Amazon Bedrock vs SageMaker AI vs Canvas vs Studio: Which AWS AI Service, Explained (2026)

Amazon Bedrock vs SageMaker AI, in one sentence: Bedrock lets you rent finished AI models (Claude, Amazon Nova, Llama and others) through an API and pay per word processed, while SageMaker AI is the workshop for building, training and hosting your own machine-learning models, paid by the hour for the machines you use. The rest of the family lives inside SageMaker AI: Canvas is its no-code, point-and-click tool for business users ($1.90 for every hour you are logged in), Studio is its workspace for developers (JupyterLab, a VS Code-based editor and RStudio), and JumpStart is its catalog of ready models that you deploy onto your own servers. Here is the twist that confuses almost everyone: since the end of 2024 there are two things called SageMaker. The original service was renamed SageMaker AI, and the plain name SageMaker now belongs to a newer umbrella platform whose workspace, SageMaker Unified Studio, brings data tools, SageMaker AI and Bedrock together. The money difference is just as sharp: for a small business chatbot, Bedrock can cost well under a dollar a month, while the same idea on a SageMaker GPU endpoint left running costs about $1,000. Below, we start from the basics and then compare all five, with real prices.

Jake's phone repair shop had two AI ideas. The first was a chatbot that answers "is my repair ready?" and "how much for a cracked screen?" on the shop's website. The second was a prediction: which repairs are likely to come back within a month, based on two years of his repair records. He opened the AWS console, searched "AI," and found Bedrock, SageMaker AI, SageMaker Canvas, SageMaker Studio, SageMaker JumpStart and SageMaker Unified Studio, all claiming to help. A tutorial walked him through deploying an open model from JumpStart, so he did. He forgot about it over a long weekend, and the next week his bill showed about $68 for a GPU server answering nobody. Ethan's first question was the one this guide answers: which of your two ideas needs a model you build, and which needs a model you rent?

The basics: models, training, and the two ways AWS charges for AI

Five ideas make every comparison in this guide easy to follow.

• A model is the trained "brain" that does the work: it reads something and produces an answer, a prediction or an image.
• A foundation model is a very large, general model trained by a lab such as Anthropic, Amazon, Meta or Mistral on huge amounts of text or images. Claude and Amazon Nova are foundation models. You use them as they are, or adjust them a little.
• Training is teaching a model from data. Training your own model from your own records, like Jake's repair history, is classic machine learning, and it needs compute power for hours.
• Inference is using a trained model to answer. Every chatbot reply and every prediction is inference.
• Serverless vs. servers. With a serverless service, AWS runs everything and you pay only per request. With a server-based one, you choose a machine size, it runs until you stop it, and you pay by the hour whether anyone uses it or not.

That last point is where most surprise bills come from. Bedrock is serverless and charges by tokens, the small chunks of text a model reads and writes; a token is roughly three-quarters of a word. Most of SageMaker AI is server-based and charges by the instance-hour, like renting a car by the day.

"So Bedrock is a taxi and SageMaker is renting the car," Jake said.

"Close enough," Ethan said. "A taxi costs you only while you ride. A rental car costs you every hour it sits in your driveway. You left the rental running all weekend."

The two SageMakers: the naming, untangled

If AWS's AI names feel like a maze, it is not you. They changed three times in two years, and old tutorials use the old names. Here is the timeline, plainly:

• November 30, 2023: SageMaker's original web workspace was renamed SageMaker Studio Classic, and a new, faster SageMaker Studio replaced it.
• End of 2024: the service everyone knew as Amazon SageMaker was renamed Amazon SageMaker AI. At the same time, AWS gave the plain name Amazon SageMaker to a new umbrella platform for data, analytics and AI together.
• March 2025: that platform's workspace, SageMaker Unified Studio, became generally available, with Bedrock's building tools inside it. What used to be called Bedrock Studio, then Bedrock IDE, now lives there as "Amazon Bedrock in SageMaker Unified Studio."

