An investor compares Chinese AI labs' use of model distillation to athletes doping. While it provides a short-term performance boost, it risks long-term harm by disincentivizing fundamental, groundbreaking research. This creates a dependency on external models rather than fostering true innovation.
China's domestic IPO market for AI has been surprisingly robust, with local LLM developers like Zhipu AI and Minimax trading at significantly higher price-to-sales multiples than Anthropic and OpenAI. This indicates strong local investor appetite and a potential valuation disconnect between the two markets.
China has implemented a comprehensive AI regulatory framework that is arguably the world's most stringent. Unlike in the West, China's primary concern is ensuring AI does not threaten social stability. The state must approve all consumer-facing AI products, inspecting models and training data beforehand.
To improve reliability of NVIDIA's Vera Rubin systems, CoreWeave developed its own hardware: 'Racky' (a rack manager) and 'Velvi' (a liquid cooling manager). These proprietary tools provide better signals and automation, helping manage the complexity of new, highly integrated GPU racks and speed up deployment.
Despite Anthropic's vocal criticism and efforts to close loopholes, such as random ID verification, the gray market for its models persists in China. Resellers continually find new ways to bypass restrictions, highlighting the difficulty of enforcing regional access controls against a determined and decentralized network.
Data center projects are often built to a target 'yield on cost,' a key real estate metric. With bond yields rising to 8-9%, the cost of financing inflates the project's denominator. This squeezes profit margins to the point where new projects may become unaffordable for borrowers.
A massive cottage industry of small-scale resellers has emerged in China to provide access to restricted Western AI models. They use reverse proxies, bulk email signups, and cryptocurrency to circumvent geographic and financial blocks, creating a sprawling gray market that is difficult to police.
The shortage of advanced NVIDIA chips from U.S. export controls prevents Chinese labs from conducting large-scale experimental training. This compute deficit is a primary driver for their reliance on illicitly distilling knowledge from U.S. models, effectively forcing them to 'copy the homework' to stay competitive.
The red-hot market for financing AI data centers is cooling. Banks are 'filled up' with exposure and becoming more selective, while bond yields for these projects have risen from ~6% to 8-9%. This dual-market pressure makes it harder and more expensive for developers to fund new capacity.