Nearly a trillion dollars in US AI investment rests on a bet that frontier models would command monopoly pricing. That bet is falling apart. Open-weight models from Chinese labs like DeepSeek, Qwen, Kimi, and GLM now sit six to twelve months behind the closed frontier on benchmarks while costing ten to thirty times less to run. The gap is closing, not widening. Shaun Warman, who wrote the analysis this piece draws from, puts it plainly: capability that a US closed lab could charge enterprise rates for in 2024 is now downloadable and deployable on rented hardware at single-digit cents on the dollar in 2026.
DeepSeek's origin story explains why this happened faster than anyone expected. The lab spun out of High-Flyer Capital Management, a Chinese quantitative hedge fund founded by Liang Wenfeng. Unlike US labs dependent on venture capital, DeepSeek was bankrolled by trading profits. By 2021 they'd already amassed roughly 10,000 Nvidia A100 GPUs. That quant DNA shows up in their architecture. They aggressively use Mixture of Experts to slash inference costs, treating model training as an optimization problem where efficiency beats brute force every time. The result is GPT-4 class performance at a fraction of the spend.
The models sit on infrastructure that's free and mostly Western-built. vLLM handles production throughput. llama.cpp runs models on laptops. Ollama wraps everything for non-technical users. LangChain and LlamaIndex provide orchestration that two years ago only existed inside OpenAI's product org. None of these tools belong to the closed labs. The infrastructure is geographically and economically agnostic. The weights are not.
Warman predicts three responses. Regulatory enclosure targeting Chinese open weights under national security pretexts. Vertical integration where frontier labs absorb their own customers by becoming operators. And a split market where US users pay premiums while the rest of the world routes around American rails. The Hacker News discussion adds a wrinkle worth watching. Several commenters argue that value is shifting to orchestration and multi-agent systems, where cheaper open models can compete through better coordination. If they're right, the moat was in the plumbing all along.