A new data analysis from Our World in Data confirms what industry insiders have long suspected: the global AI landscape is essentially a two-country race, with American and Chinese companies training nearly all of the world's most widely-used models.

The Concentration Problem

Frontier model development requires massive computational resources, specialized hardware, and elite engineering talentβ€”inputs that are heavily concentrated in a handful of organizations. OpenAI, Google DeepMind, Anthropic, Meta AI, and xAI represent the American side of this duopoly, while Baidu ERNIE, DeepSeek, ByteDance, and Tencent's AI Lab anchor China's capabilities. The geographic concentration raises eyebrows among policy analysts who worry about single points of failure in critical infrastructure. When ChatGPT, Gemini, Claude, and Llama collectively handle billions of queries daily, the fact that they all trace back to just two nations becomes an economic and national security consideration.

Compute Is Still King

Training competitive foundation models demands thousands of high-end GPUs running for monthsβ€”something only organizations with access to significant capital or state backing can afford. The US export controls on advanced chips have attempted to constrain China's training capabilities, but the data suggests American firms maintain a structural advantage in compute availability that isn't going away overnight.

What This Means for Developers

For builders and enterprises downstream of these foundation models, this concentration creates both stability and dependency risks. APIs are reliable, benchmarks are standardized, and tooling is matureβ€”but when one provider has an outage or changes pricing, half the industry feels it. The EU, India, and Gulf states are pouring money into sovereign AI projects precisely because they recognize this imbalance.

Key Takeaways

  • US companies (OpenAI, Google, Anthropic, Meta) train most globally-used models
  • Chinese firms (DeepSeek, Baidu, ByteDance) represent the other dominant cluster
  • Compute access remains a fundamental barrier to entry for new players
  • Geopolitical tensions around chip exports haven't fundamentally shifted this dynamic

The Bottom Line

This data is a reminder that AI isn't as distributed or democratized as the marketing suggestsβ€”it's two superpowers running a high-stakes training race while everyone else builds on top of their foundation models.