Key Takeaways
- Mike Cannon-Brookes calls out startups claiming they built proprietary models for $1 million, pointing out they merely fine-tuned someone else's billion-dollar foundation model.
- Foundation models improve fast enough across general benchmarks to erase temporary price-performance leads held by domain-specific models.
- Specialized models survive only under two conditions: non-commoditizable proprietary data or domains outside text and code, like robotics world models.
- Startups that spent equity capital training bespoke domain models two or three years ago have already watched generalized frontier models overtake their benchmarks.
The One-Million-Dollar Training Myth
Every enterprise pitch deck contains a slide claiming proprietary AI. Founders present custom weights, specialized pipelines, and fine-tuned checkpoints as an unbreachable technical moat. Mike Cannon-Brookes hears these claims constantly, and he refuses to buy the pitch.
“A lot of people who say they've trained their own model, I think they're being a little bit fast and loose with words as far as I'm concerned,” Cannon-Brookes told Harry Stebbings. “We could say we've trained a bunch of models at Atlassian. I say that's bullshit.”
He points to the simple math behind these claims. Training an actual state-of-the-art model requires clusters of thousands of GPUs, massive energy infrastructure, and hundreds of millions in capital. When an early-stage startup claims to train a custom enterprise model on a tiny budget, they are playing semantic games with their cap table.
“You see they train them for like a million bucks and I'm like there's no way you trained this for a million bucks,” Cannon-Brookes said. “You took someone's billion dollar model and tailored it if you want to think about it generically for a particular application.”
Taking an open-weight base, running low-rank adaptation on a few thousand domain examples, and calling the output a proprietary model is marketing spin. Worse, it misleads founders into believing they own an enduring asset when they have merely rented a snapshot in time.
The Foundation Steamroller
Fine-tuning creates a temporary price-performance bump. An adapted small model can run cheaper and faster on an isolated vertical task today. The trap is assuming that advantage lasts.
“The jury is really out on the benefits of building specialized models at the moment,” Cannon-Brookes explained. “These specialized models have to prove that they can stay ahead of the generified model for a given price performance ratio. I don't think I see many of them doing that.”
To survive, a specialized architecture must beat general foundation models on an exponential price-per-watt curve year after year. The market history of the past three years shows the opposite dynamic. General foundation models rapidly absorb vertical capabilities simply by scaling pre-training data, synthetic reasoning runs, and model architectures.
If your software company relies on language, structured reasoning, or standard code, the frontier labs will steamroll your tailored checkpoints within one release cycle. The only true exceptions are domains where foundation models lack training data, such as real-world physical dynamics in robotics, or companies with private, non-commoditizable customer data graph layers that nobody else can access.
What to Do With This
Audit your product roadmap by Friday and kill any internal task force dedicated to training or pre-training domain models. Move those engineers to building proprietary data capture and workflow integration, where user data creates actual switching costs. Test your core workflow against the newest general frontier model through standard APIs. If the frontier model matches your fine-tuned model's accuracy within 5%, deprecate your custom weights immediately.