DeepSeek open-sources an inference tool, turning up the pressure on OpenAI and Anthropic

DeepSeek has open-sourced an inference tool, a release that Gizmodo characterizes as dialing up the pressure on OpenAI and Anthropic. Inference is where closed labs concentrate their margin, and publishing software fit for production deployment puts the cost curve on the table.

From the public description, the tool targets the serving and deployment layer rather than new model weights. That distinction matters: for the past two years the open-source challenge landed mainly on models themselves, with DeepSeek's weights handing developers frontier-adjacent capability for free. Opening the inference stack touches the engineering that decides what those models cost to actually run.

The competitive logic is straightforward. OpenAI and Anthropic earn most of their revenue from token-priced inference, so a capable serving stack that anyone can download lowers the barrier to self-hosting and narrows the room for pricing power.

It also fits DeepSeek's pattern. The company built its reputation on low-cost open-weight models, and the debate around it keeps returning to the same question: as capability gaps shrink, whoever pushes unit inference costs lowest wins the accounts that are sensitive to price.

An open inference tool is not the same as an open service, though. The software addresses scheduling, throughput and production deployment; model quality, reliability, compliance and support remain the adopter's problem. Large enterprises get a cheaper way to experiment, while smaller teams still face operational complexity.

The longer-term effect is ecosystem-shaped. Once an inference framework is widely adopted, it becomes the place where hardware support and kernel optimization accumulate, and the tooling built around it can be harder to replace than the model itself. That is why open releases at the serving layer are usually read as a longer game than a single model launch.

Three things are worth watching: performance and stability data under real production load, how quickly major hardware platforms are supported, and whether OpenAI and Anthropic respond on pricing or self-hosted options. Gizmodo's headline calls it pressure; the size of that pressure depends on whether developers actually move workloads onto the tool.