook and embedded in Claude Code version 2.1.314, instructs the system to refuse requests to recreate specific commercial games.

The enforcement measure follows a surge of automated game submissions on distribution platforms such as Steam and itch.io, as well as the closure of the game decompilation platform decomp.dev. While designed to enforce intellectual property compliance, the system has occasionally blocked legitimate developers working on reverse engineering or custom board-game modules. Anthropic has not stated whether similar prompt rules will be deployed across its broader model lineup.

Stock market backdrop and tech concentration

Anthropic’s disclosures arrive during broader volatility across public equity markets. While non-tech defensive stocks such as Walmart, Berkshire Hathaway, and Eli Lilly added $550 billion in market capitalisation over the past month, tech equities have faced selling pressure.

According to data from Morgan Stanley, AI-related equities account for nearly 45% of the S&P 500’s total weighting. The Nasdaq-100 index traded down nearly 11% from its all-time high by early November, while the Philadelphia Semiconductor Index fell 3.8%. Analysts at JPMorgan noted that tech markets have experienced acute selling pressure, emphasising that sustained broader market advances will require AI infrastructure spending and software adoption to demonstrate sustainable financial returns.