On February 23, Anthropic did something the AI industry almost never does.

They named names.

In a public post, they accused DeepSeek, Moonshot AI, and MiniMax of running large-scale distillation campaigns to extract Claude’s capabilities.

Anthropic says these efforts involved roughly 16 million exchanges across tens of thousands of fraudulent accounts, coordinated through proxy infrastructure. The goal wasn’t casual experimentation. It was systematic capability harvesting.

That’s a strong claim. And the fact that it’s being made publicly is just as significant as the claim itself.

Distillation isn’t controversial on its own. Every major lab distills its own frontier models into smaller, cheaper systems. That’s normal.

What Anthropic is describing is different: using a competitor’s model as a training signal.

The examples they give are specific. Repetitive prompts designed to elicit detailed reasoning traces. Structured grading tasks that effectively turn Claude into a reward model. Extraction of agentic tool-use patterns and coding behaviors at scale.

One detail is particularly telling. Anthropic says that when they released a new model, one of the labs redirected traffic toward it within 24 hours to begin extracting capabilities from the updated system.

That suggests something operational, not incidental.

If you’ve been following the AI race over the past two years, most competition has been framed around scale. Who has more compute. Who has better data. Who can train larger models more efficiently.

This announcement shifts the focus.

It suggests that inference endpoints themselves are strategic assets. That API access isn’t just about developer growth, but about capability control.

It also complicates how we interpret model progress. When we see rapid improvements from labs around the world, we often assume independent breakthroughs. Anthropic’s framing introduces another variable: acceleration through extraction.

That doesn’t automatically validate or invalidate anyone’s progress. But it adds nuance to how we think about the competitive landscape.

There’s also the safety layer. Anthropic argues that distilled models may preserve capability while stripping away safeguards. Whether you fully agree with that framing or not, it moves the conversation into national security and policy territory.

This isn’t just product competition anymore.

There’s a quiet contradiction in modern AI.

Labs want open APIs. They want integration, ecosystem growth, and developer adoption. But as models become more capable, each output becomes more valuable as potential training data.

At normal scale, that’s usage.
At industrial scale, it becomes something else.

Anthropic describes building behavioral fingerprinting systems, traffic classifiers, and coordinated detection mechanisms. The language sounds closer to cybersecurity than SaaS operations.

We’re watching model endpoints evolve from “developer tools” into guarded infrastructure.

That’s a meaningful shift.

Shortly after the post went live, Elon Musk responded publicly, saying Anthropic has trained on stolen data at massive scale and has paid large settlements related to copyright issues. His point was simple. It is hard to accuse others of extraction when the broader industry is still facing open questions about how training data has been collected.

That response changes the tone of the conversation. Anthropic is talking about model distillation and capability control. Musk is bringing the focus back to the ongoing debate around scraping, copyright, and data consent. Both discussions sit at the core of how modern AI systems are built, and neither one is fully resolved.

After headlines spread that Claude Code could help automate COBOL modernization, shares of IBM dropped roughly 13 percent in a single session.

Markets move for many reasons, but the signal was clear. IBM’s legacy modernization and consulting business is deeply tied to mainframe and COBOL infrastructure. If frontier models can compress that work meaningfully, investors reassess revenue durability. IBM took a pretty big hit today -13% 📉📉📉

Whether Claude can actually deliver at scale is still an open question. But the reaction shows how quickly AI capability claims translate into real market pressure.

What we’re watching is not just a dispute between companies.

It’s the AI stack maturing in public.

Model access is becoming strategic. Training data is still legally unsettled. Inference endpoints are treated like infrastructure. And markets are reacting in real time to capability claims that may or may not materialize at scale.

For those of us building, investing, or operating in this space, the takeaway is simple.

The frontier is no longer just about bigger models.

It is about control, diffusion, defensibility, and who captures value once intelligence becomes widely replicable.