When you read...	It usually means
"SageMaker" in a tutorial from before 2025	What is now called SageMaker AI
"SageMaker AI"	The ML service: training, endpoints, Canvas, Studio, JumpStart
"SageMaker" in AWS material from 2025 on	Often the umbrella platform (data + analytics + AI)
"SageMaker Studio"	SageMaker AI's developer workspace (JupyterLab, Code Editor, RStudio)
"SageMaker Unified Studio"	The umbrella platform's workspace, which includes Bedrock tools and data tools
"Bedrock Studio" or "Bedrock IDE"	Old names for Bedrock's tools inside Unified Studio

So "SageMaker Studio vs SageMaker" is a fair question with a simple answer: Studio is the front door of SageMaker AI, while "SageMaker" on its own may mean either the ML service or the new umbrella, depending on the date of what you are reading.

Amazon Bedrock: rent a finished AI model

Bedrock is AWS's serverless service for using foundation models through an API. You do not train anything and you do not manage servers. Your app sends a request ("summarize this," "answer this customer"), Bedrock passes it to the model you chose, and the reply comes back. You pay for the tokens in and out.

What you get, beyond the models themselves:

• Model choice. Proprietary models such as Anthropic's Claude and Amazon Nova, open models from Meta (Llama), Mistral AI, Qwen and OpenAI's open-weight GPT-OSS, image models, and many more. Some models, Claude among them, are available through Bedrock but not through SageMaker JumpStart.
• Bedrock Marketplace, with over 100 additional models you can try.
• Building blocks for real apps: Knowledge Bases (answer from your own documents, often called RAG), Guardrails (block unsafe or off-topic answers), AgentCore (run AI agents in production), Flows (chain steps together), Data Automation (pull information out of documents, images, audio and video) and Prompt management.
• Familiar APIs. Besides its own API, Bedrock offers OpenAI-compatible endpoints, so an app written for OpenAI's API can often switch by changing the base URL and key.
• Light customization. Fine-tuning, distillation (a cheaper model taught by a bigger one), reinforcement fine-tuning, and importing a model you customized elsewhere.

Pricing: per input and output token, with prices set per model. On top of standard on-demand pricing there are service tiers (Priority for speed, Flex for cheaper non-urgent work, Reserved for committed capacity), batch inference at up to 50% off for jobs that can wait, and provisioned throughput for steady high volume. Features such as Knowledge Bases and Guardrails are billed separately. There is no Bedrock-only free tier: you pay from the first request, which at small volumes is often a few cents.

Choose Bedrock when you want AI in an app or workflow (a chatbot, document summaries, classification, an agent) and a good existing model can do the job with clear instructions and your documents. That is most business uses, and it needs no machine-learning background. If you want the full beginner picture first, our plain-English Bedrock guide starts from zero.

Try Bedrock in ten minutes

The fastest way to understand Bedrock is to use it once, in the console, before writing any code.

1. Sign in to the AWS console and pick a Region where the models you want are offered, such as US East (N. Virginia).
2. Open Amazon Bedrock and choose the Chat / text playground.
3. Select a small, inexpensive model first, such as Amazon Nova Micro, and ask it a question from your real work: "Write a polite reply telling a customer their screen repair is ready for pickup."
4. Switch to a larger model and ask the same question. Compare the answers, and compare the price per million tokens on the Bedrock pricing page.
5. Open the model's details and look at the API request example. That snippet is all your app needs to do the same thing in code.
6. Set a budget alert before you build anything bigger.

If the playground says you do not have access to a model, your IAM user or role is missing Bedrock permissions, or the model is not offered in that Region; try another Region or model. Ten minutes of this teaches more about "rent a model" than any diagram.

Amazon SageMaker AI: the workshop for your own models

SageMaker AI (formerly just SageMaker) is AWS's managed service for the whole machine-learning lifecycle: prepare data, train models, tune them, deploy them to endpoints, and monitor them. It covers both kinds of AI: classic predictive models (fraud scores, demand forecasts, "will this repair come back?") and generative models you customize or host yourself.

• Training jobs run your training code on machines you pick, and stop when done.
• Endpoints host a model so your app can call it. You choose the machine type and the number of machines; endpoints can now scale down to zero when idle, if you configure them to.
• Serverless model customization lets you fine-tune open-weight models with techniques such as supervised fine-tuning (SFT), DPO and reinforcement learning, through a guided interface, billed per token rather than per server.
• HyperPod is for very large training and inference clusters, the kind used by teams training their own big models.
• JumpStart, Canvas and Studio are all parts of SageMaker AI, covered below.

