The AI arms race between the US and China has a new scoreboard, and it’s not looking as lopsided as Silicon Valley might prefer. Chinese open-source models from DeepSeek and Alibaba have rapidly closed the gap with leading American systems, matching frontier capabilities at a fraction of the cost.

CNBC’s coverage highlights a reality that’s been building for months: Chinese AI labs aren’t just catching up. They’re doing it cheaper, faster, and with open-source models that anyone can download and deploy.

The numbers tell a stark story

DeepSeek’s R1 model, released in January 2025, outperformed several prominent US models on third-party benchmarks. That includes Meta’s Llama 3.1, OpenAI’s GPT-4o, and Anthropic’s Claude Sonnet 3.5 in complex problem-solving tasks spanning reasoning, coding, and mathematics.

DeepSeek followed up with its V4 preview, which demonstrated strong agent-based performance while keeping inference costs lower than comparable US alternatives.

By March 2026, Alibaba’s Qwen series had surpassed Meta’s Llama to capture over 50% of global open-source model downloads.

On aggregator platforms like OpenRouter, Chinese open-source AI models accounted for approximately 30% of global AI model usage by early 2026. That figure was nearly negligible in late 2024.

Why open source changes the calculus

US AI companies like OpenAI and Anthropic have built their businesses around the idea that the best models live behind paywalls. That works great when your models are meaningfully better than anything else available. It works less great when a free alternative posts comparable benchmark scores.

What this means for investors

First, the obvious: pricing pressure on proprietary US AI systems could compress margins for companies that have commanded premium valuations based on assumptions of sustained competitive moats.

Second, the open-source surge has implications for the decentralized AI narrative in crypto. Projects building around open-source model hosting, inference marketplaces, and distributed compute stand to benefit if the industry tilts further toward open models.

Third, there’s a geopolitical dimension that investors can’t ignore. The US government has been tightening export controls on advanced AI chips to China, operating on the assumption that hardware restrictions would slow Chinese AI development. The performance of DeepSeek and Qwen suggests those restrictions have been less effective than hoped, or that Chinese labs have found ways to optimize around hardware constraints.

Meta has already committed to open-source with Llama, but even its lead in that category has been overtaken by Qwen’s download numbers.