Hugging Face has quietly pulled the plug on its NVIDIA NIM API, the serverless inference service it launched with considerable fanfare for Enterprise Hub customers. Anyone hitting the docs today finds a blunt notice: the service was deprecated as of April 10th, 2025, and the company is steering users toward Inference Providers instead. No migration guide, no sunset explanation—just a redirect.
The service, announced as a way to run open models like Llama and Mistral on NVIDIA’s DGX Cloud infrastructure without the usual headache of provisioning GPUs, was pitched as pay-as-you-go simplicity. Enterprise organizations could generate a fine-grained token, grab a code snippet from a model card, and fire off OpenAI-compatible chat completions in minutes. Pricing was transparent enough: $8.25 per hour for H100 Tensor Core GPUs, which broke down to $0.0023 per second of compute time. The catch was always that usage fees landed on your Enterprise Hub bill, making it a convenience play for teams already in that ecosystem rather than a general-purpose tool.
What’s notable here is how fast the plug got pulled. The NIM API was positioned as a flagship collaboration between Hugging Face and NVIDIA, complementing the existing ‘Train on DGX Cloud’ service. For a joint product to go from launch to deprecation in under a year suggests either adoption didn’t meet expectations or the strategic priorities shifted—possibly both. Hugging Face’s push toward Inference Providers, which aggregates multiple model-serving backends, may have simply made the NVIDIA-specific integration redundant.
Developers who built workflows around the NIM endpoint now face the familiar grind of swapping base URLs and re-testing authentication. The OpenAI-compatible interface at least means the code changes are minimal, but the fine-grained token scoping that NIM required doesn’t map one-to-one onto every alternative. For teams that liked the predictability of per-second GPU billing, the replacement options may force a recalculation of costs—especially if they’re now paying per-token or per-request rates with different efficiency profiles. The bigger question is whether this signals a broader retreat from NVIDIA-specific partnerships on the Hub or just a one-off cleanup of an underperforming integration.
