Reflection AI introduced its Beam model on October 5, 2026. The sparse model features 501 billion parameters, with 23 billion active at any time. It was trained on 23.8 trillion tokens using 10,500 Nvidia GB300 GPUs over a four-week period.
The training cost for Beam is estimated at $100 million. This highlights a significant pricing gap compared to Chinese models, which reportedly cost only a few million dollars to train. Beam claims to offer 3–4x better inference efficiency than the Chinese GLM 5.2 model.
Nvidia is actively supporting US AI startups through equity investments and infrastructure access. By mid-2026, the company held approximately $99 billion in AI startup equity. Additionally, Nvidia is developing financing channels intended to mobilize over $500 billion for compute resources.
In August 2026, Nvidia finalized a $6 billion licensing and investment agreement with Poolside to foster an open-weight model ecosystem in the US. The long-term competitiveness of American labs may depend on whether these efficiency gains can offset higher initial training expenses.
