The real AI race may no longer be at the frontier

This summer, the AI landscape witnessed a significant shift as attention turned from high-profile frontier models to the rapid development of open-source alternatives. While major players like Anthropic and OpenAI captured headlines, many developers were quietly advancing their own projects, demonstrating that innovation isn't solely reliant on industry giants. Chinese open-weight models have made a notable impact, accounting for 41% of downloads on Hugging Face this spring, surpassing their U.S. counterparts. On platforms like OpenRouter, six of the top-performing models come from Chinese companies such as Tencent, Xiaomi, and Z.ai, with Anthropic's Claude Opus 4.7 lagging behind in seventh place. Data from Vercel indicates that open-source models are taking a significant share of AI application infrastructure, handling nearly a third of AI requests in June. However, this data may not fully represent the entire AI ecosystem, particularly the usage by major labs, which likely dominates the activity of OpenAI and Anthropic. This raises an essential question: How relevant are frontier models if most production AI is powered by more affordable, customizable options? Clem Delangue, CEO of Hugging Face, suggests that the future may see frontier models relegated to experimental uses or high-value tasks, while the majority of production workloads could be managed by private or open-source models. Hugging Face, known for its platform supporting open models, has seen a shift in its community's preferences. More companies are opting to own their AI models rather than rely on external providers, particularly as they face the high costs of scaling closed frontier models. Delangue emphasizes that enterprises should avoid depending on a single API provider, advocating for greater control and visibility over their core capabilities. The platform has grown significantly, with a new repository created every seven seconds, currently hosting nearly three million public models and one million datasets. Delangue notes that this trend reflects a departure from the notion of a singular dominant model, with many companies adapting various models to suit their specific needs. The rise of open models is further fueled by the continuous advancements from Chinese AI labs, which frequently release powerful open-weight models that are both cheaper and easier to customize compared to closed alternatives. For instance, Z.ai's recent model, GLM-5.2, shows strong performance in coding tasks and security vulnerability identification against Anthropic's offerings. Concerns about relying on a single model provider have been echoed by industry leaders like Microsoft CEO Satya Nadella. He emphasizes the importance of data control for enterprises using AI, advocating for a more distributed approach to learning infrastructure. This growing preference for open models has also sparked debates about their accessibility. Anthropic's CEO Dario Amodei has raised alarms about the potential dangers of widely available powerful models, fearing misuse by malicious actors. However, Delangue argues that the real risk lies in the concentration of power. He believes that enhancing transparency and leveling the playing field will ultimately make AI safer. In his view, restricting powerful models to a select few companies does not mitigate risks associated with advanced AI; rather, it may exacerbate them by creating a power imbalance. Delangue asserts that open-source models allow for better defense against known cybersecurity vulnerabilities, promoting a more robust and balanced AI ecosystem.