Enterprise LLM Compliance Compared: OpenAI Enterprise, Anthropic, AWS Bedrock, Azure OpenAI, Vertex AI, Cohere, Mistral, and Databricks — Real Certifications, Real Trade-offs (2026)

**OpenAI Enterprise** publishes its full compliance program on the OpenAI trust portal at https://trust.openai.com/. The site lists current SOC 2 Type II (covering ChatGPT Enterprise, ChatGPT Team, and the API platform), ISO 27001, ISO 27017, ISO 27018, ISO 27701, and CSA STAR Level 2 attestations. HIPAA BAAs are available for ChatGPT Enterprise and API customers on request per https://openai.com/business-associate-agreement/. FedRAMP Moderate is listed as in-progress; OpenAI for Government runs on Azure Government infrastructure inheriting Azure's FedRAMP High posture. The marketing copy to ignore: 'enterprise-grade security' — that is a category, not a certification. Ask for the report.

**Anthropic Enterprise** (the Claude for Business tier) publishes at https://trust.anthropic.com/. As of mid-2025 Anthropic became one of the first frontier-model vendors to achieve ISO 42001 — the new ISO standard for AI management systems — alongside SOC 2 Type II and ISO 27001. ISO 27017, 27018, and 27701 are listed as in-progress on the trust portal. HIPAA BAAs are available on the Enterprise tier. The default API and Claude for Business posture is no training on customer data — Anthropic has held this line longer and more publicly than any other frontier-model vendor, which matters in regulated procurement.

**AWS Bedrock** inherits the full AWS compliance umbrella at https://aws.amazon.com/compliance/. That covers SOC 1/2/3, ISO 27001/17/18/9001/42001, PCI DSS Level 1, HIPAA, FedRAMP High in GovCloud, DoD IL4/IL5, IRAP PROTECTED in AU regions, and BSI C5 in EU regions. Bedrock-specific data handling is documented at https://docs.aws.amazon.com/bedrock/latest/userguide/data-protection.html — model providers (Anthropic, Meta, Cohere, Mistral, AI21, Stability) do not see your prompts or outputs, and AWS does not use your data to train Bedrock-hosted models. The trade-off: each Bedrock-hosted model still inherits the underlying model provider's GPAI obligations under the EU AI Act.

**Azure OpenAI** runs at the intersection of Microsoft's compliance posture and OpenAI's model stack. The Service Trust Portal at https://servicetrust.microsoft.com/ lists FedRAMP High, DoD IL4/IL5/IL6 (Azure Government), ISO 27001/17/18/701/42001, HIPAA, BSI C5, and the EU Data Boundary commitment for European customers. Azure OpenAI's data, privacy, and security documentation at https://learn.microsoft.com/azure/ai-services/openai/how-to/data-privacy explicitly states that customer prompts and completions are not used to improve OpenAI or Microsoft models. Microsoft's abuse-monitoring feature stores prompts for up to 30 days; customers can apply to opt out via the abuse monitoring opt-out form.

**Google Vertex AI** publishes its certifications through the Google Cloud Compliance Reports Manager at https://cloud.google.com/security/compliance/compliance-reports-manager. Vertex AI is covered under SOC 1/2/3, ISO 27001/17/18/701/42001, HIPAA, PCI DSS, FedRAMP High via Assured Workloads, IRAP PROTECTED in AU, and BSI C5 in EU. Vertex's data governance posture at https://cloud.google.com/vertex-ai/generative-ai/docs/data-governance commits that customer data is not used to train Google's foundation models. Google was the first hyperscaler to publish an ISO 42001 attestation specific to its generative-AI services.

