This article analyzes the structural reasons behind the US lag in open-source AI models. High training costs, rapid model depreciation, and closed-source business models make it difficult for US AI companies to open-source, and Meta's shift from open to closed source further confirms this trend. Meanwhile, Chinese open-source models like Kimi K3 and Qwen are rising globally.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Money and talent are abundant, yet the US is collectively absent from open-source AI. This is no accident—it stems from deep contradictions in training costs, business models, and industrial structure.
Training a frontier model costs hundreds of millions of dollars, and its shelf life is only about 12 to 24 months. Once a new generation arrives, the previous one depreciates. This means the investment must be recouped within one to two years through API pricing, or it becomes a sunk cost. Under this pressure, closed-source is not a choice but a survival logic.
OpenAI and Anthropic's business models rely on per-token billing and per-call charges; open source would directly undermine this premise. Moreover, open-source AI is not Linux: training frontier models requires centrally orchestrating hundreds of thousands of GPUs for months—something the community cannot crowdfund.
Economist Ara Kharazian of Ramp notes that the rise of Chinese open-source models proves there is huge market demand that US model companies are currently not meeting.
Meta was once the sole US champion of open-source AI, releasing Llama 2 and Llama 3 in succession. Its logic was competitive: if open-source models were good enough, they could erode the moats of OpenAI and Google, while Meta profits from advertising and doesn't need direct model revenue.
However, Llama 4's internal performance fell short of expectations, lagging behind GPT and Claude. Meta subsequently abandoned the open-source route and pivoted to a closed-source model project codenamed 'Avocado.' The deeper reason is that Meta never truly went all-in on 'frontier performance'; once the battlefield shifted to closed-source model capability competition, it found itself behind rivals who focused on only one thing.
Analysis suggests that when the world's most committed open-source AI company abandons this path, it validates the business model that OpenAI and Anthropic chose from the start.
At least four or five US companies have the technical capability to train frontier models, but they face structural predicaments. The industry-recognized efficient method of 'distillation' requires using outputs from stronger models, but OpenAI and Anthropic's terms of service explicitly prohibit using model outputs to train competing models.
Thompson points out that US open-weight model makers comply with frontier labs' terms of service, resulting in weaker models compared to competitors that do not comply. Additionally, talent and capital are siphoned by OpenAI, Anthropic, and Google, making it difficult for remaining startups to produce equivalent work with a tenth of the resources.
On July 15, 2026, Thinking Machines Lab, led by former OpenAI CTO Mira Murati, released Inkling, a 975-billion-parameter open-weight model under the Apache 2.0 license, but acknowledged its positioning as 'a good base for fine-tuning,' not a frontier model. This demonstrates how locked the open-source path has become.
The golden age of open-source software was built on zero marginal cost: once code is written, copying and distribution costs are nearly zero. But 'copying' an AI model is inference—each answer consumes real computing power and electricity, with rigid marginal costs.
On the training side, it's a 'start from scratch' process: each new generation requires hundreds of millions of dollars in new investment, with little reuse of previous training costs. Software-era open source was like a snowball; AI-era open source is like filling a bottomless pit.
Rory noted on the 20VC podcast that Microsoft could crush competitors with bundling strategies because marginal costs were nearly zero; but every time an AI company lowers prices, inference costs do not follow. Therefore, 'open-source AI' is an unnatural existence in an economic sense.
This article is primarily based on an analysis piece from GeekPark, which synthesizes podcasts, NPR reports, and industry analysis. However, some data (e.g., training costs, model shelf life) lack original sources and should be treated with caution. Meta's pivot to the closed-source project Avocado and Thinking Machines Lab's release of Inkling are publicly reported and have higher credibility.
The US cannot produce top-tier open-source models because of a deep contradiction between AI's economic nature and the open-source model: high and rigid training and inference costs make closed-source the only commercially viable path. The rise of Chinese open-source models fills the market gap left by US companies constrained by their business models.
Primary report
Primary source