Updated on October 18, 2026.
AI vendors typically offer two billing models: a shared pool where you pay a flat fee that covers tokens, or a BYOK mode where you provide your own API key and pay your provider directly.
In standard SaaS, the shared pool is convenient. In regulated environments, it is untenable.
The problem with the shared pool
When your vendor makes your OpenAI / Anthropic / Mistral key available to all its customers via a single master key:
- You lose the accounting detail. It becomes impossible to prove to your auditor exactly how many tokens you consumed at the provider — you receive an aggregated invoice.
- You lose the contractual chain. Your data flows through the vendor's account. The data processing agreement (DPA) with OpenAI no longer covers you directly.
- You lose control of the model. The vendor can silently switch from model A to model B for cost reasons. Your benchmarks regress and you do not know why.
- You lose immediate revocation. In the event of an incident, you cannot revoke the key without asking the vendor first.
What BYOK guarantees
With BYOK:
- Your prompts and completions flow directly between your provider account and betool — not through a pooled third-party account.
- Your DPA with the provider remains the sole contractual chain. No opaque sub-processing.
- Your billing is transparent: the provider invoices you, you see every line item. betool only invoices you for orchestration.
- Your revocation is instant: regenerate the key at the provider, and the model becomes inaccessible within seconds.
The private-model option (self-hosted)
For organisations with the strictest requirements (banking, defence, healthcare), even BYOK is not enough: sending your prompts to OpenAI or Anthropic still constitutes a transfer to the United States.
The solution is a private model:
- Ollama or LM Studio on your GPU, for open-weight models (Llama, Qwen, Mistral, DeepSeek…).
- vLLM, or any server exposing the OpenAI API, on a GPU cluster for high-throughput production.
Your prompts never leave your perimeter. Latency is under your control. Compliance is total.
What it costs
The myth is that self-hosting an LLM costs a fortune. In reality, for current-generation open-weight models:
- A dual A100 80 GB GPU server can serve a model of around 70 billion parameters in production.
- At moderate usage (a few thousand exchanges per day), a single A6000 GPU is sufficient to serve a 32B model with sub-second latency.
- Monthly amortisation cost: ~$2,000 to $5,000 depending on your procurement strategy (rental vs. purchase).
Compared to an equivalent OpenAI spend over 12 months, the investment pays back in a matter of months at serious volumes — and the ROI flips entirely if you have sovereignty constraints that make OpenAI unacceptable.
Our choice
betool has been BYOK-native since day one: every organisation can connect its own keys (OpenAI, Anthropic, Google) or its private models (Ollama, LM Studio, OpenAI-compatible endpoint), and resolution always favours its own accounts.
To get started quickly, the platform also offers managed models: you use them without a key, and every token is debited in credits, model by model, on your betool invoice. It is an explicit choice, never a silent switch.
In regulated environments, our recommendation remains BYOK:
- The shared pool creates compliance debt that your auditors will ask you to justify.
- The asymmetry between "small customers" and "large customers" on sovereignty is not acceptable.
- The experience we want to build — "I can see exactly what I consume, where, and with whom" — is only complete with BYOK.
It adds one step to onboarding (15 minutes to retrieve a key from your provider). It is an investment that pays off over time.