**Cohere Enterprise** publishes its compliance program at https://cohere.com/security. SOC 2 Type II is current; ISO 27001 is current per their security page with 27017/18/701 in progress. HIPAA BAAs are available on the Enterprise tier by contract. Cohere's compliance posture is competitive with the hyperscaler-native LLM offerings for mid-market RAG workloads but trails on FedRAMP, IRAP, and C5 — if those certifications are required, Cohere is best deployed through AWS Bedrock or Azure to inherit the host certifications. **Mistral** at https://mistral.ai/security publishes SOC 2 Type II and positions itself explicitly as the EU-sovereign option, with French and EU data residency by default. **Databricks AI** (including MosaicML) at https://www.databricks.com/trust covers SOC 2 Type II, ISO 27001/17/18, HIPAA, PCI DSS, FedRAMP Moderate, and IRAP — strong for enterprise data-platform workloads where the LLM lives next to the data.

SOC 2 Type II is the table-stakes certification — every vendor on this list has it for their enterprise tier. The differentiator is scope. **OpenAI**'s SOC 2 Type II at https://trust.openai.com/ covers ChatGPT Enterprise, ChatGPT Team, and the API platform. **Anthropic**'s covers Claude for Business, the API, and Claude.ai for Work. The hyperscaler SOC 2 reports (AWS at https://aws.amazon.com/compliance/soc/, Azure at https://servicetrust.microsoft.com/, GCP at https://cloud.google.com/security/compliance/soc-2) cover the entire cloud, with Bedrock, Azure OpenAI, and Vertex AI listed as in-scope services. Always pull the latest report under NDA and verify your specific service is named in the scope section.

ISO 27001 is the broader information security management system standard; everyone holds it. **ISO 27017** (cloud security controls) and **ISO 27018** (cloud PII) are where the gaps start showing. OpenAI, the three hyperscalers, and Databricks all hold 27017/18. Anthropic and Cohere have 27001 and list 27017/18 as in-progress. **ISO 27701** (privacy information management) is held by OpenAI, the hyperscalers, and Highspot-tier enterprise vendors; Anthropic and Cohere have not yet attested. For GDPR-heavy procurement, 27701 reduces controller-side review effort meaningfully.

**ISO 42001** is the new AI management system standard published in late 2023. As of June 2026, **Anthropic** was the first major frontier-model vendor to achieve it, with **Google** the first hyperscaler. **Microsoft** and **AWS** have announced 42001 programs that are in audit or early certified scope; verify the latest at https://servicetrust.microsoft.com/ and https://aws.amazon.com/compliance/iso-certified/. **OpenAI**, **Cohere**, **Mistral**, and **Databricks** have published roadmaps but had not yet completed 42001 attestation at the time of writing. If your AI governance program requires an ISO 42001 vendor attestation as a procurement gate, Anthropic and Google are the safest bets in 2026.

**HIPAA BAAs** are available from all six providers in the table, but the operational mechanics differ. **OpenAI** signs BAAs for ChatGPT Enterprise and API customers per https://openai.com/business-associate-agreement/. **Anthropic** signs BAAs for Claude for Business and API customers per https://trust.anthropic.com/. **AWS** and **Azure** sign master BAAs that cover Bedrock and Azure OpenAI as HIPAA-eligible services — see https://aws.amazon.com/compliance/hipaa-compliance/ and the Microsoft list at https://learn.microsoft.com/compliance/regulatory/offering-hipaa-hitech. **Google** covers Vertex AI under the Google Cloud BAA per https://cloud.google.com/security/compliance/hipaa. **Cohere** signs BAAs on the Enterprise tier by contract. Mistral and Databricks signing terms vary by deployment region.

**FedRAMP** is where the hyperscalers pull ahead. **AWS GovCloud** holds FedRAMP High for Bedrock — see https://aws.amazon.com/compliance/fedramp/. **Azure Government** holds FedRAMP High plus DoD IL4/IL5/IL6 for Azure OpenAI — see https://servicetrust.microsoft.com/. **Google Cloud Assured Workloads** holds FedRAMP High for Vertex AI — see https://cloud.google.com/security/compliance/fedramp. **OpenAI** lists FedRAMP Moderate as in-progress on its trust portal; OpenAI for Government runs on Azure Government, inheriting that posture. **Anthropic** Claude models are deployable through AWS GovCloud and Azure Government, inheriting those FedRAMP High authorizations. **Cohere**, **Mistral**, and **Databricks** are not FedRAMP-authorized standalone as of June 2026.

