On Friday, a blog post by Tobi Knaup titled “Open-weight AI is having its Kubernetes moment” shot to the top of Hacker News, collecting 312 points and 255 comments by Saturday morning. The thesis was clean and flattering to its audience: just as Kubernetes won the container orchestration wars and became the boring, reliable plumbing of the internet, open-weight AI models are now winning the intelligence wars, becoming the default substrate on which the next decade of software will be built. The comments section filled with the usual hosannas — democratization, the death of vendor lock-in, the triumph of the commons.
The analogy is more revealing than Knaup perhaps intended. And not in the way the HN crowd thinks.
The Model Is Free. The Compute Is Not.
The open-weight models now leading the frontier are not small. Moonshot AI’s Kimi K2.6, released in April, crosses the trillion-parameter threshold — a scale that, until recently, was the exclusive province of proprietary systems behind closed APIs. Running inference on a model of that size at production scale is not something you do on a spare A100 in the office closet. Industry estimates put the monthly cloud compute bill for a serious deployment in the low six figures. That is before you factor in the engineering time to keep the thing from falling over.
Eiso Kant, founder of the AI coding startup Poolside, recently noted that his team spent $12 million building language models for code before the market cared — and that was for models a fraction of the size of today’s open-weight behemoths. The weights may be free to download, but the meter starts running the moment you try to do anything useful with them. And the meter is owned by Amazon, Microsoft, and Google.
The Orchestration Layer Is Where the Money Goes
Here the Kubernetes parallel cuts in a direction the celebrants are ignoring. Kubernetes won, yes. But the open-source project itself never made a dime. The money flowed to the cloud providers who wrapped it in managed services — Amazon EKS, Google GKE, Azure AKS — and charged a premium for the privilege of not having to run it yourself. The orchestration layer became the profit center.
The same dynamic is already unfolding in open-weight AI. Google’s newly released ATLAS report, an expansive survey of how organizations are actually using AI at work, is tellingly light on model training and heavy on deployment, integration, and management. The action is shifting from building models to operating them. The startups raising real money right now are not training foundation models; they are selling monitoring dashboards, fine-tuning platforms, and inference optimization tooling. The picks-and-shovels play is back, and the biggest shovel-sellers are the same three companies that won the cloud wars.
The Lock-In Nobody Talks About
A cloud architect at a mid-sized bank pinged me on Slack while the HN thread was still climbing the front page. “We’re not locked into a model vendor, sure,” he wrote. “But try migrating a fine-tuned 1T-parameter model from AWS to Azure without a team of six and a quarter-million-dollar budget. You don’t. You just pay the bill.”
He is describing the quiet reality behind the open-weight victory lap. The model weights are open. The operational knowledge — the fine-tuned adapters, the data pipelines, the prompt chains, the safety guardrails — is proprietary and deeply entangled with a specific cloud provider’s infrastructure. Switching costs are not zero; they are enormous. What you end up with is not vendor lock-in at the model layer, which the open-weight movement has genuinely solved. It is vendor lock-in at the infrastructure layer, which is far more lucrative and far harder to escape.
The Kubernetes moment for AI is real. But the lesson of Kubernetes is not that open-source wins. It is that open-source provides the raw material for a new kind of lock-in — one that is invisible to the engineers who celebrate it and immensely profitable for the platforms that own it. The weights are open. The wallets are not.