Are people worried that open-source labs will start charging fees? At least for now, they haven’t. CSPs can directly deploy open-source models and sell Tokens—it’s totally feasible and already the mainstream approach. AWS Bedrock, Google Vertex AI Model Garden, and Azure AI Studio are all listing and commercially selling Token services for open-source models like Llama, DeepSeek, Qwen, and Mistral. They don’t need to train these models themselves; they just need to: 1) Download the open-source weights and use their own chips (TPU, Inferentia, Trainium, or custom chips); 2) Optimize serving; 3) Add routing, caching, security, compliance, and enterprise-grade features; 4) Sell Tokens externally. The core competitive edge of CSPs lies in their infrastructure + platform capabilities. This aligns perfectly with what we’ve been discussing: CSPs commoditize the “model layer” and capture value in hosting, optimization, routing, and governance. Will open-source labs start charging fees in the future? The risk exists, but the impact is limited. Possible scenarios: A particularly successful open-source model lab might change its license, launch a commercial version, or charge licensing fees for ultra-large-scale users. But CSPs can flexibly switch models. If one model starts charging, they can switch to another permissive one (there are plenty of options right now). Even if some models start charging, CSPs still have a huge advantage: they can use their own chips to drive inference costs extremely low, something the original labs selling APIs usually can’t achieve. Previously, deploying open-source models was a core part of reshaping CSPs’ business models, precisely because of the shift in cost structure.