**IRAP** (the Australian Signals Directorate's Information Security Registered Assessors Program) and **BSI C5** (Germany's Cloud Computing Compliance Criteria Catalogue) are the regional certifications that matter for AU and DE buyers respectively. The three hyperscalers hold both at PROTECTED level. **Anthropic**'s Claude models inherit IRAP and C5 when deployed via Bedrock (AU regions) or Azure (EU regions). **OpenAI**, **Cohere**, **Mistral**, and **Databricks** require careful sourcing for AU and DE workloads — the certifications either flow from the hosting cloud or require custom architecture decisions. If you have a hard AU PROTECTED or DE BSI C5 requirement, deploy through AWS Bedrock or Azure as the default path.

Data residency is the first question every EU and APAC DPO asks. **OpenAI Enterprise** offers data residency in the US, Europe (Ireland), Japan, Singapore, India, South Korea, and Canada per https://help.openai.com/en/articles/9831260 as of mid-2025, with new regions added periodically. **Anthropic** offers US and EU residency natively, with additional regional reach via AWS Bedrock and Google Vertex AI deployments per https://trust.anthropic.com/. The three hyperscalers offer global region coverage by default — pick your region in the console and the data stays there subject to the documented support and abuse-monitoring exceptions. **Cohere** offers US, EU, and Japan; **Mistral** is EU-first with US as a secondary region; **Databricks** matches AWS / Azure / GCP region availability.

**Zero data retention (ZDR)** has become the contractual term that separates serious enterprise vendors from the rest. **OpenAI** offers ZDR on the Enterprise tier and the API by contract — see https://platform.openai.com/docs/models/how-we-use-your-data. Default API retention is 30 days for abuse monitoring; ZDR removes that. **Anthropic** has a default no-training stance and offers contractual ZDR via Claude for Business and API enterprise agreements per https://www.anthropic.com/legal/commercial-terms. **Bedrock**, **Azure OpenAI**, **Vertex AI**, **Cohere**, **Mistral**, and **Databricks** all default to no training on customer data; ZDR-equivalent controls are configurable per service.

Microsoft's **Azure OpenAI** abuse monitoring is a specific case worth understanding. By default, Azure stores prompts and completions for up to 30 days to support abuse monitoring per https://learn.microsoft.com/azure/ai-services/openai/concepts/abuse-monitoring. Customers handling sensitive data can apply for an exception via the abuse monitoring opt-out form. The opt-out is granted based on a documented use case — typical approvals include healthcare, financial services, and government workloads. If your DPO needs zero prompt persistence, file the form before you sign the contract, not after.

**Sub-processor lists** are public for every vendor on this list. OpenAI publishes its list at https://openai.com/policies/subprocessors/. Anthropic publishes at https://www.anthropic.com/legal/subprocessors. The hyperscalers publish sub-processor lists per service in their trust portals. The mechanics of opt-out differ materially: with **OpenAI** and **Anthropic**, you generally accept the sub-processor list as published — granular per-sub-processor opt-out is limited. With **AWS Bedrock**, **Azure OpenAI**, and **Vertex AI**, you control sub-processors implicitly by selecting which regions and services you deploy — there is no shared sub-processor pool you cannot opt out of. This is a real advantage for highly regulated buyers.

**EU Data Boundary** is Microsoft's commitment that EU customer data stays in EU data centers, including telemetry and support data, documented at https://learn.microsoft.com/privacy/eudb/eu-data-boundary-learn. Azure OpenAI is in scope for the EU Data Boundary as of 2024. Google offers similar Sovereign Controls for EU per https://cloud.google.com/sovereign-controls-eu. AWS offers the AWS European Sovereign Cloud, with general availability rolling through 2025 and 2026 — confirm current status at https://aws.amazon.com/compliance/european-sovereign-cloud/. If your contract requires EU-only data handling including support metadata, the three hyperscalers are the default path; smaller LLM vendors typically cannot match this depth.

