EXPLORINEER
en

US — TechCrunch

Anthropic is still grappling with the economics of compute

AI1 min read

Anthropic, an AI company, is preparing for an IPO (a first sale of shares to the public). A new filing shows its revenue is growing fast, but the cost of computing power is eating up more of that money.

What happened

Anthropic filed an update to its S-1 (a document companies file before selling shares to the public). The update included financial results through September 30, one quarter later than the earlier filing. The company made more money, but its compute costs grew even faster, so its margins got worse.

Key numbers

  • Revenue reached $17 billion this quarter, up from $11.5 billion the quarter before.
  • Compute costs rose to $10 billion for the quarter.
  • Compute took up 58% of revenue, up from 48% the quarter before.
  • Revenue for every dollar spent on compute peaked at just over 2x, then fell to 1.72x.
  • Anthropic pays SpaceX $1.25 billion a month for compute.

Why compute costs are high

Compute means the computer power needed to train and run AI models. Demand grew so fast that Anthropic had to buy short-term compute, which costs more than long-term deals. It also uses neoclouds, which are smaller companies that rent out computing power with flexible terms but at higher prices.

Why it matters

Compute is Anthropic's biggest operating expense, and the company expects it to stay that way. Investors may not worry while revenue grows quickly, but if demand slows, a weaker compute picture could become a serious problem. AI labs also depend on a huge buildout of data centers, which could slow if bond markets or local opposition get tougher.

What to watch

  • Whether Anthropic's compute margins improve as multi-year deals start.
  • How fast Anthropic's own compute project, called Anthropic Direct, grows.
  • Whether the data center buildout keeps pace with demand.
  • How investors react to the company's IPO.

Glossary

IPO
Initial public offering, the first time a private company sells shares to the public.
S-1
A filing that companies submit to regulators before selling shares to the public.
Compute
The computer power needed to train and run AI models.
Anthropic Direct
Anthropic's project to build its own compute base.

Ad

Add 150 words. Learn 10 a day.

Dicter word books send only the words you learn to your review notes.

More in this topic

AIZDNet Korea

AI Models Alone Are Not Enough: Snowflake Bets on Enterprise Data

Snowflake is competing in the AI agent market by connecting enterprise data and business systems, using models from OpenAI and Anthropic. Ahead of its virtual Snowflake Expedition 2026 event, the company unveiled "Cocoa-Cowork bots" that perform work based on company data, along with a feature for managing AI execution. Snowflake says it will also introduce a Business Ontology to standardize shared terms across departments, which will enter public preview.

AIThe Register

Forget retraining - AI can now tweak another AI's weights

MIT CSAIL researchers found that frontier AI models such as OpenAI's GPT-6 Astra can rewrite another model's weights directly rather than through training, altering its behavior. In one demonstration, Astra changed only 0.016 percent of Qwen3.5-0.8B's weights, improving MATH-500 performance by as much as sixfold and IFEval instruction-following roughly threefold, without gradient descent. The researchers call the approach "Weight-Space Programming" and plan to build a benchmark and harness for it.

AIThe Register

Open models can do 90 percent of agentic work at a twentieth of the cost of a frontier LLM: Gartner

Gartner analysts Mark Petty and Raymond Paquet argue that open-weight models can handle about 90 percent of routine enterprise agentic workflow steps at parity with closed models for a quarter of the price of the cheapest closed model. They say the best-case saving against a frontier model run on every step is up to a twentieth, though current open models are more verbose and retry more often, so the realistic saving is closer to five or six times. The analysts also predict that by 2030, 90 percent of enterprise operational agent workflow steps will run on open-weight models, and they argue that harnesses, which route work between models, are where the value in AI products lies.

AIZDNet Korea

Who Works Better? SK Telecom Employees' AI Agents Go Head to Head

SK Telecom held the finals of "AX Spark," an internal competition in which employees' self-built AI agents were judged on work performance and productivity gains. Five finalists, selected from 173 entries, are already used in real operations, including PYLON, which analyzes power usage and costs for about 180,000 telecom facilities nationwide, and M-TripleA, which tracks attack servers behind malicious smartphone apps. The company plans to extend the best cases across the organization, and it has over 5,000 internally developed work AI agents this year.