"Build vs. buy" in AI is usually asked at the wrong altitude. Almost no enterprise should build models — and almost every enterprise must build something. The useful question is which layer you build at. Here's the framework that resolves it.
The Four Layers, and Who Should Build Each
Layer 1: Foundation Models — Buy (Almost Always)
Training frontier models is now a capital-markets activity: the labs raised $20–122 billion rounds in 2026 and sign compute contracts denominated in gigawatts. Unless model capability is your product and you can fund it like the labs do, buy. The exception isn't "we have data" — it's "we have data, distribution, and a decade of committed capital."
Layer 2: Fine-Tuned and Private Models — Mostly Buy, Selectively Build
The hardware floor for private deployment collapsed (competitive enterprise models now run on one or two GPUs), making hosting open weights a reasonable middle path for regulated data. Actual fine-tuning earns its keep only with a stable, high-volume, well-evaluated task where prompting demonstrably plateaus. Most teams that think they need fine-tuning need better retrieval.
Layer 3: The Integration Layer — Build (This Is Where You Live)
The model gateway, evaluation harnesses, retrieval over your proprietary data, agent orchestration wired into your workflows, and the governance controls around it all — this layer touches everything that makes your company itself, and no vendor sells it off the shelf. It's also deliberately vendor-neutral: built well, it makes Layer 1 vendors swappable, which is worth real negotiating leverage annually.
Layer 4: Applications — Buy the Generic, Build the Differentiating
Buy the commodity copilots (coding, meetings, documents) — the market is efficient and switching is easy. Build only applications that encode proprietary process knowledge, where the workflow itself is your competitive edge. The test: if a competitor bought the same tool, would it matter? If no, buy it.
The Total-Cost Math Everyone Gets Wrong
Buying is priced honestly (subscription + usage) but carries hidden lines: integration, vendor lifecycle churn (forced model migrations are now roughly annual per vendor), and lock-in premium at renewal. Building is underpriced habitually: teams count construction and forget operation — evaluation infrastructure, on-call, model-migration upkeep, and security review typically run 30–50% of build cost per year. A build that isn't funded for operation is a buy decision deferred eighteen months at premium prices.
Three Questions That Settle Most Cases
Is this capability differentiating or defensive? Defensive (everyone has it, you just need parity) → buy. Differentiating → build at Layer 3–4.
Does the workflow change faster than vendors ship? If your process evolves weekly, a vendor roadmap is a leash → build. Stable process → buy.
Can you staff the operation, not just the construction? No durable team → buy, whatever the spreadsheet says.
The 2026 Default Posture
The pattern among enterprises that are visibly winning: buy models, build the integration layer, buy commodity apps, build one or two differentiating applications — and revisit annually, because the buy side improves every quarter. Build-vs-buy in AI isn't a decision; it's an operating rhythm.
