HIPAA-Compliant Inference for Claude, OpenAI, DeepSeek and GLM: Bedrock vs Azure vs Fireworks vs Anthropic Direct vs Concentrate

If you need HIPAA-compliant inference for more than one model family, such as Claude, OpenAI's GPT models, DeepSeek and GLM, you have five realistic options: Amazon Bedrock, Microsoft Azure, Fireworks AI, a direct BAA with each model maker, or Concentrate. Each can be used with protected health information (PHI) under a Business Associate Agreement (BAA). They differ in which models one BAA covers, how failover works, and how many contracts you end up managing.

What does HIPAA-compliant inference require?

HIPAA-compliant inference means every provider that receives PHI has signed a BAA and the deployment follows it. Under HHS guidance, a cloud or AI service that handles ePHI for you is a business associate. Coverage is per service and per model route, so check the exact model and endpoint, including any fallback your app uses during an outage.

Comparison: Bedrock vs Azure vs Fireworks vs direct vs Concentrate

Option	Who signs the BAA	Claude	OpenAI GPT	DeepSeek and GLM	Failover across clouds	Fees
Amazon Bedrock	AWS	Yes	Check catalog	Some open models; check each model's eligibility	No, AWS only	AWS pricing
Microsoft Azure	Microsoft	Yes, in Microsoft Foundry	Yes	DeepSeek yes; check others	No, Azure only	Azure pricing
Fireworks AI	Fireworks	No	Open-weight gpt-oss only	Yes	No, Fireworks only	Fireworks pricing

Competitor details as published on their own pages, November 2026. Always confirm coverage for your exact model and endpoint.

Amazon Bedrock: strong if you are all-in on AWS

Amazon Bedrock is on the AWS HIPAA Eligible Services list, covered by the AWS BAA, which you accept in AWS Artifact. Claude is available on Bedrock, along with a catalog of open models. The catch is scope: not every model or deployment type on Bedrock is necessarily covered, so check model by model, and everything stays inside AWS. If your app also needs GPT on Azure or a model Bedrock doesn't serve, you need another BAA and your own failover logic.

Best for: teams already running PHI workloads on AWS who mainly need Claude.

Microsoft Azure: the broadest single cloud

Microsoft includes the HIPAA BAA in its Product Terms for in-scope services; see Microsoft's HIPAA offering page. Azure serves OpenAI's GPT models, Claude through Microsoft Foundry (Anthropic says its BAA covers Claude in Microsoft Foundry), and open models such as DeepSeek. Models sold through Foundry by third parties can have different terms, so confirm each one. Like Bedrock, it is one cloud: an Azure outage has no covered place to fail over to unless you add another provider and another BAA.

Best for: Microsoft shops that want GPT and Claude in one cloud.

Fireworks AI: fast open models

Fireworks states HIPAA support and zero data retention for open models on its data security page. It serves open models like DeepSeek, GLM, Kimi and Qwen, but not Claude or OpenAI's GPT models, so a team using frontier models still needs separate BAAs for those.

Best for: open-model workloads where you already have frontier-model BAAs elsewhere.

Anthropic or OpenAI direct: one model family per contract

Anthropic's BAA covers its first-party API, but your Primary Owner signs it and then contacts Anthropic sales to turn it on, and not every API feature is covered. OpenAI also offers BAAs for its API. Each direct BAA covers only that maker's models, so Claude plus GPT plus an open model means three contracts, three integrations and failover you build yourself.

Best for: single-model products with no plans to add other providers.

Concentrate: one BAA across Claude, GPT, DeepSeek, GLM and more

Concentrate signs one BAA and serves the inference itself across 7 BAA-eligible providers: Anthropic, OpenAI, Azure, Google Vertex AI, Fireworks AI, Azure Fireworks and Baseten. As of November 2026 that covers 70 models, including Claude Sonnet 5.5, GPT-6.1 Sol, DeepSeek V4.1 Flash, GLM-5.3 and Gemini 4 Argon.

On a HIPAA organization, zero data retention is locked on, request-content logging is locked off, and uncovered routes are blocked. If Anthropic is down, a Claude request can fail over to Claude on Azure; if no covered route is left, the request is rejected rather than sent to an uncovered provider. Bring Your Own Key is off, because the BAAs do not cover your own provider accounts.

Pricing: BAA access is $499/month, and inference is billed at provider rates with no per-request fees and no markup on tokens. Without a BAA there is no subscription fee.

Best for: teams that use more than one model family and want one BAA, one API and covered failover.

Which option should you choose?

If you only need Claude and already run PHI on AWS, Bedrock works.
If you are a Microsoft shop and need GPT and Claude, Azure works.
If you only need open models and already have frontier BAAs, Fireworks works.
If you need Claude, GPT, DeepSeek, GLM or Gemini together, or failover that stays covered across clouds, use Concentrate.

The full list of covered models and providers is in Model Fortress. Related reading: Bedrock vs direct BAAs vs a gateway BAA, Top 5 AI gateways for healthcare and GLM-5.3-Flash BAA.

FAQ: HIPAA-compliant inference

Is Amazon Bedrock HIPAA compliant?

Bedrock is HIPAA-eligible under the AWS BAA. Your deployment still has to follow the BAA, and you should confirm each model's eligibility.

Is Azure OpenAI HIPAA compliant?

Azure OpenAI is covered by Microsoft's HIPAA BAA in its Product Terms. GPT models are also BAA-eligible on Azure and OpenAI through Concentrate.

Can I use DeepSeek or GLM with PHI?

Yes, on BAA-eligible providers. On Concentrate, DeepSeek and GLM are covered on Fireworks AI, Azure Fireworks and Baseten, and some DeepSeek models on Azure.

Can one BAA cover Claude, GPT and open models?

On Concentrate, yes. One BAA covers every model on a BAA-eligible provider.

Does Concentrate sign a BAA?

Yes. BAA use is $499/month, with no per-request fees and no markup on tokens. Non-BAA use has no subscription fee. Start at concentrate.ai/compliance.