Foundation model
The pre-trained base every AI feature sits on. Switching it is harder than vendors admit and easier than internal politics suggests.
A foundation model is the large pre-trained neural network that everything else in your AI stack sits on top of. GPT, Claude, Gemini, Llama, Mistral — these are foundation models. Your fine-tunes, your retrieval pipelines, your agents, your prompts are all downstream of one.
The word “foundation” is precise. It’s the geological layer. Change it and what’s built on top has to be re-tested. Don’t change it and you eventually inherit its weaknesses.
Why it matters now
Foundation-model selection used to be a once-a-year procurement question. In 2026 it’s a quarterly one. New frontier models arrive every few months; each one resets price-per-token curves and capability ceilings. Decisions you locked in last quarter look defensible or expensive within ninety days.
The strategic question is no longer “which model is best?” but “how easy is it to switch?” — and the answer depends on choices your team made eighteen months ago. Tightly tuned prompts, vendor-specific function-calling syntax, and embedding spaces that only work with one provider’s encoder are the three quiet sources of lock-in.
The misuse
“We can’t change foundation models — our prompts are tuned for this one” is the common defense for inertia. It’s technically true and strategically wrong. Re-tuning prompts takes a week. The real lock-in is in the org chart: the team that picked the current model would rather not be told their pick is now mid-tier.
A good test: ask whether anyone on the team has run the same workload against the top three foundation models in the last quarter, with current pricing. If the answer is no, “we can’t switch” means “we haven’t tried.”