Prefer to listen? Here's the narrated myth-vs-reality briefing.
Listen to Concentration of power
Plus unlocks audio streaming. Pro adds downloadable audio, video, certificates, and more.
The Myth
"Only big labs can build AI" was true for roughly 18 months in 2023. It's no longer accurate. Open-weight models (Meta's Llama, Mistral, DeepSeek, Qwen, Gemma, Moonshot's Kimi) are no longer a tier below the closed frontier — in July 2026 Moonshot published Kimi K3, the largest open-weight model ever released, and a Chinese open model was the one that repelled the OpenAI agent's intrusion at Hugging Face when US closed models balked. Hugging Face alone hosts 1,000,000+ open models, and the cost to train competitive models has dropped dramatically.
The Reality
Compute concentration IS real: training frontier models still requires capital and data center access most organizations lack. A handful of cloud providers (AWS, Azure, GCP) host the majority of production AI workloads. App-layer concentration around ChatGPT, Claude, and Gemini creates legitimate questions about who controls the defaults that hundreds of millions of users see. There is also a subtler risk worth naming: safety is a real argument that can also be a convenient one. As Washington weighs restrictions on Chinese open-weight models and the industry debates how to pace itself, the same reasoning that justifies caution can justify shutting out competitors — a tension OpenAI's Sam Altman named himself in July 2026, warning against anything that "feels like regulatory capture" or "collusion among the frontier labs." Watch who benefits from each proposed rule, not just what the rule promises.
The Positive Path
The open-weight movement, sovereign-AI initiatives (France's Mistral, UAE's Falcon, China's DeepSeek), and edge-AI hardware (Apple Silicon, NVIDIA Jetson) all push back against concentration. Learning to use both closed and open models — and understanding which to use when — gives you optionality the panic narrative says you don't have.