7 min. read As of: 22.04.2026

Anyone in DACH looking to deploy Claude, GPT, or Llama productively today has three options: the Anthropic API directly, AWS Bedrock via EU inference profiles, or self-hosted on your own hardware. On 2 August 2026, full enforcement of the EU AI Act takes effect. That turns the question “where does inference run?” into a compliance question. Locking in your architecture now saves you a costly re-platforming later.

Key Takeaways

• Deadline is set: From 2 August 2026, the EU AI Office has full enforcement authority over GPAI providers, including fines and model withdrawals (EU AI Act Implementation Timeline).
• Anthropic lacks an EU region: The direct Anthropic API still only offers “us” and “global” as inference geographies — no dedicated EU cluster.
• Bedrock is the shortest path to Claude with EU data residency: AWS has served Claude Opus 4.7 from Ireland and Stockholm since April 2026, with Frankfurt attached via a cross-region inference profile.
• Self-hosted is closing the gap: Llama 4, Mistral Small 4, and Qwen 3.6 now trail the closed-source flagships by just 3 to 5 percentage points on MMLU-Pro, while inference costs have dropped 40 to 60 percent.
• The architecture decision isn’t a matter of faith: Data classification, latency budget, and team skills beat any vendor recommendation.

RelatedOpus 4.7 vs GPT-5.4 in the EU Cloud Benchmark  /  Deploying Gemma 4 locally

The Situation in April 2026

What is AI inference? Inference is the productive operation of a trained model: text, an image, or a table goes in, an answer comes out. With large language models (LLMs), this happens on specialized GPU hardware that generates tokens sequentially. Anyone building AI features in the DACH region makes a critical decision at exactly this point — where these inference computations actually run. This is far from a minor detail, because personal data, business logic, and customer interactions are regularly processed here.

Three developments are squeezing DACH teams simultaneously. First: the EU AI Act has applied to newly placed GPAI models since August 2025, and from 2 August 2026 the AI Office gains full enforcement power and can impose penalties (see analysis for mid-market tech teams). Second: Anthropic’s Opus 4.7 is the most capable model on the market, but there is no dedicated EU region in the Direct API (Anthropic Privacy Center). Third: open-source models are now benchmark-level peers to GPT and Claude. Anyone who only needs text classification or retrieval-augmented generation (RAG) responses can run that within their own cloud.

The result: in DACH enterprise settings, the question “which model?” is asked first less and less often. The first question is “where does the inference run?” — and in 80 percent of cases, the model choice follows almost automatically from the answer.

Source: European Commission, AI Act Implementation Timeline

What this means for architecture decisions: each route carries a different compliance footprint, a different latency curve, and a different team overhead. The three paths below are not mutually exclusive — many teams end up running a combination. They are, however, the three clean starting points.

Path 1: Anthropic Direct API without EU Residency

Integrating directly against platform.claude.com is the fastest way to reach Claude Opus 4.7. No hyperscaler account coupling, no IAM role theater — the SDK call is written in four lines of Python or TypeScript. The trade-off: Anthropic’s Direct API currently offers only the geographies “us” and “global.” A dedicated “eu” inference region has not been announced.

For DACH environments, this means three things. Teams processing only publicly available data, marketing copy, or code generation generally get on fine. Anyone sending personal data in the sense of the GDPR through the API needs either a solid legal basis under Art. 44–49 GDPR plus the EU–US Data Privacy Framework, or a different route entirely. Teams that fall under the EU AI Act as deployers of high-risk systems should have an answer ready — one that does not amount to “we call the US API” — at the very latest now.

Typical use case: internal developer tooling, code review automation, content generation for marketing assets. That works with minimal friction. The journey becomes painful the moment customer data, HR data, or a regulated process enters the picture.

Path 2: AWS Bedrock via EU Inference Profiles

Bedrock is the most pragmatic way to run Claude in the EU in 2026. Claude Opus 4.7 was enabled for Ireland and Stockholm in mid-April 2026, with Paris and Frankfurt accessing it via cross-region inference profiles (AWS Weekly Roundup, April 20, 2026). For teams with an existing AWS footprint, this is a one-hour integration: enable model access, adjust the IAM policy, and fire an API call against bedrock-runtime using the EU inference profile.

The compliance gains are real: data in transit and the inference itself stay within AWS EU regions, the Data Processing Addendum is signed, and the audit trail is clean. Teams already running a AWS-centric policy framework will avoid the usual back-and-forth with legal entirely.

The trade-offs: Bedrock carries a markup over Anthropic’s direct pricing, which adds up noticeably at high volumes. New Claude versions also tend to land in US regions first, reaching EU regions a few weeks later. Teams already deeply invested in Azure or GCP will need to weigh the network hop to AWS. For Google Cloud teams, Vertex AI is the equivalent path – ten EU regions, same data residency logic.

In practice, a team with an existing AWS account structure gets started like this: request model access for Claude Opus 4.7 in the Bedrock console, attach an IAM policy granting bedrock:InvokeModel on the EU inference profile ARN, and pass the profile rather than a specific region ID in the client SDK. AWS then automatically routes to the nearest available EU region, guaranteeing that request and response data never leaves the geography. Cross-region logs land in CloudWatch, giving you a documented audit trail for compliance assessments.