The practical procurement advice across all eight vendors is the same: get the data residency commitment in the master services agreement, not the marketing page. Confirm specifically: (1) where prompts and completions are processed at inference, (2) where they are stored for retention or abuse monitoring, (3) where support and telemetry data lives, (4) how to opt out of any abuse monitoring or sampling, (5) the full sub-processor list with effective dates. Get the answers in writing before legal counsel reviews — vendor sales engineers will hedge in conversation; the DPA either commits or it does not.

The EU AI Act's General-Purpose AI (GPAI) obligations came into force on 2 August 2025 for new models and become enforceable for previously placed models on 2 August 2027. The Act creates two tiers: standard GPAI obligations (documentation, transparency summary, copyright policy, downstream-deployer support) and GPAI with systemic risk obligations (additional model evaluation, adversarial testing, incident reporting, cybersecurity). The systemic-risk threshold is currently set at models trained using more than 10^25 floating-point operations — which captures the largest frontier models from OpenAI, Anthropic, Google, Meta, and Mistral. The EU AI Office maintains the current designations; verify at https://digital-strategy.ec.europa.eu/en/policies/ai-office.

**OpenAI Enterprise** publishes its EU AI Act compliance summary on https://trust.openai.com/ including the model documentation and training data transparency summary required under Article 53. GPT-class frontier models are flagged as GPAI with systemic risk. Downstream deployers using OpenAI through ChatGPT Enterprise or the API can rely on OpenAI's documentation to support their own Article 13 transparency obligations to end users — but you still hold the deployer-side obligations for the specific application you build. See our EU AI Act compliance checklist for the deployer-side gating questions.

**Anthropic Enterprise** publishes its EU AI Act program at https://trust.anthropic.com/. Claude frontier models (Opus, Sonnet) are flagged as GPAI with systemic risk per the Article 51 threshold. Anthropic's Responsible Scaling Policy doubles as the model-evaluation and incident-reporting framework expected for systemic-risk models. The ISO 42001 attestation makes Anthropic one of the easier vendors to defend to an EU regulator during a conformity assessment — the AI management system standard is broadly aligned with the AI Act's quality-management requirements.

**AWS Bedrock**, **Azure OpenAI**, and **Google Vertex AI** are interesting cases under the Act. The hyperscalers are technically deployers when they offer hosted models; the underlying GPAI obligations rest with the model providers (Anthropic, Meta, Mistral, AI21, Cohere, Stability for Bedrock; OpenAI for Azure OpenAI; Google for Vertex AI Gemini). This means as a downstream enterprise buyer, you contract with the hyperscaler for the deployment, but the GPAI documentation flows from the model provider through the hyperscaler's marketplace. Confirm both layers when scoping conformity assessments.

**Cohere**, **Mistral**, and **Databricks** publish GPAI compliance plans on their security and trust pages. Cohere's Command models, Mistral's Large and Medium models, and Databricks' DBRX-class models are not currently designated systemic-risk under the 10^25 FLOP threshold, which reduces the obligations to the standard GPAI tier (documentation, transparency, copyright policy). If the EU AI Office revises the threshold downward, this could change — track designations at https://digital-strategy.ec.europa.eu/en/policies/ai-office.

Procurement practice for EU AI Act risk-tiered workloads: (1) classify your application under the Act's risk tiers (prohibited, high-risk, limited-risk, minimal-risk) before vendor selection — most enterprise LLM applications are limited-risk or minimal-risk, but RAG over HR data, credit decisions, or medical triage may be high-risk; (2) for high-risk applications, require the model provider's Article 53 documentation in the master services agreement; (3) for systemic-risk models, require notification rights if the EU AI Office issues incident-reporting findings; (4) maintain your own deployer-side documentation including the Article 26 fundamental rights impact assessment for high-risk uses. Vendors will not own your deployer-side obligations — that is your job.

