Moat or Commons

The US AI industry has poured massive capital into the premise that frontier models would become a monopoly business. But that premise is collapsing as open-weight models narrow the performance gap to roughly 6-12 months and cut inference costs by 10-30x.

The issue isn't simply technological competition but capital structure. OpenAI, Anthropic, Google and Meta's model businesses, and major cloud partners are operating on the assumption of about $1 trillion in AI capex over the next four years, but if the assumption that monopoly pricing is possible breaks down, that investment logic no longer holds.

The open ecosystem has already pushed forward both models and infrastructure together. DeepSeek drew attention with a training cost of about $5.6 million, while the comparable US closed-source model was estimated at $500 million-$1 billion. Since then, Qwen, Kimi, GLM, MiniMax, Llama, Mistral combined with vLLM, llama.cpp, Ollama, LangChain, LlamaIndex have become de facto shared infrastructure that anyone can download and deploy.

If prices rise, users can easily walk away. Even with a $250 monthly subscription or an enterprise API, if the alternative is cheap or nearly free, the lock-in effect weakens sharply. For capital to restore scarcity, it must artificially manufacture a moat through regulatory blockades, vertical integration, and bundling, and the author believes US policy is likely to harden in that direction over the next 18-36 months.

• Under the banner of security, Chinese open-weight models get classified as a supply-chain risk, with restrictions expanding to federal agencies, contractors, and even critical infrastructure.
• Frontier labs rise from selling models to becoming operators, selling outcomes in areas like law, software, drug discovery, and financial analysis.
• A dual market solidifies, split between an expensive closed-source market at home and workaround routes abroad.

The auto industry is a mirror for this. In 1980, US automakers held about 80% of the US light-vehicle market, but by 2024 that had fallen to under 40%. Export restrictions, bailouts, and even a 100% tariff aimed at BYD failed to stop the trend, and the protective shield only bred weaker competitiveness rather than protecting producers.

If the same outcome repeats in AI, the cost falls on consumers, independent developers, startups, and in the long run, America's global influence. Protected margins may hold briefly, but product competitiveness and overseas markets will be lost even faster.