AI Infrastructure in the Cloud Guide

This is an original planning guide. It is not a vendor manual, an official standard, or legal advice. Example addresses use documentation ranges. Review every command before you paste it.

A GPU is a meter that keeps running

AI infrastructure here is the machine or the managed service that would run a model, and the decision to pay for it. The GPU comparison lines up types you type. The platform comparison lines up managed services you are considering. Neither tool launches a machine or uploads data.

Write the stop time before the start time. A GPU lab without a stop is the cost mistake this page is here to prevent. The cost guide records the estimate. This page records the shape and the data you are allowed to use.

The data in a test is invented, or it is data you already have permission to place in that service. A customer extract is not a convenient sample. A prompt or a file you paste into a service you do not administer is a copy you did not mean to make.

A vector store, if the design needs one, belongs to the database desk's vector guide. This page does not stand up a database as a side effect of trying a model.

One shape, a stop time, invented inputs

Decide managed service or your own GPU, and write why. If you choose a GPU, use the comparison to pick a lab size, not the largest. If you choose managed, name the region and what the service is allowed to retain. If you cannot answer retention, you are not ready to send anything but invented text.

Write the stop time and the person who will check it. A calendar reminder is part of the ticket. The instance guide's family lesson still applies: specialised hardware is a justification, not a default.

Prepare an invented input of a few lines. Run nothing in this writing hour unless a separate lab window is already open and the stop time is in force. When that lab ends, the machine is stopped or the job is deleted. A stopped GPU is the proof.

Do not leave a notebook server running because the tab is still open. The tab is not the meter. The meter is the instance or the endpoint. Write which one you stopped.

# AI lab note. Stop time before start time.
# Invent the input. No customer extract.
# A managed service needs a retention sentence.
# The largest GPU is not the lab.
# A vector database is a different guide.

A change you can finish

The lab start is its own window with the stop time already written. This writing hour does not launch.

Rollback is stopping the lab. There is no rollback for data you already sent to a service, which is why the input stays invented.

Choose managed service or your own GPU.
Write the reason and the lab size.
Write retention if a service would keep the input.
Write the stop time and the person who checks it.
Prepare invented input only.
Hand the price to the cost guide.
Stop the lab at the time you wrote, when that window happens.

Mistakes that outlive the command

AI infrastructure fails in the same few ways. Each one looks finished on the day you type it. The list is here so the ticket can forbid them before the window, not so you can recognise them in an incident review.

No stop time.
The largest GPU as a first try.
A customer extract as the sample.
Pasting private text into a service you do not run.
A notebook left running overnight.
A vector database created as a side effect.
Skipping the cost estimate.
A production endpoint as the experiment.
Retention unknown and real data used anyway.
Two shapes launched to compare the bill live.

Decisions to write down before the window

One decision for AI infrastructure: One shape is chosen. The reason is a sentence. The size is a lab size. The stop time is written. A person owns the stop. The input is invented. Retention is known, or no real data is sent. The cost guide has the guess. Nothing launched in the writing hour. A vector store is a separate ticket, or it is not needed. The account is yours. And one experiment is this window.

The hour you stop

Write the account, the project or subscription, the region, and the clock time before you change anything for AI infrastructure. If this account is serving people and you do not have a way back, this hour is a plan. A plan you can read tomorrow is a success. A change you cannot undo is not.

Use a lab account, a lab project, or a lab subscription you administer. You do not point a CLI, a pipeline, or a browser tool at an account you do not run. The tools on this site run in the browser. They draft text. They do not log in, and they do not scan.

prompt-for-this is a reminder that a secret is missing. It is not an access key, a token, a password, or a kubeconfig you save. The real secret stays in the password store or the platform's secret store. If a file, a log, a ticket, or a pipeline still contains a live secret, stop and rotate it. Then delete the copy.

One change. A network, a virtual machine, a policy, and a pipeline do not share a window. If two things move and the outcome is wrong, you will not know which one did it. The rollback is the previous file, the previous plan, or the resource you can delete because it was only a lab.

A second person reads the scope aloud: which account, which region, which name. If they cannot find it in the ticket, the ticket is not ready. Silence is not agreement.

Documentation addresses stay in the note. 192.0.2.0/24, 198.51.100.0/24, 203.0.113.0/24, and 2001:db8::/32 are not prefixes you type into a production network. example.com is not your zone. A generated file that opens 0.0.0.0/0 is a file you edit before it is applied. A source you can name is the rule you keep.

Do not apply a draft you have not read. A cost estimate, a policy, a manifest, and a Terraform plan are reading material. Apply, a resize of a live machine, and a delete of a bucket are later decisions with their own proof. Read the names. Then decide.

Keep the previous file. A generator draft is not the previous file. The previous file is what you put back if the new one is wrong. You practised the path back, including the serial console or a second admin, before you changed a network or a lockout setting.

If a step fails twice, stop. The failure is the result. Adding a command from a forum beside a half-applied stack produces an account nobody can explain. Rebuild the lab from the copy you kept.

When the checks pass, the next account waits until someone else can repeat the proof from the notes. AI infrastructure is done for this window when the proof is written, not when a CLI printed a success line.

This page is a planning guide. It is not a vendor manual, not an official standard, not legal advice, and not an audit opinion. If an auditor must see the decision, they see the ticket and the evidence, not this paragraph.

Hand the finished note to the neighbouring guide that owns the next step, and leave this account alone until that guide has its own window.

AI checks before a GPU or a job starts

Confirm this before you call the change done: The stop time is in the ticket. The owner of the stop is named. The size is not the largest by default. The input is invented. Retention is written, or the service is not used. The cost figure exists or is the next ticket. No customer data is in the note. The account is a lab account. A vector database was not created here. A second person can say when it stops. The start has not happened, or the stop is already in force. The note has no key. The hostname in the prompt matches the hostname in the ticket for AI infrastructure. The result line includes the time and the person, not only the command. You can point at the rollback value without searching a chat history. And done is a sentence a second person can repeat: the proof for AI infrastructure happened, and the next host is not started early. If one of those fails, stop.

Tools for this guide

These tools run in the browser on this site. They do not log into your machines.

Guides in this cluster

These supporting pages belong to this pillar.

Companion: DBA: Vector Databases Explained.