Vendor compliance pages are sales documents. The real procurement work is forcing each vendor to commit to specific clauses in your master services agreement and data processing addendum. The minimum ask list for any enterprise LLM contract in 2026: (1) latest SOC 2 Type II report under NDA, (2) ISO 27001 certificate and statement of applicability, (3) signed DPA with EU SCCs and UK addendum, (4) HIPAA BAA if you handle PHI, (5) data residency commitment naming specific regions, (6) ZDR clause specifying retention period and abuse-monitoring opt-out, (7) sub-processor list with notification rights, (8) audit log retention commitment, (9) breach notification SLA in hours, (10) EU AI Act Article 53 documentation for any GPAI model in scope.

Question to ask **OpenAI Enterprise**: 'Confirm in writing that ChatGPT Enterprise prompts, completions, and uploaded files are not used to train OpenAI models, are retained for zero days under our ZDR addendum, and that EU data residency applies to prompt processing, completion processing, and audit log storage.' Verify the answer against https://trust.openai.com/. Get the regional data residency commitment specifically named in the order form, not just the master agreement.

Question to ask **Anthropic Enterprise**: 'Confirm that our Claude for Business and API usage is covered by the default no-training stance, that ZDR is contractually applied with zero-day retention, and that our deployment falls under the ISO 42001 attestation scope.' The 42001 question matters because some AI governance programs require attestation scope mapping. Verify at https://trust.anthropic.com/. Ask for the most recent ISO 42001 certificate with the statement of applicability.

Question to ask **AWS Bedrock** and **Azure OpenAI**: 'Confirm which Bedrock-hosted models / Azure OpenAI models are HIPAA-eligible under our master BAA, confirm the abuse-monitoring opt-out status for our workload, and confirm the specific regions in scope for our deployment.' For Azure, file the abuse-monitoring opt-out form before signing, not after. For Bedrock, confirm the per-model availability list at https://docs.aws.amazon.com/bedrock/latest/userguide/model-ids.html for your selected region.

Question to ask **Google Vertex AI**: 'Confirm the ISO 42001 attestation scope covers our Gemini deployment, confirm our region selection is within the EU Sovereign Controls boundary, and confirm Vertex AI Workbench data does not flow to training pipelines.' Verify at https://cloud.google.com/security/compliance/compliance-reports-manager. The Workbench question matters because notebooks can persist sensitive data if you do not configure CMEK and VPC Service Controls correctly.

Question to ask **Cohere Enterprise**, **Mistral**, and **Databricks**: 'Confirm SOC 2 Type II scope covers our deployment, confirm timeline for ISO 27017/18/701 completion, and confirm BAA and FedRAMP roadmap relevant to our use case.' For Cohere, verify at https://cohere.com/security. For Mistral, verify at https://mistral.ai/security — specifically confirm whether you are using La Plateforme (Mistral-hosted) or self-hosted weights, because the compliance posture differs materially. For Databricks, verify at https://www.databricks.com/trust — Databricks' compliance is broader than the LLM piece because the platform includes Unity Catalog data governance, which simplifies the overall posture.

Some regulated enterprises ask whether they can skip vendor compliance reviews entirely by self-hosting an open-weight model (Llama 3, Mistral Large weights, Mixtral, Falcon, Qwen) on infrastructure they already control. The compliance math is appealing: if the model runs in your VPC on your existing FedRAMP High AWS GovCloud account, the model provider's SOC 2 status becomes irrelevant — you inherit your own infrastructure compliance posture. There is no third-party data flow to disclose, no sub-processor list to opt out of, no abuse-monitoring opt-out to file.

The trade-off is real engineering cost. Running Llama 3 70B at production load requires roughly 4 to 8 H100 or H200 GPUs per replica per https://huggingface.co/meta-llama, plus vLLM or TensorRT-LLM serving infrastructure, plus a model-ops team that can patch, observe, and rotate. Per-token cost can land at 2 to 5 times the equivalent Bedrock or Vertex AI inference cost at moderate volume per the OpenAI API cost calculator and analogous benchmarks. The break-even point is typically 100M+ tokens per day with a stable workload and a real ML platform team.

