Find XO: The New Entry Barrier Set by Vertical Integration: Now You Must Make Your Own Chips (techcrunch.com)

- OpenAI and Broadcom completed tape-out of their first inference-only chip Jalapeño within nine months of design. - Following Google's TPU and Amazon's Trainium, AI companies are solidifying the trend of going from models down to silicon. - Actual deployment and cost advantages are yet unconfirmed, and training still depends on external GPUs. - Nonetheless, startups feel the pressure first. - Prepare structures in advance to avoid being subject to the pricing and policies of model operators.