DeepSeek just open-sourced its inference engine

Commenters treated the release less as a routine code drop than as a stress test for the economics and concentration of AI inference. The bullish case is that DeepSeek has published optimization work capable of pushing enormous token volume at striking reported revenue per accelerator hour, while supporting both NVIDIA and Huawei hardware. Several readers see that as a way for independent providers to inherit hard-won serving optimizations rather than each rebuilding a proprietary engine, potentially lowering the cost of deploying capable models and weakening NVIDIA’s software-and-hardware grip. One user’s switch from a premium Claude subscription to DeepSeek’s API was offered as practical evidence that the cheaper model is already fast and useful enough for everyday work. The excitement was tempered by implementation details and geopolitical suspicion. This is not a complete turnkey serving stack: key-value cache management, prefill/decode disaggregation, and replica load balancing remain outside the release, so running a reliable production service is still far beyond simply pointing a small GPU cluster at the repository. Readers also challenged the profit claims because the cited revenue figure lacks enough deployment detail, including a clear accelerator comparison, latency, concurrency, and utilization. Debate over motive stayed unresolved: some framed openness as a deliberate way to commoditize rivals’ infrastructure, while others saw ordinary commercial reputation-building and compute constraints, not centralized state strategy. Subscription pricing and predictable monthly spend also remain areas where US incumbents retain practical appeal.