kvrun vs Together AI and Fireworks: UK jurisdiction compared

Together AI and Fireworks AI are two of the best open-model inference platforms available: fast, inexpensive and quick to adopt. Both are US companies. kvrun is a UK company (KVRUN LTD, registered in England and Wales) that runs open models on dedicated GPUs in London. For a UK buyer the useful comparison is less about speed and more about tenancy, retention and where the data legally goes. This page covers all three, and says when serverless is the right call.

This reflects each provider's public documentation at the time of writing. Their terms change, so confirm the current position in Together's and Fireworks' own documentation and in your contract.

Summary

Shared serverless vs dedicated tenancy

A serverless endpoint is a pool of GPUs that serves many customers at once. You send tokens, pay per token, and never think about capacity. It is very efficient, because the provider keeps the GPUs busy across everyone's traffic.

The trade-off is that your requests share a serving engine with other organisations. Providers isolate customers in software, and the leading ones take that seriously. But the KV cache, and especially prefix caching, is where shared engines have leaked in published research: timing differences from a shared cache can reveal what another user sent. Is prefix caching safe? goes through the papers.

A dedicated endpoint gives you GPUs that serve only your traffic. Both Together and Fireworks offer one, at a higher price than serverless. kvrun only offers dedicated deployments, and adds a governed KV cache on top: each block sealed with AES-256-GCM under a per-tenant key, prefix reuse that never crosses tenants, and an audit log of access and erasure.

Retention

Both US providers have improved here, and their positions differ:

• Together AI documents zero data retention as an organisation-level setting that is off by default. Once you turn it on, it applies from that moment forward, and data processed earlier has to be deleted separately. Using your data for training is opt-in.
• Fireworks AI documents zero data retention by default for open models, so prompts and generations are not logged or stored. The exceptions it documents are prompt caching, which can hold data and its KV cache in volatile memory for several minutes, opt-in features that log prompts, and its Responses API, which keeps conversation data for 30 days unless you set store=False.

If you choose either one, write the retention setting into your records of processing, check it is on before any personal data flows, and make sure your team knows which API features change it.

The UK data-transfer position

Sending personal data to a provider that processes it in the United States is a restricted transfer under UK GDPR. It is lawful, but you need a transfer mechanism:

1. The UK–US data bridge. If the provider is certified under the UK Extension to the EU–US Data Privacy Framework, you can rely on it. Check the provider's certification on the official Data Privacy Framework list, not a marketing page.
2. Contractual clauses. Otherwise, you need the ICO's International Data Transfer Agreement, or the UK Addendum to the EU Standard Contractual Clauses, plus a transfer risk assessment.

A US-headquartered provider may also be subject to US legal process for data it controls, even when processing happens in an EU region. Many data protection officers weigh that, and some public-sector and regulated buyers rule it out entirely. It is a judgement for your DPO and counsel, not a reason to panic.

With processing in London by a UK company, there is no international transfer to document for inference. That removes a section of your DPIA rather than adding one. UK GDPR and LLMs: a checklist for hosting open models covers the rest of it.

When serverless is the right call

Serverless from Together or Fireworks is often the better choice, and it is worth being plain about when:

• Prototyping and evaluation. Trying ten models in an afternoon is far cheaper per token than standing up a GPU for each.
• Low or spiky volume. A dedicated GPU that sits idle overnight is expensive. Per-token pricing charges nothing when you are idle.
• Non-personal data. Public documents, synthetic data and internal code with no personal data in it carry much less transfer risk.
• Breadth of models. Both catalogues are far larger than kvrun's, and new open models usually appear there first.

When dedicated, governed tenancy wins

• Prompts contain personal, privileged or regulated data, and your DPO wants no international transfer.
• Your contract requires UK-only processing by a UK entity.
• You need proof of isolation and erasure, not only a retention policy: per-session erasure certificates and an audit log.
• Volume is steady enough to keep a GPU busy, so dedicated pricing is competitive.