fairlane.systems

DEEPSEEK · LLM PROVIDER

DeepSeek in Swiss practice: PRC provider, self-host option and revDSG reality

DeepSeek V3.x and R1 are extremely cheap and technically strong – but data flow goes to China. Not recommended for client data via API. Self-host via HF weights as alternative.

Researched & fact-checked by: · As of: 2026-05

What is DeepSeek?

DeepSeek is a Chinese LLM provider headquartered in Hangzhou. In May 2026 two model lines are active: the DeepSeek-V3 family (V3 and the V3.2 update, general chat, OpenAI-compatible, MoE architecture) and DeepSeek-R1 (reasoning model with chain-of-thought, much-discussed in January 2025).

Two traits made DeepSeek famous: first, the price. DeepSeek V3.2 costs around USD 0.28 per 1M input tokens and USD 0.42 per 1M output tokens (as of May 2026, api-docs.deepseek.com). DeepSeek-R1 as a reasoning model is at USD 0.55 / USD 2.19. With prompt caching input tokens drop to USD 0.03 – a factor of 10 below OpenAI. Second, the open-weight strategy: weights are on Hugging Face under a permissive licence. Whoever has the GPU hardware can run DeepSeek-V3-671B or R1 locally.

Legally DeepSeek is a special case. Servers are in the People's Republic of China; the company falls under China's 2017 National Intelligence Law – Chinese authorities can order data disclosure without notifying the affected user. The Italian DPA issued a ban within 72 hours in 2025; the EDPB set up an AI Enforcement Task Force; at least 13 EU jurisdictions are investigating. In March 2025, under Korean pressure, DeepSeek added an opt-out for training data and named an EU representative in its data privacy annex – that is not GDPR-compliant, just a compromise.

Why it matters

DeepSeek is the most important open-weight competitor to Llama. Three points make it interesting – and one makes it problematic.

First: DeepSeek-R1 is among the cheapest reasoning models on the market. Anyone needing chain-of-thought for legal analysis or math proofs pays 5-10x more at OpenAI o4 or Claude Sonnet reasoning. R1 sits on par with o1/o4 on MATH-500 and GSM8k and costs a fraction.

Second: V3.2 as a general model is top-3 on many code and reasoning benchmarks – alongside the current top GPT model and Claude Sonnet. Whoever generates code, delegates backend refactoring or automates schema conversions gets comparable quality at substantially lower cost.

Third: open weights. Whoever can afford the hardware (8x H200 or 16x H100 for V3 671B Q4) can run DeepSeek entirely in-house. That removes the PRC data flow. It is the only variant in which DeepSeek is acceptable for Swiss client data – and it is expensive.

The problem point: the DeepSeek API is effectively unusable for client data under revDSG/GDPR. Even with the training opt-out, inference logs remain in PRC jurisdiction. A fiduciary office bound by professional secrecy cannot use this without explicit, documented consent from every client – and realistically no client will sign that. For internal, non-confidential workloads (marketing copy, public FAQ, internal code review) the API makes sense. For client data: no.

How it works

API path: account at platform.deepseek.com, add a card, generate an API key. The call is identical to OpenAI (api_base = https://api.deepseek.com/v1). LiteLLM gateway recognises DeepSeek natively. Model names: deepseek-chat (= V3.2) and deepseek-reasoner (= R1). Model aliases can change with new releases – verify the current mapping on api-docs.deepseek.com before production use and plan migrations.

Pricing May 2026: V3.2 USD 0.28/0.42, R1 USD 0.55/2.19 per 1M tokens. Cache hit gives 90% discount on input (USD 0.03), company-wide since April 2026. Context window: V3.2 around 128-256k tokens, R1 128k.

API data retention: inference logs are stored; opt-out for training data is available (deactivate "Improve the model for everyone" in account settings). DeepSeek claims its privacy policy now meets GDPR-mapping standards – EU data protection authorities disagree. Servers are physically in PRC, data flow EU -> CN without adequate safeguards.

Self-host path: weights on Hugging Face (deepseek-ai/DeepSeek-V3, deepseek-ai/DeepSeek-R1, deepseek-ai/DeepSeek-V3.2 etc.). Licence: DeepSeek License Agreement, commercial use permitted. Hardware for V3 671B Q4: 8x H200 or 16x H100 (~USD 36/hour cloud rental at Lambda or vast.ai). Distillations (DeepSeek-R1-Distill-Qwen-14B, -32B) run on a single A100 80GB and reach 80% of R1 quality. For SME self-host the distillations are the realistic variant.

CIO decision: use DeepSeek or not?

  1. 01Data classification: does the workload contain client data, personal data or confidential business data? If yes: DeepSeek API is out. Check whether self-host or distillation is feasible.
  2. 02Run a Transfer Impact Assessment (TIA) for third-country transfer if API use is being considered. Realistic outcome in May 2026: risk not acceptable for client data.
  3. 03For self-host: model the hardware budget. 8x H200 for V3 671B = enterprise. 32B distillation on A100 80GB = realistic for an SME with high volume.
  4. 04Check the alternative model: Llama 3.3 70B delivers comparable quality without PRC questions and with better tooling.
  5. 05If the API is used anyway (internal tools, public data): set the training opt-out flag, document it, strictly scope the application.
  6. 06Configure LiteLLM routing: DeepSeek only for tagged "Tier-0 public" workloads, hard block on everything else.
  7. 07Quarterly review: monitor EU DPA status to see whether bans threaten the application.

When to use DeepSeek

DeepSeek API is the right choice when (a) the data is not confidential and (b) reasoning or code generation is needed at very low cost. In practice: public FAQ generation, marketing copy without client references, internal code reviews without PII, mathematical modelling with public data, benchmark tests before an architecture decision.