Pricing: mostly per instance-hour for notebooks, training and endpoints, plus storage; prices depend on machine size. Savings Plans cut compute costs by up to 64% with a one- or three-year commitment. New users get a two-month free trial that includes 250 hours of ml.t3.medium notebooks, 50 hours of m4.xlarge or m5.xlarge training, 125 hours of m4.xlarge or m5.xlarge real-time inference, and 150,000 seconds of serverless inference.

Choose SageMaker AI when you need a model that does not exist yet: one trained on your own data for your own prediction, or a foundation model changed more deeply than prompting allows. Also choose it when you must control the hardware, cost and speed of hosting yourself. Jake's "which repairs come back" idea is a SageMaker AI job; his chatbot is not.

SageMaker Canvas: machine learning without code (and what it costs)

Canvas is the part of SageMaker AI built for people who do not write code: analysts, operations staff, small business owners. You bring a spreadsheet or a table from a database, pick the column you want to predict, and Canvas builds, tests and explains a model with point-and-click steps. It handles tabular predictions (will this customer churn, what will sell next month), forecasting, and image and text models. It also offers ready-to-use models for tasks such as reading documents or detecting sentiment, and access to foundation models for chatting with your data.

For Jake's prediction, Canvas is a natural start: upload two years of repair records, choose "came back within 30 days" as the target, and let Canvas try models and show which factors matter most.

Build a prediction in Canvas, step by step

Here is the path Jake followed for his "will this repair come back?" model, in plain steps:

1. In the SageMaker AI console, set up a domain with the quick setup option the first time. This creates the workspace and its permissions.
2. Open Canvas from the console or from Studio.
3. Import your data: upload a CSV file, or connect to a source such as an S3 bucket. Jake exported two years of repair records with columns such as phone model, repair type, parts used, technician and "came back within 30 days."
4. Create a new model, choose the dataset, and pick the target column, the one you want to predict.
5. Choose a Quick build to get a first result fast, or a Standard build for a more accurate model that takes longer.
6. Read the Analyze results: Canvas shows how accurate the model is and which columns influence the prediction most. For Jake, one supplier's battery batch stood out immediately.
7. Use Predict on new records, one at a time or in a batch.
8. Log out from the Canvas menu when you finish, so the session charge stops.

Two pieces of advice that save frustration: clean obvious errors out of the spreadsheet first (blank rows, mixed date formats), and keep a few months of data aside to test the model on records it has never seen. A model that looks brilliant on the data it learned from can still be wrong about next month.

How much does SageMaker Canvas cost?

The average cost of Canvas depends mostly on one thing: how many hours you are logged in. The prices, in US East (N. Virginia):

What you pay for	Price
Workspace session (being logged in)	$1.90 per hour
Training image or text models	About $2.03 to $4.89 per hour of training
Training tabular models	Billed by the SageMaker compute hours used
Predictions on tabular data up to 5 GB	No additional charge
Foundation models inside Canvas	Per token, like Bedrock
Free tier	160 workspace hours a month for the first two months

Worked example: someone who uses Canvas for two hours a day on 22 working days spends 44 hours, which is 44 × $1.90 = $83.60 a month in session charges, plus any training. Training an image model on about 1,000 images costs roughly $1.68. The trap is the session meter: it runs until you log out. Closing the browser tab is not enough. A Canvas session left logged in for a whole month is 730 hours × $1.90, about $1,387. Logging out stops the workspace charge, and your models and data stay saved.

SageMaker Studio: the workspace for ML developers

SageMaker Studio is the web-based workspace inside SageMaker AI where data scientists and ML engineers write code. It gives you a choice of tools that open in your browser: JupyterLab for notebooks, Code Editor (based on the open-source Visual Studio Code), RStudio, and the older Studio Classic for teams that have not moved yet. From Studio you can also see your training jobs, deploy and monitor endpoints, and browse JumpStart models, all in one place.

Studio is organized into domains (a team's setup) and spaces (a user's workspace with a chosen machine size). The workspace itself is not where the cost comes from; the machine behind each running space is, charged per hour. A small ml.t3.medium space forgotten for a month costs about $36, and a GPU space far more, which is why Studio offers idle shutdown. Turn it on.

SageMaker notebooks vs SageMaker Studio

Before Studio, SageMaker offered notebook instances: a single Jupyter server on a machine you start and stop yourself. They still exist and work fine for one person and one notebook. Studio is the modern choice for anything more: several tools, shared team setups, faster start-up, and everything about training and deployment in one interface. If a tutorial says "create a notebook instance," it is older; the same steps work in a Studio JupyterLab space.

