Các công ty khởi nghiệp AI của Mỹ đua nhau phát triển các phiên bản thay thế giá rẻ hơn các mô hình Trung Quốc giữa bối cảnh khó khăn trong huy động vốn

The Huoxing Finance report states that on August 2, according to The Wall Street Journal, as Chinese open-weight models such as Kimi, Qwen, and DeepSeek approach the performance of top U.S. models at significantly lower costs, Silicon Valley and Washington are increasingly concerned that Chinese models may long-term compress profits for U.S. AI companies. U.S. startups like Arcee AI, Reflection AI, and Poolside are racing to develop local alternatives to meet user demand for low-cost, downloadable, and customizable models. However, U.S. open-weight model companies face difficulties in fundraising. Some investors question whether freely open models can generate stable revenue and worry that such technologies may undermine the value of their investments in OpenAI and Anthropic. In the first quarter of 2026, AI startups raised a total of $255.5 billion, with nearly two-thirds coming from just three funding rounds by OpenAI, Anthropic, and xAI. Despite limited funding, Arcee AI used 2,048 NVIDIA Blackwell B300 chips with a budget of approximately $20 million to complete 33 days of pre-training and launch Trinity Large. The model remains smaller than top-tier models and lags behind OpenAI and Anthropic on multiple benchmarks, but the company plans to develop larger models through a new round of funding. NVIDIA has become a major supporter of the U.S. open AI ecosystem, not only developing the Nemotron series of models but also investing in Reflection AI, Poolside, and Thinking Machines Lab. Industry insiders note that the U.S. open-weight ecosystem remains relatively small, and Chinese models still hold an overall advantage.