Where self-hosting genuinely beats vendor compliance: (1) classified or air-gapped workloads where no external network connectivity is allowed, (2) workloads under jurisdictions where US-headquartered vendor SCCs are insufficient (some defense and intelligence agencies, some sovereign cloud requirements), (3) workloads where the EU AI Act systemic-risk designation creates downstream-deployer obligations you want to avoid by using a non-systemic-risk open model, (4) workloads where the prompts themselves are the trade secret and even a ZDR contract is insufficient governance.

The hybrid pattern that works in 2026: use vendor frontier models (Claude Opus, GPT-5, Gemini Ultra) for high-complexity reasoning where quality matters, and self-host smaller open models (Llama 3 70B, Mixtral 8x22B, Mistral Large open weights) for high-volume routine inference where the prompts contain regulated data. The vendor models handle the 10 percent of calls that need frontier quality; the self-hosted models handle the 90 percent of calls where Llama-class quality is sufficient and the data sensitivity makes the round-trip not worth the compliance overhead. **Databricks**' platform is built around this pattern.

Mistral occupies a unique position because it offers both: La Plateforme as a managed vendor service per https://mistral.ai/security, and open weights for self-hosting per the Apache 2.0 and Mistral Research License terms on https://huggingface.co/mistralai. If EU sovereignty is the dominant compliance requirement, the Mistral managed offering plus self-hosted Mixtral fallback is the cleanest single-vendor story. The same does not apply to OpenAI, Anthropic, or Google models, which are not available as open weights.

The bottom line on build-vs-buy for LLM compliance: vendor compliance is what you pay $0.003 to $0.015 per 1K input tokens for. The SOC 2 Type II, the ISO 27001, the BAA, the EU Data Boundary, the abuse-monitoring opt-out, the 30-day breach SLA — those are the deliverables. Self-hosting open weights gives you absolute control over the compliance posture in exchange for owning the operational burden. For most regulated enterprises in 2026, the right answer is vendor-hosted for the majority of workloads plus a self-hosted Llama or Mistral fallback for the highest-sensitivity data.

**OpenAI Enterprise** procurement typically runs 4 to 8 weeks from first contact to production access for a 200-seat ChatGPT Enterprise rollout, longer for an API-first deployment with custom DPA red-lines. Plan: 1 week for security review (pull SOC 2 and ISO from trust portal under NDA), 2 weeks for DPA negotiation, 1 to 2 weeks for SSO/SCIM provisioning, 1 week for user training. Verify the latest pricing and SLA at https://openai.com/enterprise. For API-only enterprise deployments without ChatGPT Enterprise, the timeline compresses to 2 to 4 weeks because there is no end-user provisioning.

**Anthropic Enterprise** (Claude for Business) procurement runs 4 to 6 weeks typically. The shorter cycle versus OpenAI reflects (1) the smaller seat counts most early Claude for Business deployments start at, and (2) Anthropic's faster legal turnaround on standard enterprise agreements as of 2025-2026. Verify current SLA and pricing at https://www.anthropic.com/enterprise. If your security review requires the ISO 42001 certificate, request it in the first week — Anthropic provides it under NDA.

**AWS Bedrock** deployment is the fastest if you are already an enterprise AWS customer — most teams have Bedrock available within their existing AWS account on day one. Time to production is dominated by model-specific procurement, especially Anthropic-on-Bedrock or Meta-on-Bedrock if your AWS Enterprise Discount Program does not already cover them. Plan 1 to 3 weeks for model availability sign-off plus standard application-level deployment. Verify at https://aws.amazon.com/bedrock/.

**Azure OpenAI** deployment is dominated by the access application process. Azure OpenAI requires a one-time approval for each subscription per https://learn.microsoft.com/azure/ai-services/openai/. Approval typically lands in 1 to 5 business days for enterprise Azure customers. After approval, file the abuse-monitoring opt-out if needed (1 to 3 weeks for processing), provision Entra ID role-based access controls, configure private endpoints, and you are in production. Plan 3 to 6 weeks end-to-end for a regulated workload.

