Open-source AI as geopolitical weapon: the case for and against American open models
May 14, 2025
Key Points
- Meta faces a shareholder math problem: $40 billion annual AI capex on open-source models that generate zero proprietary data and can't justify returns beyond a certain capability threshold.
- China deploys DeepSeek and Manus as free geopolitical weapons, gravitationally pulling mid-tier countries toward its economy while the U.S. risks either handing rivals a direct capability copy or ceding the free tier globally.
- The historical pattern of open-source monetization—Red Hat, Elasticsearch, Databricks—suggests free distributions eventually shift to paid tiers, making today's 'free' AI models a temporary competitive illusion.
Summary
Open-source AI divides into two economic questions and one geopolitical question. The economic logic may not hold, but the geopolitical logic is gaining weight.
Meta's Llama models work fine for enterprise use cases, but Meta will eventually face shareholder pressure to justify its $40 billion annual AI capex. Once training costs climb past $10 billion, investors will demand clear ROI. Open-source models capture no feedback loop. When users run Llama, the data doesn't return to Meta for retraining, unlike ChatGPT's thumbs-up/thumbs-down signals. Over the next decade, model differentiation will come from proprietary data, not internet-scale training runs. Meta's Llama strategy makes sense today because it serves internal use cases such as recommendation algorithms, feed ranking, and image generation that don't require frontier-grade model quality. But as capex scales, Zuck will face an awkward math. Why spend $10 billion or more on open models that can't improve Meta's products beyond a certain point and produce zero proprietary data?
Historically, open-source software companies follow a predictable arc. Red Hat moved CentOS behind a subscription wall. Elasticsearch changed licensing after seeding competition. Databricks owns the IP that accelerates Apache Spark, forcing customers to license their tools. The pattern holds: free distributions eventually monetize. For developers choosing open-source models today, "free" is an illusion. Inference costs, either paying a middleman for GPU management or absorbing direct GPU depreciation, often exceed API call pricing. Paid versions are generally better than their open equivalents. Llama 3 reached GPT-4 class performance a year after GPT-4 shipped. That gap will likely persist because the best engineers cluster around paying customers, not loss-leading distributions.
The geopolitical layer inverts the picture. China is distributing DeepSeek and Manus as part of an "AI belt and road initiative." These models are open-source not from conviction but as a distribution weapon. They're already deployed across Europe. A mid-tier country facing a binary choice between paying OpenAI or using a free Chinese model will gravitationally slide toward China's economy. A strong American open-source AI ecosystem could counterweight that pull.
The tension is sharp. If the U.S. open-sources its best AI and truly leads in capability, it hands near-peer rivals a direct copy. If it doesn't, Chinese competitors occupy the free tier globally and win hearts through cost. The question becomes whether a one-generation capability gap between deploying GPT-6 versus a Chinese open-source equivalent of GPT-5 matters enough to justify the geopolitical trade-off. That is a call for policymakers, not market forces.