SageMaker AI vs SageMaker Studio

This is a question of service vs. window. SageMaker AI is the service: the training, hosting and tools. Studio is the window you use to work with it. You can use SageMaker AI without Studio (from code, the CLI or the console), but you cannot use Studio without SageMaker AI.

SageMaker JumpStart vs Bedrock: two ways to use the same kind of model

This comparison causes the most expensive mistakes, so it deserves care. Both give you ready-made models. The difference is who runs the server.

• JumpStart is SageMaker AI's model hub, with nearly 1,000 models: open models from Hugging Face, Meta, Stability AI and Amazon, plus some proprietary ones. You pick a model and deploy it to an endpoint you own, on a machine you choose. You pay for that machine every hour it runs, used or not. In return you get full control: the exact model version, the hardware, the network, fine-tuning on your own terms.
• Bedrock runs the model for you. No endpoint, no machine, no idle cost: you pay per token only when your app actually calls it.
• Bedrock Marketplace sits in between: you pick from its catalog through Bedrock, but the model runs on a dedicated endpoint, so it is billed more like JumpStart.

For Jake's chatbot, Ethan did the arithmetic on paper. Suppose the shop's website gets 2,000 questions a month, each sending about 1,500 tokens (the question plus instructions and shop details) and receiving about 300 tokens back:

Option	How it is billed	Monthly cost
Bedrock, Amazon Nova Micro	3M input tokens × $0.035/M + 0.6M output tokens × $0.14/M	About $0.19
JumpStart model on one ml.g5.xlarge, always on	730 hours × $1.41	About $1,029
Same endpoint, shop hours only (8 h × 22 days)	176 hours × $1.41	About $248
Jake's forgotten long weekend	48 hours × $1.41	About $68

Prices are US East on-demand list prices, and a stronger Bedrock model costs more per token than Nova Micro. But even a model priced many times higher stays far below an always-on GPU at Jake's volume. The endpoint wins only at large, steady volume, or when you need a specific open model with full control. Our Bedrock pricing guide works through costs for stronger models.

SageMaker Unified Studio, and "Bedrock in SageMaker Unified Studio"

SageMaker Unified Studio is the workspace of the new umbrella SageMaker platform. It is built for teams whose AI work and data work overlap: in one place, people can find and query data (with tools such as Athena, Redshift, EMR and Glue), build ML with SageMaker AI, and build generative AI apps with Bedrock's tools, all organized into shared projects with common permissions. Amazon Q Developer is built in to help write SQL and code.

"Amazon Bedrock in SageMaker Unified Studio" is how Bedrock's building tools appear there: a governed place to try models, build chat apps, and add Knowledge Bases, Guardrails, Agents and Flows, then share them within the team. It is still Bedrock underneath, with the same models and the same per-token pricing. Unified Studio is where you work; the services you run inside it are what you pay for.

Do you need it? For a single developer adding a chatbot to a website, no; the Bedrock console and API are simpler. For a company where analysts, data engineers and AI developers share data and need one governed workspace, it is designed exactly for that. Organizations often connect it to their company sign-in through IAM Identity Center.

All five side by side

	Bedrock	SageMaker AI	Canvas	Studio	JumpStart
What it is	Serverless API for ready models	Service to build, train and host models	No-code ML inside SageMaker AI	Code workspace inside SageMaker AI	Model catalog inside SageMaker AI
Who it's for	App developers, businesses	ML engineers, data scientists	Analysts, non-coders	Data scientists, ML engineers	Teams hosting open models
You pay	Per token	Per instance-hour, mostly	$1.90 per logged-in hour, plus training	Per hour of each running space	Per hour of each endpoint
Idle cost	None	Yes, for anything left running	Yes, until you log out	Yes, until the space stops	Yes, until you delete the endpoint
ML skill needed	Little	Moderate to high	None	Coding and ML	Some
Jake would use it for	The website chatbot	A custom model, if Canvas is not enough	The repair-return prediction	If he hires a data scientist	Not yet

Can Bedrock do predictions like forecasting or churn?

