Anthropic’s computing push faces challenges beyond scale
Anthropic’s IPO prospectus disclosed approximately USD 900.9 billion in long-term commitments and leases, nearly all tied to computing infrastructure. Shares of the Claude developer rose 34% at Thursday’s market open, valuing the company at roughly USD 2 trillion, according to Reuters. The investment could expand experimentation, but researchers caution that computing capacity alone cannot resolve uncertainty about which approaches will improve AI.
The company’s agreement with Amazon provides for up to five gigawatts of computing capacity, while an expanded partnership with Google and Broadcom will bring additional infrastructure online starting in 2027. IBM researchers say greater capacity enables more testing of model designs and training data. Smaller, specialized models and better use of varied hardware could also reduce costs and improve efficiency.
Algorithmic innovation remains essential, according to Jürgen Schmidhuber of King Abdullah University of Science and Technology and the Swiss AI Lab IDSIA. He highlighted recursive self-improvement, in which AI systems help improve their own capabilities, while noting that full self-improvement would also require hardware advances. IBM Research’s Dan Gutfreund identified interpretability as another major obstacle: understanding how models produce outputs could help researchers identify promising approaches without relying on so many expensive experiments.