Anthropic · LLM WIKI

AI safety company founded in 2021 by Dario Amodei, Daniela Amodei, and other former OpenAI researchers. Its best-known product line is Claude.

Origins in the OpenAI Split

Anthropic emerged from an internal conflict at OpenAI over safety and company direction. The founding team believed OpenAI was moving too quickly and becoming too commercial. The result is a company with a strong research DNA, which creates a distinctive tension: the people running the company are researchers who would often rather be doing research.

The Compute Bottleneck (2026)

Anthropic’s biggest structural weakness is not technical. It is a GPU allocation problem. Three groups are competing for the same compute:

Research needs GPUs to train new frontier models, reportedly including a “god model.”
Product needs GPUs for Claude.ai, Claude Code, and enterprise tooling.
Users are paying customers consuming the capacity they bought.

The issue is that Anthropic invested too late and too little in its own GPU infrastructure. While OpenAI bought nearly everything available, Anthropic and Google are relying more on partnerships such as Amazon collaboration and Google-financed data centers. The result is chronic undercapacity, visible in 95% uptime, frequent outages, and the 2026 rate-limit crisis.

Project Glasswing

Project Glasswing is Anthropic’s strategic answer to part of that security landscape: not simply broader model access, but controlled deployment of high-end capability to defensive teams. The core idea is uncomfortable but important. If a frontier model can find vulnerabilities faster than most humans, the bottleneck is no longer just model quality. It becomes access policy and deployment policy.

That makes Glasswing less of a classic product launch and more of a coordination mechanism: partnerships, compute credits, limited access, and an explicitly defensive purpose. For this wiki, it is a useful example of how “AI security” is not just a slogan but an organizational question.

Valuation and Growth

2024: $100 million revenue
2025: $1 billion revenue
2026 ongoing: $14 billion annualized
Customers paying more than $100k per year: 7x growth over the last year
Fortune 10: 8 of 10 are Claude customers

At a secondary-market valuation of $600 billion, Anthropic is growing faster than OpenAI, but the infrastructure is not keeping up.

Why Anthropic Is Not Being Commoditized

Ben Thompson argues in Agents Over Bubbles (March 2026) that agents require an integrated architecture of model plus harness, and that is exactly Anthropic’s moat. Claude Code is not just an API wrapper around Claude. It is software that steers the model, verifies results, and iterates. That integration is not easily swapped out.

You can see the same dynamic in Microsoft’s dependence on Claude Code/Anthropic for Copilot Cowork, even though Microsoft publicly wants to remain model-agnostic. Enterprise agents need more than raw models.

Critique: Anthropomorphization as Brand Strategy

AI Anthropomorphization is especially pronounced at Anthropic. @atmoio shows in a video analysis of the Claude Mythos system card how the “Impressions” section, 20 pages of employee amazement at model output, is marketing rather than science, and how the model produces consciousness-adjacent language because Anthropic’s own blog posts about consciousness are part of the training data. Circular validation packaged as a finding.

Connections

Claude Code - flagship developer tool; the integrated harness is the differentiator
OpenAI - comparison point and former employer of the founders
Project Glasswing - controlled defensive use of frontier capability
Dario Amodei - the main executive behind the safety and research agenda
Nina Schick - shows how the topic is interpreted externally as an AI inflection point
Theo Browne - critical but fair observer
Ben Thompson - argues for Anthropic’s structural edge via model-plus-harness integration
AI Anthropomorphization - structural pattern in Anthropic’s communication
@atmoio - video critique of the Claude Mythos system card