A Roundup of ZDR Policies Across AI Platforms
tetsuoai · x · 2026-07-14
This post compiles the current ZDR (Zero Data Retention) / data retention status across various model/API platforms, emphasizing that "retention" isn't always bad: it aids in debugging agent toolchains, enables batching and caching, and supports the availability of some cutting-edge models.
Key Takeaways
- SpaceXAI: True ZDR; can be toggled via /privacy in the CLI and supports retroactive changes. Certain features (remote session lists, cloud agents, handoff) require data sync to be enabled.
- OpenAI / Anthropic / Azure / Google Vertex / Gemini API / Cohere / Groq / OpenRouter / Bedrock / Mistral / DeepSeek: These range from "partial support," "requires approval," "only available on certain endpoints," "configured per project/org," to "only structurally ZDR if self-hosted."
- The author views Bedrock as the "only truly enforced" solution, though it isn't enabled by default.
- DeepSeek is explicitly marked as having "no ZDR".
Main Points
- Vendors differ wildly in their definition of "zero retention." Often, it merely means disabling certain logs or training rather than truly not writing to disk.
- For actual use, pay attention to: store:false, ContentLogging:false, dataretentionmode:none, whether retroactive disabling is supported, and if capabilities/models are silently restricted.
More from Infra
- Encrypted GPT-2 speedrun challenges FHE inference: best is 3.031 s/token, goal sub-1s — tensorqt · 2026-10-27
- Open-source Latch compiles PyTorch into training megakernels, 2.75x faster on H100 — sguada · 2026-10-27
- NVIDIA claims HGX Rubin NVL8 delivers 10x tokens per megawatt vs B200 — thursdai_pod · 2026-10-27
- NVIDIA Claims HGX Rubin NVL8 Delivers 10x More Tokens per Megawatt Than B200 — thursdai_pod · 2026-10-27
- Together lets multiple LoRA post-training experiments share the same GPU capacity — togethercompute · 2026-10-27
- Commoditization dilemma: either human labor or GPUs become the commodity — _AustinCalvert_ · 2026-10-27