Inference, at the level of isolation you need
Summary
Every Infer service tier runs on the same foundation: zero data retention, encryption in transit, a gateway that admits or refuses rather than queues, and no path by which personnel can see customer content. What changes between tiers is tenancy: whether your requests share replicas with other customers, run on compute reserved for you alone, or run in an environment that is entirely yours. If your inference must not co-mingle with anyone else's, we can provide compute where it does not.
Shared capacity through public routing platforms
The default way to reach Infer. Requests arrive through our routing partner and are served on a shared pool of replicas behind Infer's gateway. Capacity is multi-tenant at the replica level: requests from different customers are interleaved on the same hardware, in the same way any public inference API operates.
- Data handling: zero retention applies in full. Content exists only in volatile memory for the life of the request and its session cache; nothing is stored or logged.
- Performance: served against the published speed floor, with admission control shedding load rather than degrading admitted streams.
- Best for: general workloads with no tenancy requirement, reached through the platform your application already uses.
- Regulated data: not eligible. Shared capacity is not offered for protected health information or other data requiring a dedicated agreement.
Reserved capacity
Replicas allocated to a single customer. Your requests are served on GPUs that carry no one else's traffic, while the gateway, admission control, and operational tooling remain Infer's shared control plane. Compute is not co-mingled; the platform around it is shared.
- Isolation: dedicated replicas and dedicated transient cache. No other customer's content is ever resident on your hardware.
- Capacity: a reserved throughput envelope sized with you, so your headroom is yours and admission decisions are made against your reservation, not the public pool.
- Access: through our routing partner, or directly against Infer's gateway with credentials issued to you.
- Regulated data: eligible for the agreements regulated workloads require, subject to the program status published on our HIPAA and compliance pages.
Private tenant
A separate installation of the Infer stack for one customer: dedicated gateway, dedicated replicas, and an isolated network and cloud environment that no other tenant shares. Nothing about the deployment, including its control plane, is common with the public service. This is the tier designed for regulated workloads.
- Isolation: separate project, network, credentials, and logging pipeline. Your environment is administratively and cryptographically distinct from every other deployment we operate.
- Placement: region and infrastructure provider selected with you from providers that have completed our subprocessor compliance evaluation, so inherited certifications and agreement coverage are known before deployment.
- Connectivity: private connectivity and IP allowlisting to your systems; public exposure is optional, not assumed.
- Regulated data: the intended home for protected health information and similar data, under the applicable executed agreements.
Customer-hosted deployment
Infer's gateway and serving stack installed in your own cloud account or data center. Content never leaves an environment you control; Infer provides the software, the deployment, and operational support, with access to your environment scoped and revocable by you. An air-gapped variant, with no outbound connectivity from the serving environment, is available by engagement.
- Data boundary: your perimeter. Infer's zero-retention design still applies inside it; the difference is that no request ever crosses out of your environment.
- Operations: operated by Infer under your access controls, or handed over to your team with runbooks.
- Regulated data: governed by your own program, with Infer's controls documented for your assessors.
Model choice
Shared capacity serves the models we list publicly. On reserved, private, and customer-hosted tiers you choose the open-weight model, including your own fine-tuned weights, served under your license and disclosed exactly as deployed: model, version, quantization, context length, and operating limits.
At a glance
| Tier | Compute | Control plane | Data boundary | Regulated data |
|---|---|---|---|---|
| Shared capacity | Multi-tenant replicas | Infer, shared | Infer environment | Not eligible |
| Reserved capacity | Single-tenant replicas | Infer, shared | Infer environment | Eligible under agreement |
| Private tenant | Single-tenant replicas | Dedicated to you | Isolated Infer-operated environment | Designed for it |
| Customer-hosted | Your hardware or cloud | In your environment | Your perimeter | Your program, our controls |
Zero data retention, encryption in transit, metadata-only logging, and the absence of any personnel path to content apply on every row. The security overview describes the controls; the compliance program describes the standards each tier is documented against.
Scoping an engagement
Shared capacity launches first, through our routing partner. Reserved, private, and customer-hosted deployments are scoped by engagement now: tell us the workload, the data classification, the regions in play, and the throughput you need, and we will propose the tier and the agreements that fit.
contact@infer.industries