Can you own it?

Data governance - retention, ZDR & training

Correction (2026-09-21): a rewritten Terms of Service (effective 2026-08-17) makes Zero Data Retention contractual and controlling - Sec 7(b): "Provider will not retain, store, or log any Customer Data submitted to or generated by the Services beyond the period strictly necessary to process and return the applicable request... ('Zero Data Retention')", and Sec 5(a) states this "controls over the Privacy Policy and any other Provider policy." The old open-ended debugging carve-out is replaced by bounded exceptions: support diagnostics deleted within 30 days, operational metadata (excludes content), and legal-compliance records. Training remains never by contract, properly scoped: "Provider will not use Customer Data to train, fine-tune, or otherwise improve any model, except as necessary to provide the Services."

The decisive caveat is scope: this covers DeepInfra's own open-model serving only. The same documents state that when you route requests to the closed models it re-sells - Anthropic Claude, Google - DeepInfra transfers your data to those endpoints to fulfil the request, and that vendor's storage and training policy applies (Google stores the output per its Privacy Notice; Anthropic per its Trust Centre). This is the key finding for anyone assuming a single blanket data guarantee: the guarantee is model-dependent, so segregate open-model and closed-model traffic when governance matters.

Data & IP ownership

For DeepInfra's own open-model serving, the ownership story is clean and grounded in the Terms of Service, which state you retain any intellectual property rights over your Submissions. Correction (2026-09-21): the rewritten Terms add Sec 17, a genuine bilateral confidentiality duty (three-year survival) - replacing the prior "will remain private" non-commitment. Combined with zero-retention and no training on submitted data, your inputs, outputs and any derived knowledge stay yours - customer_retains. But ownership is not uniform across the catalogue. The closed models DeepInfra re-sells - Anthropic Claude, Google - are transferred to those vendors and route your data under their storage and training terms rather than DeepInfra's, so whether your data stays proprietary depends on which model you call. The sub-processor list is not disclosed. Segregate open- and closed-model traffic and pin the terms per model.

Compliance & attestations

DeepInfra's data-privacy documentation states SOC 2 and ISO 27001 certification plus GDPR and HIPAA technical/organisational measures. The finer SOC 2 Type 1 (point-in-time, not Type II) designation, the sub-processor list and ISO 27701 status live on the Sprinto-powered trust centre, which is unreachable - documented but unverified. HIPAA is framed as measures, not a BAA. For healthcare or continuous-controls procurement, confirm the SOC 2 Type and the HIPAA posture on the trust centre directly.

Data residency & jurisdiction

DeepInfra operates US-based data centres only - no EU data-residency option was surfaced and region pinning is unconfirmed. For workloads with EU residency requirements this is a hard constraint that no plan upgrade removes. As a US-headquartered company it also carries standing US CLOUD Act exposure.

Security controls

Baseline controls are evidenced by ISO 27001 and SOC 2 certification, and the retrieved data-privacy docs confirm batch data is held encrypted on disk then deleted. The assessment is held to partial because the SOC 2 Type-1-vs-Type-II distinction rests on the unretrieved trust centre and no independent penetration test was surfaced.

Pricing & cost model

Cost is DeepInfra's standout: it is among the cheapest providers on the market. Pricing is mixed - pay-as-you-go per-token (indicatively Llama 3.1 8B ~$0.02/M, Llama 3.3 70B ~$0.35/M, gpt-oss-120B ~$0.08/M blended) plus dedicated GPU by the hour (A100 $0.89, H100 $1.79, H200 $2.19, B200 $2.79). Rates are approximate, so confirm live pricing, but the order of magnitude is the reason to consider DeepInfra.

Reliability posture

No incidents have been reported, but no public uptime SLA or failover documentation was independently verified. Reliability is therefore assessed conservatively: adequate on available evidence, but without an observable status/SLA record to lean on for availability-critical workloads.