Why We Created Cloud World Model?

Simulating the cloud, without the cloud, so people and AI agents can design, test, and decide without paying for consumption.

Simulate the cloud without the cloud.

When we started taking Canvas Cloud AI to users, I expected the conversation to be about diagrams, about learning paths, about how to actually internalize cloud architecture. And we did talk about all of that. But underneath almost every conversation, there was a different thread that kept surfacing.

The cloud is too expensive. The cloud is too expensive. I heard it multiple times.

It came up from solo learners trying to practice. It came up from architects who couldn’t spin up the topology they actually wanted to test. The cost of learning the cloud and the cost of testing the cloud had quietly merged into the same problem.

That kept bothering me each time I thought about it. And the more I sat with it, the more I started to see a second pattern hiding behind the first.

The lock-in nobody talks about openly

On paper, customers have more choice than ever. AWS, GCP, Azure, OCI, a growing tier of specialty providers, the menu is long. But if you look at how things actually play out over a decade, most customers end up deeply committed to one vendor. Not because they evaluated the alternatives and chose decisively. Because evaluating the alternatives is itself expensive.

To truly compare providers, you have to build on each of them. To build on each of them, you have to pay each of them. So most teams don’t compare. They commit. And once they commit, the switching cost compounds quietly in the background until “we could move” becomes “we can’t really move.”

That’s the structural reason I think the major cloud providers have never publicly shipped a serious simulator of their own platforms. I might be wrong about the motive, I don’t know what’s discussed in their internal rooms but the incentive is hard to ignore. A great public simulator would make it easy to model “what would this look like on the other guy.” That is not a feature you build when your business depends on beating the other guy.

We come at it from the other direction. We want people to know their options. We want them to be able to design, compare, and decide before they spend a dollar. A cloud world model is, to me, a necessary step toward that kind of choice.

Simulate the cloud, without the cloud

So we built one. Cloud World Model lets you simulate AWS, GCP, Azure, OCI and DigitalOcean architectures and see how they actually behave, cost, performance, scaling, failures, without provisioning anything real.

The thesis is simple. You should be able to design and test a cloud architecture without paying for the cloud while you’re still figuring it out. Pricing should be a number you can model, not a number you discover on your invoice. Resilience should be a property you can stress-test, not something you find out about during an outage. Scaling decisions should be ones you can rehearse.

Take a customer I think about a lot. It’s May, and they know that in December they’re going to see roughly four times their normal workload. Today, the honest answer to “will the system hold?” is to run performance tests against the real environment, which itself increases consumption, and which still doesn’t really simulate December conditions, because you can’t safely simulate a four-times spike in production. So they guess. They overprovision. They under-provision. They find out in December.

That’s the world we’re trying to replace. A world where you can pre-play December in May. Where you can swap one provider for another in the model and see the cost and latency shift in seconds. Where you can inject a zone outage, watch your architecture fall over, fix it, and re-run, all without a single real resource being touched.

Why this matters even more for AI agents

The user feedback started this. But the second customer for a cloud world model is, increasingly, not a person at all.

AI agents are starting to make real decisions inside real enterprises. They are choosing instance shapes, deciding when to scale, picking between providers, refactoring topologies. And the way an agent gets good at these decisions is the same way a human does: by trying things and seeing what happens. The difference is that an agent can try thousands of variations a day.

If every one of those variations has to be tried against a real cloud account, the bill is unbounded and the blast radius is real. The only sane place for an agent to learn or to plan, or to compare is inside a model of the cloud that doesn’t charge for exploration.

That is the lane a cloud world model is uniquely built for. Decisions without consumption. Reasoning without invoices. We even support running it headless, so an agent can sit on top of the simulator and explore an architecture space without anyone watching the canvas. Actually, we hope over time that there is less usage of the actual canvas. The future of cloud engineering needs this layer. Humans need it to learn faster, and agents need it to act responsibly.

What a cloud world model actually is

If I had to define it cleanly:

A cloud world model is a simulation engine that mirrors how real cloud infrastructure behaves, cost, performance, scaling, and failures, so you can design, test, and optimize architectures without ever provisioning, or paying for, the real thing.

It’s the cost calculator, the performance test bed, the chaos experiment lab, and the multi-cloud comparator, fused into one model you can drive interactively or by API. It is not a replacement for production. It is the thing that should sit before production, so that what you eventually push to a real account is something you’ve already proven out in the model.

The tie back to Canvas Cloud AI

Cloud World Model is a Canvas Cloud AI product. The two fit together the way a textbook fits with a lab.

Canvas Cloud AI is where you learn cloud architecture visually concepts, patterns, the way the pieces compose. Cloud World Model is where you actually pick up the pieces and run them. Learners get a place to practice without a cloud bill. Teams get a place to compare designs without committing to a vendor. Agents get a place to train and plan without setting money on fire.

That’s why we built it. Not because the simulation was the goal, but because everything else we wanted, open learning, real choice, safe AI decision-making was blocked behind the same wall. The wall that says you have to pay the cloud to understand the cloud.

We don’t think that wall needs to be there anymore.

Try it. If you want to design, stress-test, or compare a cloud architecture without provisioning a thing, you can start in the Cloud World Model platform. Bring an idea you’ve been afraid to model because of the cost. That’s exactly the kind of thing it’s for.