Rent Your AI Like You Could Leave

01 Rent or own: the bill side by side

Mistral, a French AI lab, opened a preview of its new model, Mistral Large 4, on October 6. It says that by the end of October anyone will be able to download the model itself: one very large file. That is what "open weights" means. You can run the file on your own servers instead of renting it.

Renting means you pay for what the model reads and writes, counted in tokens, little pieces of words. Mistral lists this model at $1.36 per million tokens in and $4.18 per million out.

Owning means a server. Even compressed, the file is over 500 gigabytes, and all of it has to sit in the chips at once. You would need a server with eight of the most expensive AI chips, which rents for about $36 an hour at the middle of the market.

Option Rough monthly cost

Rent: 1,000 emails a day, about 2,000 tokens in and 500 out each About $145 (under $5 a day)

Own: the server running around the clock About $25,000

Own: switched off at night and on weekends Near $6,300

Own: cheapest chip price found / Amazon's cloud About $12,000 / $45,600

At the middle of that range, owning only breaks even at roughly 178,000 of those emails a day, before you pay the people who keep it running. And Mistral is selling the rented version at half price for its launch.

Almost nobody should download it. But the download changes the AI you already rent.

02 Three questions that sort your work

Run each kind of AI work in your business through these. Tick the ones it passes.

The data has to stay inside. Client files, yes. The newsletter and meeting notes, no.

It would hurt if the model changed or refused halfway through a job. This matters most to a security team that can't have its tools refuse mid-incident.

The volume is huge and steady. Think the 178,000 emails a day from section 1, every day.

Who passes all three? A bank. A ministry. A security team. That's who the file is really for.

Worked example: a 40 person accounting firm

Work Where it goes

Reading a client's return and drafting the summary for the partner Something the firm controls

Newsletter, meeting notes, everything else Stays rented

"Something it controls" doesn't have to be a trillion parameters. Google's Gemma 4 came out in April under a fully open licence, and its 31 billion parameter version fits on a single data center chip. Or Mistral will set up a big model in a private cloud or on the firm's own servers, so the data stays in its environment.

03 Rent like someone who could leave

When a model's file is public, the company renting it to you isn't the only one who can run it. That's an exit. If the price jumps or the terms change, you can move the same model somewhere else, if the licence allows it. Start with three questions for the provider you rent from today.

Three questions for your AI provider

1. Where does it run? In which country, and under which law?
2. Can you train on our data? Have we opted out?
3. Do we have zero data retention?

Asked on [DATE] by [NAME]. Answers: [PASTE]

For reference: Mistral says it will serve this model from a European deployment it runs itself, under European law. Its own data agreement says it can train on customer data unless the customer has opted out, and it mentions a setting called zero data retention: the provider doesn't keep what you send. Find out which side of each line you're on.

04 Your own ten-example test

The third check is whether you can switch. Keep your instructions and ten real examples of your work written down outside the tool. When something better or safer shows up, you can test it on your work in an afternoon.

Ten-example test

Job: [THE WORK, E.G. "DRAFT A REPLY TO A CLIENT EMAIL"]
Instructions we give the AI: [PASTE THEM HERE]

Example 1
Input: [A REAL EXAMPLE, PRIVATE DETAILS REMOVED]
A good answer looks like: [WHAT YOU WOULD ACCEPT]
...
Example 10

Model tested: [NAME] Date: [DATE] Good answers: [ ] of 10

05 This month, and the end of October

Do this month

Check your AI provider's training setting. Write down the one kind of work whose data should never leave. Save ten real examples as your own test.

Watch when the file comes out

The licence. It isn't public yet. Mistral's pricing page says commercial use of its open models needs a Mistral licence. Which compressed version ships. It changes the hardware you'd need by up to four times. Independent tests. Almost every number so far is Mistral's own. On one independent index, Vals, the model placed 32nd out of 44.

If you do decide to own. Running it yourself moves the guardrails, and the blame, onto your own team.

Sources

Mistral: Introducing Mistral Large 4 Mistral Docs: Mistral Large 4 Mistral: Pricing Mistral: Data Processing Addendum Yotta Labs: Mistral Large 4 hardware requirements Beam: NVIDIA H200 pricing Technology.org: Mistral Large 4 arrives with 1 trillion parameters The Decoder: Gemma 4 under Apache 2.0 VKTR: Mistral Large 4 debuts as a 1T-parameter open-weight model

Mathew's LibraryLibrary card

Free access

Keep reading free with your library card

One email opens this checklist and all 3 resources in Mathew's library. No account, no password.