Not in the way people hope. Bedrock is built around generative models that read and write text and images. You can ask one to "guess" whether a customer will leave, but it is not trained on your history and cannot measure its own accuracy against your data. Numeric predictions from your own records (forecasts, churn, fraud scores, which repairs will come back) are classic machine learning, and they belong in SageMaker AI or Canvas, where the model is trained on your data and tested on records it has not seen. Many real projects use both: a SageMaker AI or Canvas model produces the prediction, and a Bedrock model turns it into a clear explanation or a customer message. That is exactly how Jake's two ideas eventually met: the prediction flags a risky repair, and the chatbot drafts a friendly check-in message for the customer.

Which one should you choose? Five questions

1. Can an existing model do the job with good instructions and your documents? Chatbots, summaries, writing help, document Q&A, classification: start with Bedrock.
2. Do you need a prediction from your own records (churn, returns, demand, fraud)? If you do not code, start with Canvas; if you do, use SageMaker AI through Studio.
3. Did prompting and documents stop being enough, so the model's behavior itself must change? Try Bedrock's fine-tuning first, then SageMaker AI serverless customization.
4. Must you run a specific open model on hardware you control, for compliance, latency or very high steady volume? Use JumpStart endpoints, and plan the hours.
5. Is this a team effort across data and AI, with shared data and governance? Work in SageMaker Unified Studio, which uses all of the above underneath.

AWS itself suggests the same path for generative AI: start with Bedrock (prompting plus your documents), move to SageMaker AI customization when you need to change how a model behaves, and use both together in production. Models customized in SageMaker AI can even be imported back into Bedrock for serverless use.

Fine-tuning: Bedrock or SageMaker AI?

Sooner or later, someone asks whether to fine-tune a model, meaning train an existing model a little further on your own examples so it adopts your style, terms or format. Both services can do it, and the choice follows the same rent-or-build logic.

• Fine-tune in Bedrock when the model you want offers customization there and you want the result to stay serverless. You provide examples, Bedrock trains a private copy, and you use it through the same API.
• Customize in SageMaker AI when you need more control: open-weight models, techniques such as DPO or reinforcement learning, or your own evaluation. The serverless customization option keeps this from becoming an infrastructure project, and the result can be imported into Bedrock for serverless use.

Before either, try two cheaper things: better instructions in the prompt, and a Bedrock Knowledge Base with your documents. Most "the model doesn't know our business" problems are solved by giving it your information at question time, not by retraining it. Fine-tuning is for changing how a model answers, not what it knows.

Where Bedrock AgentCore and Hugging Face fit

Two more names come up in the same conversations. Bedrock AgentCore is Bedrock's service for running AI agents in production: programs that use a model to plan steps and call tools, such as looking up a repair status and then emailing the customer. It handles the runtime, memory, tool access and monitoring, and it can use models from Bedrock or from SageMaker AI endpoints. If your chatbot grows into something that takes actions, that is where it goes next; our AgentCore guide explains it from the beginning.

Hugging Face models, the huge public library of open models, can reach AWS through several doors: JumpStart, Bedrock Marketplace, or your own SageMaker endpoint. The same rule applies: if Bedrock offers the model serverlessly, that is usually the cheapest way to start, and dedicated endpoints are for when you need control.

How to delete a SageMaker endpoint properly

Because endpoints are the most common source of surprise bills, here is the exact clean-up, in the SageMaker AI console:

1. Open Inference → Endpoints, select the endpoint, and choose Delete. This is the step that stops the hourly charge.
2. Open Inference → Endpoint configurations and delete the matching configuration, so nobody recreates the endpoint by accident.
3. Open Inference → Models and delete the model entry if you no longer need it. The model files in S3 remain until you delete them there too.
4. Check the next day's bill under Billing → Bills, filtered to SageMaker, to confirm the hourly charge has stopped.

The bills that surprise people, and how to avoid them

• Endpoints left running. A JumpStart or SageMaker endpoint bills every hour until it is deleted. After testing, delete the endpoint (and its endpoint configuration) in the SageMaker AI console.
• Canvas left logged in. Log out from the Canvas menu when you finish; the session meter is $1.90 an hour.
• Studio spaces and notebook instances left running. Turn on idle shutdown, and stop spaces you are not using.
• Big Bedrock models for small jobs. Test with a small model first; many tasks do not need the largest one, and the price difference is large.
• No alarm on the bill. Set a budget alert before you experiment. Our billing-alerts guide takes ten minutes and would have caught Jake's weekend.

"I stopped everything, so why is my SageMaker bill not zero?"

