Track 4 · Strategy · Lesson 19 of 25

Cost, risk and control

The cloud sells you power and reach for a bill that keeps coming. Local sells you certainty for money paid up front.

20 minutes 7 sections Not started yet

What you will be able to do

  • You will be able to describe the shapes of the bill rather than the prices
  • You will be able to name the real risks on each side, including the security ones
  • You will be able to say what you gain and give up with each choice

How each side is paid for

Cloud AI is rented, and most people rent it as a plan. Anthropic sells Claude as a free tier and a range of paid monthly ones, for one person or for a team. Each comes with its own limits on how much you can use. You pay the same every month, and the limits are what tell you to stop.

A limit is not a bill, and it still costs you something. It arrives as a message telling you to come back later. That is paid out of your afternoon rather than your account.

The other shape is a meter, and it is what you get if you build something on a model maker's own service. The price is published per million tokens, where a token is a chunk of a word. Reading costs one rate and writing costs another. Nothing to begin with, then it climbs with use.

Local AI is bought instead. A machine that can hold the model, some power, and your own time. You pay for it whether you use it once a month or two hundred times a day.

The numbers on all three move fast, and a course that printed them would be wrong within a year. So this lesson teaches the shape. Look up the figures on the day you decide, from the vendor's own page.

Backed by Claude model pricing, Claude pricing: the Free, Pro and Max plans.

Where the lines cross

Draw all three as lines on a page. The metered line starts at nothing and climbs with use. The bought line starts high and stays flat. A monthly plan is a low step that repeats for as long as you keep it.

They cross somewhere. Light or occasional use sits on the cheap side of the metered line. Heavy, steady, day-after-day use sits on the cheap side of the flat one.

You can find your own crossing point without knowing a single price first. Count instead. How many of these jobs a month, and roughly how much text goes into each one? Take those counts to the vendor's page on the day you decide, and set the answer against the machine spread over three years.

Do that twice, for a quiet month and a busy one. If the two answers disagree, your choice turns on volume, and volume is then the thing to watch.

Each line hides something. A metered bill grows quietly, a few pennies at a time, and surprises people at the end of a quarter. A plan is paid in full in the months you barely open it. The bought machine costs you hours of setup and tidying that nobody invoices you for.

Your own time is the cost most often left out of this comparison. Put it in. It is the difference between a fair answer and a flattering one.

Risk on the cloud side

Cost is what you pay. Risk is what can change without asking you.

The model you built a habit around can be retired. Anthropic's own deprecation page says it retires older models as newer ones launch, that a retired model stops working, and that requests to one will fail. It promises at least sixty days of notice for public models. That is fair warning. It is still a change you did not choose.

Sixty days is not long if the thing you built leans on one model's exact habits. Prices change too, in both directions. So can the rules about what happens to what you type.

None of this makes the cloud a bad choice. It makes it a rented one. You are a tenant, and tenants get letters from the landlord.

Backed by Claude model deprecations, Is my data used for model training?.

The other cloud risk: after you press send

That first risk is the service changing under you. The second is what becomes of the words once they have arrived.

Good vendors publish this, and it is worth reading once. Anthropic's retention page says a deleted conversation leaves your history at once, and leaves its storage within thirty days. A chat its systems flag as breaking the usage policy can be kept for up to two years. If you turned on the setting that lets your chats improve the model, they may be kept in a de-identified form for up to five years.

None of that is a scandal. It is a clear answer to a question most people never think to ask, and the clarity is the good part.

The risk is the shape of it. This is a promise you can read and cannot operate. You can ask for a deletion. You cannot watch one happen. And a policy can be rewritten, the same way prices and models are.

There is a second link as well. Providers use other providers. The United Kingdom's data regulator says a company handling data for you must put an equivalent contract in place before passing it on again. Your own country will have its own version. So the chain is meant to hold. It is still a chain, and you did not pick every link in it.

Backed by How long do you store my data?, Contracts and liabilities between controllers and processors: when is a contract needed?.

Risk on the local side

Owning is not the safe option. It is a different set of risks, and this course is not going to pretend otherwise.

Your machine can break, and it is yours to fix. Nobody updates your setup while you sleep. If something stops working on a Sunday, that is your Sunday. There is no queue to join and nobody to escalate to.

You can also drift. A model file does not look older as it ages. People run the same one for a year without noticing that much better ones arrived, because nothing ever forced them to look.

And the ceiling is real. When a job grows past what your hardware holds, that is a purchase decision, not a setting you change. The cloud lane has no such wall, which is a large part of what it is for.

The local risk people forget

Local is not the safe option on security either, and a lesson called cost, risk and control has to say so.

Everything sits in one place, and that place has a handle on it. A stolen bag takes the model, the chat history and the files in one go. There is no account to suspend and no support line to ring.

Encryption answers that particular version, and lesson five works through it, so this lesson will not repeat it. The part lesson five does not cover is backup.

The United Kingdom's National Cyber Security Centre is direct about why. Your device might be lost, stolen, or stop working. Its advice for small organisations is to keep a copy of everything the business needs to operate. It suggests two copies, in case one of them is lost or fails.

A local-only lane with no backup is not private. It is fragile. Those are different words, and they get muddled because both feel like being careful.

Notice the symmetry with the other side. A break-in at a provider can reach many people at once. A stolen laptop reaches one person completely. Neither of those is the safe one. They are two shapes of the same bad day.

Backed by Small organisations guide to cyber security: backing up your data.

So what are you actually buying?

Put plainly, in both directions.

Read those three back as risks as well as costs. The cloud's risks arrive as letters you did not ask for. The local ones arrive as a Sunday you had other plans for. Neither list is shorter than the other.

  • With the cloud you buy capability and room to grow. You pay every month or by the token, and you depend on somebody else's decisions.
  • With local you buy certainty and privacy. You pay up front, in money, in your own time, and in a hard ceiling on what you can run.
  • With both, run on a written rule, you buy the right to send each job to the side that suits it. You pay by having to maintain two things instead of one.

Before you move on

Check yourself

Two or three questions to make you recall it. Nothing is deducted for a wrong answer, and you can try again as often as you like.

  1. 1. Cloud AI is rented in two shapes. Which two are they?
  2. 2. What is a real risk of building a habit on one cloud model?
  3. 3. What does a local-only setup with no backup give you?
  4. 4. What does local AI cost you that nobody sends you a bill for?

Where this comes from

  1. Claude model pricing Anthropic documentation checked 8 September 2026
  2. Claude pricing: the Free, Pro and Max plans Anthropic documentation checked 8 September 2026
  3. Claude model deprecations Anthropic documentation checked 8 September 2026
  4. Is my data used for model training? Anthropic documentation checked 8 September 2026
  5. How long do you store my data? Anthropic documentation checked 8 September 2026
  6. Contracts and liabilities between controllers and processors: when is a contract needed? Information Commissioner's Office documentation checked 8 September 2026
  7. Small organisations guide to cyber security: backing up your data National Cyber Security Centre documentation checked 8 September 2026

Stuck on this lesson, or think something here is wrong? Email info@redroostertec.com and a person will answer.