DeepSeek maintains a 70% API margin while pricing tokens below third-party platforms, leveraging proprietary implementation details and operational know-how for greater inference efficiency on identical hardware. This efficiency advantage allows Chinese companies to release model weights without selling at a loss or failing to recoup training costs, despite lacking sufficient compute capacity for all inference demand.
Proprietary operational efficiency creates a durable cost advantage for frontier model providers, even with open-weight releases, challenging the long-term sustainability of pure open-source AI models.