AI Sector Daily Digest: July 21, 2026

In brief

1. China considers tighter controls on AI models and chips

• Chinese authorities are considering adding AI and semiconductor technologies to a future catalogue of restricted or prohibited exports. 1
• The proposals could limit overseas access to model weights and training data, restrict foreign production of advanced chips based on Chinese designs, and cover strategic technologies such as agentic AI.
• The measures are not final: Reuters said the report came from the Financial Times, regulators are still weighing industry feedback, and Reuters could not independently verify it.

2. Samsung creates a robotics division reporting to its CEO

• Samsung Electronics said its new RX, or "Robotics eXperience," division will report directly to the chief executive and own the company's robotics strategy, core technology, and commercial execution. 2
• Executive Vice President Lee Dongkun will lead the robotics strategy team after working on robotics strategy at Hyundai Motor Group, including the direction of Boston Dynamics.
• Samsung plans research hubs in the United States, China, and Japan, with humanoid robots aimed first at manufacturing and later at homes and retail.

3. Moonshot pauses new Kimi K3 subscriptions after a capacity crunch

• Moonshot AI temporarily stopped new consumer subscriptions after requests for its newly launched Kimi K3 exceeded forecasts and approached the limits of its existing compute clusters. 3
• Kimi K3 is a 2.8-trillion-parameter open-weight model focused on coding and agent-style tasks; Moonshot said current paid users would remain unaffected.
• The service issue arrives as Moonshot seeks up to $2 billion in new capital and prepares for a possible Hong Kong listing, with Reuters reporting a June valuation of $30 billion.

4. Anthropic gets final approval for a $1.5 billion copyright settlement

• A federal judge in San Francisco approved Anthropic's settlement with authors who accused the company of using their books to train Claude without permission. 4
• The deal follows an earlier ruling that model training on the books was fair use but that Anthropic infringed authors' rights by storing more than 7 million pirated books in a central library.
• More than 91% of covered authors and publishers had claimed their share, but authors who opted out have separate lawsuits still pending.

5. A Cornell-CMU model finds weak AI rules can backfire

• Researchers from Cornell and Carnegie Mellon modeled how safety requirements affect general-purpose AI providers and companies that build products on top of their models. 5
• In the model, a low safety bar applied only to downstream companies made products less safe than under no regulation because upstream providers could shift the safety burden downstream.
• The paper found that requiring both sides of the pipeline to meet defined safety targets could improve safety and profits, but the authors say the model is simplified and has not yet measured real-world policy effects.

What to watch