DeepInfra
Inference API serving models from a catalog; input handling and underlying model rights are separate layers.
Score calculation
(8) ÷ 1 = 8.0
1 of six dimensions rated. Unrated dimensions and BDS status are excluded.
Full scoring method ↗Setup details
DeepInfra operates hosted inference; selected models may be third-party/open-weight.
Energy
↑DeepInfra describes operating GPU infrastructure across multiple U.S. data centers, but its public service materials do not allocate facility electricity, water, or emissions to a model, region, or request.
Reasoning & sources
DeepInfra’s infrastructure disclosures establish that it operates GPU capacity, not the energy or water consumed by a particular inference route. Ask for a workload- and region-linked footprint before comparing deployments; corporate scale or efficiency claims cannot substitute for that boundary.
Request-level electricity, emissions, and water use remain unknown.
Training rights
↑DeepInfra’s terms leave third-party model rights to each provider’s license and disclaim responsibility for third-party content; no catalog-wide training-corpus provenance or rights audit is disclosed. The host’s no-training rule applies to customer data, not upstream model corpora.
Reasoning & sources
DeepInfra’s terms make model-specific licensing the customer’s responsibility and do not provide an audited corpus-rights ledger, so the host cannot establish permissions for every model it serves. Before use, inspect the selected model’s license and provenance separately; the host’s no-training promise protects your inputs, not the model’s training set.
Underlying model corpus provenance, licenses and consent remain model-specific.
Privacy
↑DeepInfra’s terms, last modified October 15, 2026, promise zero data retention for customer data and prohibit using it to train or improve models, except as necessary to provide the service. They exclude customer-authorized support retention (deleted within 30 days after resolution), non-content metadata, legal/security/abuse records, and opted-in features such as batch, asynchronous outputs, third-party integrations, or paid cache; third-party-hosted models follow their own terms.
Reasoning & sources
The 8/10 rating rests on DeepInfra’s current terms: content is not retained after request processing or used to train models, with stated legal, security, support, and customer-selected-feature exceptions. Choose a DeepInfra-hosted model and review the selected feature’s terms; the contract explicitly sends third-party model data to the provider’s own policy.
Current terms expressly promise service-level ZDR and no customer-data training, with bounded, enumerated exceptions; the contract distinguishes third-party models and customer-selected storage features.
Confirm the chosen model is hosted by DeepInfra rather than a third party and avoid opting into batch, paid cache, or other persistence without reviewing that feature’s retention terms.
Openness
↑The catalog exposes a broad set of upstream open-weight models, but the service terms treat DeepInfra’s inference service as proprietary and preserve third-party model licensing separately. Access to open weights therefore does not establish that DeepInfra’s serving stack is inspectable or portable.
Reasoning & sources
The catalog confirms open-weight choices, while the service terms keep DeepInfra’s runtime under provider control and assign third-party model licensing separately. Leave the hosted-service openness score unrated until the serving stack’s inspectability and portability are documented; choose and self-host a specific model if those controls matter.
The reviewed sources do not document a reproducible server/runtime implementation or a path to move the managed inference stack; model availability and license differ by entry.
Worker treatment
↑DeepInfra’s public service and company materials describe inference infrastructure and customer-use terms, but do not report worker outcomes for data work, contractors, or data-center operations.
Reasoning & sources
The available company materials bound the business as model hosting on GPU infrastructure but do not identify its labor suppliers, pay, working conditions, or audit results. Treat worker protections as unverified and ask for current contractor and facility-vendor disclosures before relying on a labor claim.
Conditions for contractors, annotators, and model evaluators remain unknown.
Government & military
↑DeepInfra’s terms restrict certain weapons-system uses, but the reviewed company materials do not name a defense contract or establish an actual military deployment.
Reasoning & sources
The use policy is a customer-facing restriction, not evidence about procurement or enforcement in a specific deployment. Keep the contract relationship unrated unless the provider identifies a customer, award, model or endpoint scope, and applicable controls.
No inference about undisclosed customers or contracts is possible from silence.
BDS status
↑We did not find DeepInfra, Inc. among the named targets in the two official BNC campaign lists. This is a limited list check, not BDS approval.
Reasoning & sources
DeepInfra, Inc. is not a named target in the checked lists, but BDS uses selective campaigns, so absence cannot establish a clean record. Your selected model, cloud and billing route can introduce other companies into the purchase.
This review covers the named targets in the BNC consumer/pressure guide and No Tech campaign, not every BDS campaign, local group, investment or supplier. Lists and ownership can change. A company’s absence from these lists is not an endorsement or proof that it has no relevant relationships.