DeepSeek self-host is the right choice when (a) volume is high enough for owned GPUs to pay off and (b) PRC data flow must be avoided. Realistically that applies to enterprises or specialised SMEs with an 8x H200 / 16x H100 budget, not a 5-person fiduciary.

DeepSeek distillations (R1-Distill-Qwen-14B/32B) are the right choice when an SME wants to run a reasoning model on an A100 80GB. Quality lands at 70-80% of R1, with a hardware investment of around CHF 600-1000/month (own Hetzner GPU or cloud rental).

Vs alternatives: Anthropic Claude Sonnet has better legal reasoning at roughly 30x the cost. OpenAI o4 matches on coding benchmarks but at 10-20x. For applications where 95.0% answer quality is enough instead of 99.5%, DeepSeek has the best price-performance ratio. That is more applications than people first assume.

When not to use

DeepSeek API is the wrong choice whenever client data, personal data or confidential business data is in play. Clear no for: fiduciary mandates, legal mandates, AML-relevant facts, VAT data, payroll, KYC, every professional secrecy (Art. 321 SCC). Cannot be justified cleanly via Third-Country Transfer Impact Assessment (TIA) either – in May 2026 the PRC is not recognised as a country with adequate data protection, neither by the Federal Council nor by the European Commission. Standard contractual clauses are theoretically available but structurally undermined by the National Intelligence Law.

Further cases: anyone building an application marketed publicly in the EU should avoid DeepSeek API – Italy, France, Germany have active proceedings, a ban can come at any time. A Swiss service provider with DeepSeek in the backend could lose EU customers.

DeepSeek self-host has no legal problems but the SME hardware reality: 8x H200 is still not a 5-person office investment in 2026. Distillations are a good middle ground, but then you are comparing with Llama 3.3 70B self-host – and Llama has the broader tooling stack.

For politically sensitive topics (China history, Taiwan, Tibet, Hong Kong) DeepSeek exhibits known censorship patterns. Anyone needing texts or analyses on such topics gets skewed output – DeepSeek is not neutral here.

Trade-offs

STRENGTHS

  • Aggressively cheap: V3.2 USD 0.28/0.42, R1 USD 0.55/2.19 per 1M tokens
  • Cache-hit discount 90%: input tokens from USD 0.03 per 1M
  • Open-weight via Hugging Face – full self-host possible
  • R1 reasoning at o1/o4 level for a fraction of the price
  • Code benchmarks top-3 alongside the current top GPT model and the current top Claude model

WEAKNESSES

  • PRC data flow: not recommended for client data under revDSG / Art. 321 SCC
  • National Intelligence Law: authority access without user notification possible
  • Active investigations by multiple EU DPAs, risk of ban
  • Political censorship on China-sensitive topics – bias documented
  • Self-host for full models (671B) requires enterprise hardware (8x H200)

FAQ

Can I use DeepSeek with client data if the client consents?

Theoretically yes, practically no. You would need explicit, written consent from the client covering the third-country transfer to the PRC, including the National Intelligence Law warning – per mandate, per data category, revocable at any time. Realistically no client signs that if the information is complete. And without complete information the consent is invalid (Art. 6(6) revFADP). Recommendation: do not attempt.

What does a self-hosted DeepSeek-R1 distillation cost?

DeepSeek-R1-Distill-Qwen-32B on an A100 80GB: at Hetzner GPU around CHF 600-700/month (GEX44), at a Swiss provider (Infomaniak Public Cloud, Exoscale GPU) about CHF 1200-1800/month. At 2-5M output tokens/month self-host beats the API. Below that threshold the API is cheaper – once the compliance question is settled.

Does DeepSeek have an EU representative?

Yes, since March 2025, named in the EU privacy annex of the policy. That formally satisfies Art. 27 GDPR but says nothing about the lawfulness of the transfer itself. The pending proceedings in IT/FR/DE/NL/IE/ES turn on exactly that question: is the transfer permissible? As of May 2026 the prevailing DPA answer is: no, without additional safeguards.

Is DeepSeek-V3.2 censored?

On politically sensitive China topics yes, documented. Tiananmen, Taiwan status, Tibet, Hong Kong, Xinjiang – the models refuse or give CCP-aligned answers. For typical fiduciary and code workloads you do not notice it, but anyone producing market analyses, geopolitical texts or compliance analyses involving China will see bias. That is also relevant under revDSG: advisory work based on a skewed source is professionally problematic.

Related topics

TIA · COMPLIANCEThird-country transfer and Transfer Impact Assessment (TIA): Swiss data in US and PRC cloud LLMsrevDSG · COMPLIANCErevDSG / revFADP and AI: what the revised Swiss Data Protection Act means for LLM useSELF-HOSTED OLLAMA · LLM PROVIDERSelf-hosted Ollama as an LLM provider: when does it replace OpenAI, Anthropic or Gemini?SELF-HOSTED VS. CLOUD · AI CONCEPTSelf-hosted vs. cloud LLM: a decision framework for SMEs and fiduciariesROUTING · AI CONCEPTMulti-LLM routing: which model when, for how muchMETA LLAMA · LLM PROVIDERMeta Llama in Swiss practice: open-weight model, self-host or provider

Sources

  1. DeepSeek API – Pricing Page (USD) · 2026-05
  2. DeepSeek Privacy Policy (data residency PRC) · 2026-04
  3. EU Regulators Scrutinize DeepSeek for Data Privacy Violations (Usercentrics) · 2026-02
  4. IAPP – DeepSeek and the China Data Question (extraterritorial enforcement) · 2026-01
  5. DeepSeek-AI Model Weights on Hugging Face (V3, R1, distillations) · 2026-05

FITS YOUR STACK?

What this looks like in your business – a 30-minute intro call.

Book a call