Avery Pennarun, co-founder and CEO of Tailscale.
Every few weeks, someone announces a cheaper way to run AI: better chips, smaller and faster models, new cloud providers offering better prices for the GPUs and infrastructure that power AI applications. This is how technology is supposed to work: things get faster, costs come down, and customers benefit.
At least, that’s the theory.
That cheaper compute can come with a catch: if your AI locks you into one provider’s models, storage, networking, and other services, moving to another LLM might mean rebuilding half your infrastructure.
The computing industry has somehow created an AI infrastructure market where the underlying technology keeps getting cheaper, but buyers never see the benefit. If your supplier knows you can’t easily leave, there’s not much incentive to offer you a better deal.
A very expensive problem
Character.ai’s bankruptcy last month offers a particularly painful example of the economics of AI.
“Knowing where your computers are doesn’t tell you who controls them.”
The company let people create AI characters and have conversations with them. Millions of people loved using it, often for hours a day. Unfortunately, those conversations cost money to generate, whether the person having those conversations pays or not.
According to its bankruptcy filings, Character.ai was spending roughly $10 million USD ($14 million CAD) a month on compute in 2025. That’s an excellent business model for whoever is selling the computers.
The team clearly understood the problem and worked hard to solve it. They stopped developing their own foundation models, adopted open-weight alternatives, and moved some workloads to cheaper infrastructure. In one deployment, they doubled inference throughput. They substantially reduced costs, but couldn’t get the overall economics to work.
I have a lot of sympathy for the engineers involved. They were doing exactly what good engineers should do: finding better technology, improving efficiency, and trying to make the system sustainable.
Character.ai managed to move some workloads to cheaper infrastructure, but whether you can take advantage of a better price depends on how much of your application you’ve built around your existing provider. Changing computers is one thing. Changing everything connected to those computers is another.
That’s the problem the cloud industry has spent the last twenty years making worse.
The easy way in is the hard way out
I’ve been working with computers long enough to recognize this business model. Back in the mainframe days, one company supplied the computers, operating system, and software. Customers got a system that worked, but changing suppliers meant replacing much of it.
Cloud computing offered a different approach to the same lock-in. AWS started with S3, its data storage service, and EC2, its cloud compute rental offering. Amazon offered the same basic building blocks companies already needed, without buying servers or maintaining a data centre.
But the problem with selling storage and compute is that someone else can sell them too.
RELATED: Defence tech founder warns Canada is getting “wishy-washy” about defining sovereignty
So AWS added more services. Lambda runs code without servers to manage. DynamoDB takes care of your database. EKS manages Kubernetes. These services save engineering time and money. They also give customers more reasons to lock in. Lambda and DynamoDB have their own APIs. Even Kubernetes, designed to be portable, can become tied to AWS networking, storage, and identity systems.
When you’re building an application, using those services makes sense. Why spend six months building something your cloud provider can give you today?
Then, a few years later, another provider offers you compute or storage for a fraction of the price. Great. Except your application now depends on half a dozen services that the cheaper provider doesn’t offer. Moving means replacing those services, rewriting integrations, and testing everything again.
AI is following the same path.
Providers are bundling models, inference, storage, networking, and management into integrated platforms. You start with a model, add the provider’s tools to manage it, connect your data, and build your application around the whole package.
Then someone releases a better model, or offers compute at half the price. You’d like to switch, but the rest of your application depends on services that don’t exist outside your current provider.
The computers have changed, the software has changed, but we’re rebuilding the same dependencies that made mainframes so profitable. Essentially, we’ve made it back to 1970, with better autocomplete.
Sovereignty should include an exit
You’ve probably heard the term “sovereign AI” thrown around a lot lately, usually in discussions about keeping data and computing within national borders. That’s important, but I think it’s a rather narrow definition of sovereignty.
Knowing where your computers are doesn’t tell you who controls them.
Sovereignty means control over the systems you depend on; choosing where your data lives, which models you use, where your software runs, who you trust to operate it, and being able to change your mind later.
A Canadian company has good reasons to keep its data in Canada. Another might need local data but want to use a model hosted elsewhere. Both should be able to make those choices without being forced into an entire vendor’s technology stack.
A server inside the border is useful, but vendor lock-in, now with a Canadian postal code, isn’t a sovereignty strategy.
“A server inside the border is useful, but vendor lock-in, now with a Canadian postal code, isn’t a sovereignty strategy.”
This doesn’t mean everyone needs to build a data centre or train their own model. Managed cloud services are often the right answer. I wouldn’t recommend buying a rack of GPUs just to prove a point.
What we need is infrastructure that lets the pieces work together without requiring everyone to buy a single package.
Open models help. So do standard interfaces, portable data, and networking and access controls that work across providers. You should be able to replace a model without rebuilding the network, or move a workload without redesigning how it connects to everything else.
There will still be work involved. Different models behave differently. Hardware has different performance characteristics. Moving data takes time and money. None of this is magic. But those should be engineering decisions, made on their merits.
Character.ai’s engineers cut costs substantially and the company still couldn’t make the business work. Portability alone couldn’t guarantee a different outcome. But its experience shows why the next generation of AI companies needs every opportunity to take advantage of better technology and lower costs.
RELATED: Canada should stop acting like the waiting room for Delaware
AI infrastructure is still being invented. We can build it as a collection of parts that work together, where companies can change models, move compute, keep their data where they need it, and maintain the same networking and security controls. Or we can let the big players build tightly bundled systems that are convenient to adopt and expensive to leave.
When customers can switch providers, suppliers have to keep competing on price, performance, and quality. When they can’t, the benefits of technological progress don’t necessarily reach the companies building on top of it.
We have a chance to build an AI ecosystem in which companies can choose the best technology for their needs, wherever it comes from, and change those decisions as technology improves. That would give the next generation of builders a much better chance of succeeding.
A market that only lets you choose your infrastructure once isn’t much of a market. And it certainly isn’t sovereignty.
The opinions and analysis expressed in the above article are those of its author, and do not necessarily reflect the position of BetaKit or its editorial staff. It has been edited for clarity, length, and style.
Feature image courtesy Web Summit Vancouver. Photo by Sam Barnes via Sportsfile.