This question arrives a week after almost everyone's first experiment, and it has a short list of answers. Stopping the expensive parts does not delete the cheap parts that hold your work:

• Studio storage. A SageMaker AI domain keeps each user's files on a storage volume (Amazon EFS for Studio home folders, and per-space storage for newer spaces). It costs a little per gigabyte per month for as long as the domain exists.
• S3 data and model files. Datasets you uploaded, training outputs and Canvas projects sit in S3 buckets and are billed for storage until you delete them.
• Logs. Training jobs and endpoints write to CloudWatch Logs, which keeps logs forever unless you set a retention period.
• Something still running. A notebook instance, a Studio space, an endpoint in another Region, or a Canvas session that was never logged out. Check each Region you used; the console only shows one Region at a time.

To find the exact source, open Billing → Bills, expand SageMaker (and S3, EFS and CloudWatch), and read the usage lines: they name the resource type and the Region. If you are completely done, deleting the domain (after its users, spaces and apps) and emptying the buckets brings the bill to zero. If you plan to come back, a few cents of storage is the price of keeping your work.

Regions: why a model is "not available"

Neither service offers everything everywhere. Bedrock models are offered Region by Region, and a new model often appears first in a few large Regions such as US East (N. Virginia) and US West (Oregon). Bedrock also offers cross-Region inference profiles, which route your request to a Region with capacity; they often carry a different price than single-Region use, and they matter if your data must stay in one country. SageMaker AI instance types vary by Region too, especially GPUs. When a tutorial's model or instance "doesn't exist" for you, check the Region selector in the top corner of the console first; it is the most common reason.

Canvas's built-in generative AI vs Bedrock

Canvas is not only for predictions. It also lets non-coders chat with foundation models, compare answers from several models, and ask questions about their own uploaded documents, all from the same point-and-click interface. Behind the scenes, those models are the same ones Bedrock offers (and some from JumpStart), billed per token, plus the Canvas session hour.

So which should a non-coder use? For trying models and asking questions about documents yourself, Canvas is friendly and keeps everything in one place. For putting AI into something other people use, such as a website chatbot or an automated email reply, you need Bedrock's API, or a developer to build on it. Jake uses Canvas to explore his repair data and Bedrock to power the website, and that split is a good default for any small business.

Learning these services: where to start

If you are learning AWS AI services for work or for a certification, the order that makes sense is the order of the decision above. First, Bedrock in the console's playground: try a few models on the same prompt and compare answers and prices. Second, Canvas with a public dataset, to understand what training a model actually means without writing code. Third, Studio and a JupyterLab notebook, once you want to see what Canvas does underneath. AWS's AI Practitioner and ML Engineer certifications follow roughly that path, and our free AWS series covers each service in the same plain-English style. Use the free trials, and set a billing alert first.

For IT admins: governance, data and cost controls

When teams start using these services, a few decisions up front prevent most later problems:

• Data handling. Bedrock does not use your prompts and responses to train its base models and does not share them with the model providers. SageMaker AI keeps your training data and models in your account. Keep sensitive data in your own S3 buckets with encryption, and use VPC endpoints to keep traffic private.
• Access. Control which teams can call which Bedrock models, and who can create SageMaker endpoints, with IAM policies; GPU endpoints in particular should not be something anyone can launch.
• Safety. Apply Bedrock Guardrails to customer-facing apps, so off-topic or unsafe answers are blocked consistently.
• Cost. Use AWS Budgets per team, tag endpoints and spaces by project, enable idle shutdown in Studio, and review running endpoints weekly. Savings Plans make sense only after usage is steady.
• Workspace. For cross-functional data and AI work, SageMaker Unified Studio with company sign-in gives one governed place instead of a scatter of consoles.

Bedrock vs SageMaker: frequently asked questions

Jake's two ideas ended up in two different places, which is exactly how it should be. The website chatbot runs on Bedrock for well under a dollar a month, and his repair-return prediction is a Canvas project he works on for an hour on Sunday evenings, logging out when he is done. The $68 weekend became the most useful lesson of the whole exercise. If AWS's AI names have made you feel lost, you were never meant to understand them on the first read; once you sort them into "rent a model" and "build a model," the maze turns into two clear doors.

Revision note. Written October 5, 2026. If AWS's AI names left you more confused than when you started, that was the naming, not you; may your next bill hold no surprises.