Protect your data. Depend less on one AI provider.

When all your AI work goes to one provider, you put four things in its hands: your data, your access, the model you tested and your ability to leave.

Will our data be stored or used for training?

Prompts carry whatever your people paste into them: source code, contracts, customer records, credentials. Once a provider holds that data, its terms decide how long it keeps it and what it does with it. Those terms change.

  • 2025Consumer chats used for training and kept for five years, unless users opted out

    2025

    Consumer chats used for training and kept for five years, unless users opted out

    In August 2025 one lab asked its consumer users to let their chats train its models, and extended retention to five years for those who agreed. Its API and business accounts were excluded.

    Source: Anthropic, 28 August 2025.

  • 2025A court ordered one lab to keep chats its users had deleted

    2025

    A court ordered one lab to keep chats its users had deleted

    In May 2025 a US court ordered one lab to keep chat logs that its users had deleted. The duty ended in September 2025, but logs saved before then stayed accessible.

    Source: Engadget, 6 June 2025; Engadget, 11 October 2025.

  • 2026Chats made with history off kept for two months instead of 72 hours

    2026

    Chats made with history off kept for two months instead of 72 hours

    In November 2026 one AI provider said it would keep chats made with history switched off for about two months, up from 72 hours.

    Source: TechCrunch, 17 November 2026.

  • 2026A $1.5 billion settlement over books one lab downloaded from pirate sites

    2026

    A $1.5 billion settlement over books one lab downloaded from pirate sites

    In July 2026 a US court gave final approval to one lab's $1.5 billion settlement with authors and publishers over books it had downloaded from pirate sites.

    Source: TechCrunch, 20 July 2026.

On Shoreheld

Used to train a model?
Never
Not your prompts, not your responses.
Kept after we reply?
No copy
Batch files are kept only until the job is done.
Prompt text in our logs?
Metadata only
Time, key, model, region, token counts, status.
Readable by our staff?
No
Logs, error logs and backups hold no prompt text.
Sent to the model's maker?
Nothing
Model servers cannot connect out.
Where is it processed?
Your region
UK, EU, Switzerland or Singapore.

These commitments cover the Shoreheld API. Requests you send to another provider, including through our gateway, follow that provider's terms. How we handle your data

Your AI access can be restricted or cut off

A provider can suspend an account, withdraw a model or ration requests when demand outruns its GPUs. When it does, the work that depends on it stops, often with little notice.

  • 2025A coding tool's access cut with less than five days' notice

    2025

    A coding tool's access cut with less than five days' notice

    In June 2025 one lab cut nearly all of a coding tool's direct access to its models with less than five days' notice, according to the tool's maker.

    Source: TechCrunch, 3 June 2025.

  • 2026Usage limits tightened at peak hours to manage demand

    2026

    Usage limits tightened at peak hours to manage demand

    In March 2026 one AI provider tightened session limits on its free and paid consumer plans at peak hours, to manage growing demand.

    Source: The Register, 26 March 2026.

  • 2026Two newest models switched off for every customer

    2026

    Two newest models switched off for every customer

    In June 2026 a US export-control directive led one lab to switch off its two newest models for every customer. One of them returned 19 days later.

    Source: Anthropic, 12 June 2026; Anthropic, 30 June 2026.

On Shoreheld

Can a model's maker withdraw a model we use?

It can shut its own hosted service. The weights it has published stay published, and we serve the version you use from our own copy.

Can Shoreheld suspend our account?

Only for non-payment, abuse or a legal order.

What happens when demand outruns the GPUs?

On the shared API each key has rate limits. Over a limit, or when shared capacity is full, the API returns a 429 with a Retry-After header. We don't queue requests.

What does reserved capacity buy?

GPUs allocated to you alone. The order names the GPU count, the region or country, the start date and the term. We agree an availability target with you; there is no public SLA yet.

Can we leave during an incident?

Only if you have prepared. The same open-weight models run with other hosts or on hardware you run, and our API uses the OpenAI format. Set up and test that alternative before you need it.

The model you tested can change under the same name

A provider can update a hosted model's weights or change how it serves them, and answers you tested start to drift.

