You cannot see a private GPU cloud's finances, so check what you can. Work out how much money you would lose if it stopped. Then check how fast you could move your traffic to another provider.
Commit only as much money as you can afford to lose. Sign a long contract only once you know you could move within days.
On our sales calls, buyers ask whether we will still be around. One put it like this: "My boss is going to ask me, you know, are you guys going anywhere?" It is a fair question. Often the buyer has to answer it for whoever approves the budget.
Why neocloud staying power is a fair question
A neocloud is a company that rents out GPUs and AI services but is not one of the big cloud providers. We are one of them, so this question is also about us. We run open models for you as an API: you send a request and pay per token, and we run the GPUs.
Analysts expect a shake-out. Gartner predicts that more than 80% of neocloud providers will cease operating independently by 2030. The reasons it gives are mounting costs and offerings that are hard to tell apart. Gartner adds that this does not mean they all fail: some could be acquired or merge.
So the event you plan for is not only a shutdown. A sale, a merger or a dropped product can change your price, your contract or your endpoint just as well.
Lyceum is a sovereign European AI inference provider for open-source models: a company built and run in Europe that serves open models. Our serverless models run in Paris and Finland. Your prompts and outputs are processed, not stored, and never used for training. Serverless means you send requests and pay per token, without renting or managing any machine. We won't ask you to trust a claim about our finances in an article. The checks below work on any provider, us included.
What you actually lose if a provider stops
Two things are at stake: money you have already paid, and the time it takes to move. Money is the easier one to measure, so start there.
Take a hypothetical team that spends €20,000 a month on inference. In the third option below, the team tops up its credits every two weeks, so its balance never exceeds half a month of spend. It can pay in three ways, and each leaves a different amount exposed if the provider stops after month four.
Monthly inference spend (hypothetical): €20,000
Prepaid 12-month commitment: 12 × €20,000 = €240,000
Provider stops after month 4: 8 × €20,000 = €160,000 unused
Monthly billing, at most 1 month: 1 × €20,000 = €20,000 at risk
Credits kept at two weeks (half a month): €20,000 / 2 = €10,000| Way of paying | Money at risk if the provider stops after month 4 |
|---|---|
| 12 months prepaid | €160,000 |
| Billed monthly | €20,000 |
| Pay-as-you-go credits, two weeks of use | €10,000 |
The example assumes the worst case: a full stop, no refund and nothing recovered in an insolvency. A sale or merger may keep the service running. It also leaves out two things: the engineering time to move, and the lower price a longer commitment can buy.
The second cost is time. How long you need to move depends on your models, your API and your contract. We cover the hours and steps in what switching inference providers actually costs.
Five checks that need no financial statements
A private company rarely shows outsiders its accounts. You don't need them to judge your own risk. Ask these five questions of any provider, including us.
- Money at risk. How much is prepaid or committed, and for how long? Run the calculation above with your own spend.
- Time to move. Are the models open and the API standard, so another provider can serve the same model? An OpenAI-compatible API (one that accepts requests in OpenAI's format) keeps your code unchanged.
- Data to get back. What does the provider store that you would need? We don't keep your prompts, and outputs aren't kept either.
- Change notice. Will you hear before someone new processes your data, and can you object? A subprocessor is a company that processes your data on the provider's behalf. With us, you can object to a new subprocessor within 10 business days of our notice, and you can ask for the list.
- Support and status. Who answers when something breaks, and where do you see incidents? Business customers get a direct line to the engineers who run our platform, with contractual response times. Incidents show on our status page.
Each answer either lowers your exposure or tells you how long a move would take. Together they replace the balance sheet you can't see.
Open models make leaving a three-line change
Time to move depends first on whether another provider can serve the same model. A closed model is sold only by its maker and the partners it picks. An open model is one whose weights, the trained numbers that make up the model, are published. Any provider can download them and serve the same model.
Our serverless endpoints are OpenAI-compatible. You point any OpenAI client at our base URL and swap in our API key and a model name. Streaming, tool calling and usage accounting then work the same. Here is the whole connection:
from openai import OpenAI
client = OpenAI(
api_key="lk_...",
base_url="https://api.lyceum.technology/openai/v1",
)
resp = client.chat.completions.create(
model="z-ai/glm-5.3",
messages=[{"role": "user", "content": "Summarise this ticket."}],
)
print(resp.choices[0].message.content)To leave, you change the same three values: base_url, api_key and model. We want you to have this easy exit. It is also what makes starting with us low-risk. Before you move a production workload, evaluate the model on the new endpoint. Providers can run the same model with different settings, and that can change its answers.
A product can disappear while its company carries on. Lambda, a US GPU cloud, is one example. Its inference page says its Inference API winds down and points customers to renting NVIDIA GPU machines instead. We checked that page on 6 November 2026. We serve open models such as Kimi, GLM, DeepSeek and Qwen through one OpenAI-compatible API. It is billed per token with no base fee. We checked our side on the same day.
Write the exit plan before you sign
EU financial firms already have to do this, and their rules make a good template. DORA, the EU Digital Operational Resilience Act, has applied since 17 January 2025. DORA applies to banks, insurers and other financial firms. They need an exit strategy for each outside IT provider that supports a critical or important function, meaning a service the business cannot run without. The strategy must cover a possible failure of the provider.
DORA also says what a good exit plan contains. It must be documented, tested and reviewed periodically. The firm must name alternative providers and plan how to move the services and the data. You are probably not a bank, but the same three parts work for any team:
- Name the alternative. Write down which provider would serve your models if yours stopped.
- Test the switch. Send a day of real traffic through the alternative and compare the results.
- Review it. Repeat the test when you add a model or change providers.
The EU Data Act points the same way. It requires cloud providers to meet minimum requirements that make switching to another provider possible. Whether it covers a given inference API is a question for your lawyer. For the full mechanics of a planned exit, see how to avoid AI vendor lock-in.
When a longer commitment is worth it
A longer commitment can lower your price. It is worth it when two conditions hold: your traffic is steady, and your exit plan has passed a test. If either condition is missing, a lower rate is not worth prepaying money you could lose, like the €160,000 in the example.
We recommend pay-as-you-go or monthly billing until your exit test has passed. With us, you can keep a small balance and top it up as you go. Our credits have no expiry date, and we set no minimum spend. When traffic steadies, you can reserve GPUs or a whole cluster, a group of linked GPU machines, to run your own models. We hold that capacity for you and price it per contract. If you would rather not run the model yourself, a dedicated endpoint is a model we serve on GPUs reserved only for you. For dedicated endpoints, our commitment pricing is agreed per contract, so you know the cost before you launch.
For a wider list of what procurement asks, use our AI vendor risk assessment checklist.
Your next step: run a week of your workload on our serverless inference with a small credit balance. Test your exit in the same week, and decide on a commitment after that.
Frequently asked questions
Does Lyceum offer an uptime SLA (a written availability promise), and what availability guarantees apply to serverless inference, dedicated endpoints and GPU capacity?
We don't publish a standard uptime SLA for serverless inference yet, and it's on our roadmap. For dedicated endpoints and reserved GPU capacity, we agree service levels with you as part of your contract.
Can I test Lyceum for free before committing, for example with trial credits or a proof of concept?
For serverless inference, you pay as you go with no commitment, so you can start small and scale when you're ready. For dedicated inference, we run time-boxed proofs of concept with success criteria we agree on before you commit. The customer pays for these proofs of concept. They usually run for a few days, up to a week, depending on the workload.