Path 3: Self-Hosted Open-Source Inference

This is the path that almost nobody took seriously twelve months ago. By 2026, the calculation looks very different. Meta’s Llama 4, Mistral Small 4, Alibaba’s Qwen 3.6, and DeepSeek V4 trail GPT and Claude by only a few percentage points in reliable benchmarks. The gap is consistent enough that in many workloads nobody notices the difference. For classification, summarization, structured extraction, RAG retrieval, and tool use, open-source has arrived in production. For edge cases like long-context agent orchestration or highly creative writing tasks, Claude Opus and GPT still pull ahead.

The technical stack has matured: vLLM with PagedAttention as the inference engine, Hugging Face TGI or BentoML as alternatives, and Triton for multi-model serving. Over the course of 2025, vLLM established itself as the de facto standard for high-throughput scenarios, delivering 14 to 24 times the throughput of naive Transformers implementations depending on the workload.

The hardware question is no longer trivial in 2026, but it is tractable. A single H200 or two A100 80 GB cards are sufficient for a 70-billion-parameter model at 4-bit quantization; two L40S cards handle smaller variants. Organizations that want to avoid co-location can get preconfigured GPU instances from German IaaS providers. The cost break-even against Bedrock sits at roughly 150 to 250 million tokens per month for many workloads — below that threshold, Bedrock is generally cheaper and simpler. For a practical guide to model selection, see the CM comparison RAG vs. Fine-Tuning vs. Prompt Engineering.

Decision Matrix for DACH Teams

The choice doesn’t hinge on the model — it comes down to three questions. First: what data class flows through the inference? Public data and marketing assets can run on any route. Personal data, financial data, health data, or sensitive business data forces you toward Bedrock or Self-Hosted. Second: what latency budget does the use case allow? For chatbots requiring sub-second responses, Self-Hosted with an EU region is fastest; anyone needing streaming outputs will do well with Bedrock Claude. Third: what can your team actually deliver? A web team with solid AWS experience gets Bedrock running in a week. An MLOps team with GPU ops expertise builds a vLLM production environment in six to ten weeks.

If you have neither in-house, don’t start with Self-Hosted. That’s not an admission of weakness — it’s a sober assessment. A poorly operated LLM cluster creates more compliance risk than a clean Bedrock integration, because missing monitoring, unpatched CUDA stacks, and unsecured inference endpoints quickly become attack vectors. With Bedrock, AWS absorbs those operator obligations as part of the Shared Responsibility Model. Teams that still want to keep that path open should start with a smaller model in a controlled environment and build operational experience before moving to production. For the sovereignty debate in leadership forums, the DC article on local AI provides a useful framework for the conversation.

The window isn’t huge — but it’s not closed either. Starting in May gives you three months for a clean decision plus rollout. Waiting until June turns the summer into a compliance sprint. And anyone who still hasn’t answered the question in July will be explaining to an auditor in August why their inference is running on a US endpoint.

Conclusion

The three paths aren’t mutually exclusive. Many DACH teams will run Bedrock for sensitive workloads, Anthropic Direct for internal dev tools, and Self-Hosted for high-volume, low-sensitivity classification. The point isn’t “Route A or B” — it’s this: make a deliberate, documented decision for each workload. That’s also the audit standard the EU AI Act enforces from August 2026 onward. Teams that write the inventory list today will have an answer in August. Those who don’t will receive their answer from the outside. That’s rarely the cheaper option.

Frequently Asked Questions

Is Claude Opus 4.7 already available in Frankfurt?

Opus 4.7 does not launch directly in eu-central-1. Activation in April 2026 runs through Ireland and Stockholm, with Frankfurt gaining access via the EU Cross-Region Inference Profile. For most compliance requirements, that’s sufficient — inference is guaranteed to remain within EU regions.

When does self-hosted beat Bedrock on cost?

Rule of thumb: somewhere between 150 and 250 million tokens per month, depending on prompt length and output tokens. Below that threshold, Bedrock almost always wins on total cost of ownership — GPU operations and MLOps headcount are expensive. Above it, the math flips.

Is the EU-US Data Privacy Framework sufficient for the Anthropic Direct API?

For many use cases, yes — provided the legal basis under Art. 44–49 GDPR is properly documented and the provider is certified under the framework. For high-risk AI systems within the scope of the EU AI Act, the answer is thinner and requires additional safeguards.

Which open-source models hold up in EU production environments?

Llama 4 from Meta, Mistral Small 4, Qwen 3.6 from Alibaba, and DeepSeek V4 all trail closed-source flagships by only a few percentage points on MMLU-Pro and comparable benchmarks. For classification, RAG, and tool use, the gap is barely noticeable in practice — but for long-context agents, GPT and Claude pull further ahead.

How significant is the risk of violating the EU AI Act after 2 August 2026?

It depends on how your use case is classified. Organizations using GPAI models purely as deployers without building a high-risk system can get by with documentation and transparency measures. Those offering a high-risk AI system themselves will need risk management, logging, human oversight, and a conformity assessment. The AI Office’s enforcement powers — including fines — kick in from 2 August 2026.

More from the MBF Media Network

• EU AI Act from August: What Mid-Market Tech Teams Need to Clarify
• Chief AI Officer 2026: Real Role or the Next C-Level Title?
• On-Premise AI as a Security Strategy

Cover image source: Pexels / panumas nikhomkhai (px:17489157)