Even Microsoft can't afford the token tax.

Last month, Microsoft got caught trying to swap the AI behind Copilot for a cheaper model called DeepSeek. Microsoft has put more than $13 billion into OpenAI. It built Copilot on OpenAI's models. It wants you to keep paying for them. And it's already looking for the exit.

The reason comes down to how you pay for AI. You pay per token, which is roughly a word the model reads or writes. A quick answer is cheap. But the "agents" everyone's building now don't answer once and stop. They remember the conversation, look things up, use tools, and work in steps. All of that is more words, and you pay for every one, every time. The bill only goes up.

The gap is enormous, and the cheaper models handle work most businesses can't tell apart. Microsoft ran the numbers on its own books and decided the premium wasn't worth it. So it's switching to a cheaper model and planning to charge you by usage instead.

That's cheaper for Microsoft. Now look at what it does for you. Nothing.

Microsoft found a cheaper landlord. You're still a tenant.

You're still renting. The meter still runs. Your files still sit on their servers, under their rules, and they can swap the model or change the price whenever they want. And the models that make all of this cheaper, DeepSeek, Qwen, Gemma, are free for anyone to download. You're paying a toll to rent something you could own.

That's the whole idea behind Vault Cloud. Instead of renting AI, you own it. We put a dedicated GPU in your hands and run those same open models on it, inside a private cloud only your company can touch. Your agents run all day. Your bill is the same every month, busy or quiet, because there's no meter. And your data never leaves your walls to train someone else's product.

Microsoft is shopping for a cheaper way to rent. We think the smarter move is to stop renting.

Sources: Microsoft's testing of self-hosted DeepSeek-V4 for Copilot Cowork and the shift to consumption-based pricing, as reported by Axios and Windows News.