Recently, with Anthropic releasing Fable5.1 and OpenAI launching the Astra large model, the arms race in the global large model field has escalated again. Although Astra has sparked widespread online discussions due to its outstanding performance and cost-effectiveness, the massive financial and hardware consumption behind it has once again highlighted the significant gap between China and the United States in terms of computing infrastructure for large models.
According to reports from U.S. media, although OpenAI has not disclosed specific specifications of Astra, the model relies on its vast StarGate program, using over 100,000 NVIDIA graphics cards for training. Based on the current mainstream B200 graphics card price (approximately $50,000 to $80,000), the hardware cost for just purchasing graphics cards alone is as high as $6 billion. If all related costs, including servers, high-speed networks, power supply, cooling systems, and daily R&D, are taken into account, the total investment in the AI infrastructure is estimated to be between $10 billion and $25 billion.
Compared to this, the investment scale of domestic large models in computing infrastructure still faces objective gaps. Taking the DeepSeek project's new AI data center construction in Inner Mongolia as an example, the project plans to purchase 160,000 Huawei AI graphics cards, with a total investment of about $2.56 billion. Even at the full budget level, this scale is only a quarter of the conservative investment scale of Astra-related infrastructure.
Not only is there a gap in the absolute amount of capital investment, but the computing architecture and ecosystem also face multiple challenges. On one hand, domestic AI chips still need to catch up with top-tier hardware in single-card computing power and ecosystem maturity; on the other hand, industry-wide estimates show that some domestic chips still have a certain generation gap in actual computing efficiency compared to NVIDIA's high-end products, and overall procurement and supply remain in a continuous tight state. In the past two years, U.S. companies like OpenAI have continuously maintained a vast computing power reserve of several ten thousand to even hundreds of thousands of top-tier graphics cards, and subsequent models with stronger performance are already being trained intensively.
This arms race, driven by billions of dollars and massive computing power, clearly reflects the harsh reality of competition in the deep waters of large models. Facing the stage-specific gap in computing resources and infrastructure, the domestic AI industry and related chip supply chain still need time to accelerate technological breakthroughs and ecological improvements.
