A technician in a cleanroom suit walking through a high-tech semiconductor laboratory.

TSMC approves up to $1 billion in Anthropic and OpenAI stakes

TSMC’s board cleared investments of up to $500 million each in Anthropic and OpenAI, alongside $40.9 billion in new capacity spending.

TSMC’s board on 10 November approved investments of up to $500 million in Anthropic common shares at the company’s IPO price, and up to $500 million in OpenAI preferred stock — a combined ceiling of $1 billion. The same meeting cleared about $40.93 billion in capital appropriations for advanced process capacity, packaging, fabs and 2027 research, plus up to $500 million to set up an investment company for semiconductor-ecosystem bets. The foundry that manufactures the industry’s most advanced chips is now taking equity in two of the labs that will keep buying them. Anthropic is preparing a public listing; OpenAI remains private, so TSMC’s stake there would be preferred stock rather than listed shares.

What else matters

Google DeepMind’s AlphaGenome 2 predicts variant effects in specific cells

Google DeepMind on 11 November released AlphaGenome 2, which predicts how DNA variants affect specific human cell types and states rather than tissue-wide averages. Trained on more than 100 million cell profiles across 68 tissues, it finds nearly twice as many expression-changing variants as the prior model. It is available for non-commercial research via the AlphaGenome API, and is not approved for clinical diagnosis.

Perplexity ships Decider v1.3 with span extraction

Perplexity on 12 November released Decider v1.3 and Decider Flash v1.3, open-weight models for classification, routing and ranking. A new locate feature returns matching text spans with character offsets. Both are available on Hugging Face under an Apache licence, and through OpenRouter and Perplexity’s Decisions API, at the same prices as v1.2.

Baidu opens ERNIE 5.5 Flash Preview

Baidu on 11 November released ERNIE 5.5 Flash Preview, a 280 billion-parameter mixture-of-experts model with about 14 billion parameters active per token and a 1 million-token context. The API is live on Baidu’s Qianfan platform. Weights are promised under Apache 2.0 with no date, and Baidu calls this an intermediate checkpoint as post-training continues.