**Google Vertex AI** deployment for Gemini models is similar to Bedrock — available to enterprise Google Cloud customers on day one. The compliance configuration (Assured Workloads for FedRAMP, Sovereign Controls for EU, VPC Service Controls for data exfiltration prevention) adds 2 to 4 weeks if you have not already deployed Assured Workloads in your org. Verify at https://cloud.google.com/security/compliance/compliance-reports-manager.

**Cohere**, **Mistral**, and **Databricks** timelines vary based on deployment mode. Cohere-on-Bedrock or Cohere-on-Sagemaker compresses to AWS timelines (1 to 3 weeks); Cohere direct enterprise is 4 to 8 weeks. Mistral La Plateforme direct is 2 to 4 weeks; Mistral self-hosted on your infrastructure is 6 to 12 weeks plus ML platform engineering work. Databricks AI on an existing Databricks workspace is 1 to 2 weeks; net-new Databricks deployment is 6 to 10 weeks. In all cases, the compliance review work runs in parallel to the technical deployment — do not let your DPO be the critical path by starting paperwork after technical pilot is complete.

If I were a US-headquartered enterprise SaaS company picking one frontier-model vendor tomorrow for an internal employee productivity rollout — knowledge management, drafting, summarization, code assistance — I would buy **OpenAI Enterprise**. The breadth of certifications, the maturity of the Enterprise admin tooling, the SCIM and SAML support, and the trust portal at https://trust.openai.com/ are the most polished in the category. The cost premium over Anthropic is small at typical seat counts. ChatGPT Enterprise is the easiest tool to defend to a CISO in 2026.

If I were a regulated US enterprise (healthcare, financial services, legal, public-sector adjacent) starting an API-first or RAG-heavy deployment, I would buy **Anthropic Enterprise** or **Claude on AWS Bedrock**. The ISO 42001 attestation, the default no-training stance, the Responsible Scaling Policy as a published systemic-risk evaluation framework, and the Claude model's strong refusal behavior on adversarial prompts make this the easiest stack to ship through a serious regulatory review. Verify at https://trust.anthropic.com/.

If I were a Microsoft-shop enterprise with EU operations, federal customers, or DoD-adjacent workloads, I would buy **Azure OpenAI**. Nothing else on this list matches FedRAMP High plus DoD IL5 plus EU Data Boundary in one stack. The trade-off is being locked to OpenAI's model release cadence on Azure, which can lag the OpenAI direct platform by 4 to 12 weeks for new model versions. For most regulated workloads that lag is acceptable. Verify at https://servicetrust.microsoft.com/.

If I were an AWS-native enterprise with mixed-model needs (some Anthropic, some Meta, some Cohere, some Mistral), I would buy **AWS Bedrock**. The single trust umbrella, the model-provider neutrality, the FedRAMP High via GovCloud, and the model-choice flexibility are unique to Bedrock. Verify at https://aws.amazon.com/bedrock/. For Anthropic-heavy workloads on AWS, the Claude-on-Bedrock path is usually preferable to Claude direct because you inherit the AWS compliance umbrella.

If I were an EU-headquartered enterprise with strict data sovereignty requirements, the order of preference in 2026 is **Azure OpenAI** under EU Data Boundary, **Google Vertex AI** under Sovereign Controls EU, **Mistral La Plateforme**, and **AWS Bedrock** under EU regions. Mistral is the only EU-headquartered option with a credible frontier model — if French or EU sovereignty is a hard board-level requirement, Mistral is the right answer. For most other EU enterprise buyers the hyperscaler EU offerings are easier to justify.

The one thing I would not do in 2026 is sign a single-vendor multi-year enterprise LLM contract without an exit clause. The category is moving fast — model quality, pricing, and compliance posture all shift quarterly. Negotiate annual renewal points, portable workload architectures (use an abstraction layer or an AI gateway), and the ability to mirror workloads to a second provider for resilience. The compliance posture you buy in June 2026 is not the compliance posture you will need in June 2028.