Arcee AI says it trained Trinity Large on 2,048 Nvidia Blackwell B300 chips with a roughly $20 million budget, as Chinese open-weight rivals intensify pressure on U.S. AI economics.
Arcee AI, a little-known Silicon Valley startup, used much of its remaining cash in late 2025 to build what it described as the strongest open-weight AI model it could. The company said it completed the effort in a 33-day pretraining run, training Trinity Large on 2,048 Nvidia Blackwell B300 chips with a roughly $20 million budget. The result is a low-cost AI system that users can download and customize, an approach that can broaden adoption by letting developers adapt the software to their own needs. The push comes as low-cost Chinese open-weight models including Kimi, Qwen and DeepSeek draw closer to leading U.S. systems, raising concerns in Silicon Valley and Washington about mounting pressure on American AI profits. Other U.S. startups including Reflection AI and Poolside are also pursuing domestic alternatives, though funding remains tight.