  • 2025Bugs degraded one lab's answers for weeks before they were fixed

    2025

    Bugs degraded one lab's answers for weeks before they were fixed

    Between August and early September 2025, three infrastructure bugs intermittently degraded one lab's answers until its engineers found and fixed them.

    Source: Anthropic Engineering, 17 September 2025.

On Shoreheld each model ID points to one published version of the weights, at one precision. We don't change either behind the ID. A new version gets a new ID, and we give at least 90 days' notice before we retire one.

We still update the software that serves the models, and a bug there could change answers until we fix it. Outputs can also vary from one run to the next, as with any model.

How model versions change on ShoreheldThe model ID you use serves the same weights at the same precision until it is retired. A new version arrives under its own ID, and you move once you have tested it. If we retire your ID, we give at least 90 days' notice first.The model ID you useRetiredSame weights, same precision90 days' notice or moreA new versiongets its own ID. Move to it onceyou have tested it.

Prices can rise. You need a way to leave.

Once your systems depend on one provider's API, it sets the price and you pay it. Moving then means rewriting integrations under pressure.

We publish our token prices and ask for no minimum spend. Our API follows the OpenAI format, so most clients need only a new base URL, key and model ID. The models are open-weight: anyone may download and run them under their licences, so you can take the same work to another host if our prices or service stop suiting you.

Hosted tools, connectors and other provider-specific features don't carry over, and a different model can behave differently. Test before you switch.

Ways to run the same workYour code calls the API in the OpenAI format. The same calls can go to Shoreheld, to another host serving the same open model in the same format, or to hardware you run with the same model files. To move, you change the base URL, the key and the model ID. Features that only one provider offers stay behind.Your codeCalls in the OpenAI formatShoreheldOpen models, your regionAnother hostThe same open modelHardware you runThe same model filesTo move, change the base URL, keyand model ID. Features only oneprovider offers stay behind.

Have a tested alternative if your AI provider fails

Financial regulators expect you to document and test a plan for leaving a critical supplier. A plan with no working alternative is hard to test.

  • United KingdomDocument and test an exit plan for each material outsourcing.

    United Kingdom

    Document and test an exit plan for each material outsourcing.

    The PRA expects a firm to document and test an exit plan for each material outsourcing, covering a stressed exit such as the provider's failure.

    Source: PRA Supervisory Statement SS2/21, paragraphs 10.1 and 10.10.

  • European UnionKeep documented, tested exit strategies for critical ICT services.

    European Union

    Keep documented, tested exit strategies for critical ICT services.

    Financial entities must have exit strategies for ICT services that support critical or important functions. The plans must be documented and tested, and must allow a move to another provider or back in-house.

    Source: Regulation (EU) 2022/2554 (DORA), Article 28(8).

  • SwitzerlandMake sure an outsourced function can move or come back in-house.

    Switzerland

    Make sure an outsourced function can move or come back in-house.

    An orderly return of an outsourced function, or its migration to another provider, must be ensured.

    Source: FINMA Circular 2018/3, margin no. 18.1 (English translation).

  • SingaporeKeep meeting your obligations if a provider fails.

    Singapore

    Keep meeting your obligations if a provider fails.

    A bank should make sure it can keep meeting its obligations if a provider fails or an outsourcing arrangement ends unexpectedly.

    Source: MAS Guidelines on Outsourcing (Banks), paragraph 3.7.2.

Shoreheld can be the alternative in that plan, and you can test it before you need it: the same open models, a named region for each request, and capacity you can reserve in the country you choose.

We are a supplier too, so plan your exit from us. The models are published weights and the API follows the OpenAI format, so the same work can move to another host or to hardware you run.

Choosing a region or reserving capacity does not make you compliant. This is a summary, not legal advice.

Questions buyers ask

Do we need the gateway to use the API?

No. The API works on its own with any client that speaks the OpenAI format. The gateway is for teams that also use other AI providers and need to control what reaches them.

How do we pay?

Pay for tokens by card through Stripe, with no minimum spend and no fixed term. Our terms of service still apply. We quote dedicated endpoints, reserved nodes, gateway set-up and hosting, and Lab work.

What assurance do you offer?

Independent certification is not in place yet. Your security team can review the gateway's design and run it in your own account, but that review does not cover the hosted API.

Try it on your own work

Create a key and pay by card. For reserved capacity or an exit plan, talk to us.