Episode Details

Mistral's September 2026 raise of $3.5 billion made open-weight models a credible enterprise option, and most data leaders read that as a signal to stop buying and start building. The position here is the opposite: capability parity raises the ceiling, not the floor, and the hard part was never the model. It is the data plumbing and the retrieval layer, a point O'Reilly's 2026 adoption research supports.

You get the three conditions that justify self-hosting, the token volume crossover point DBT Labs published (with the caveat that they sell data tooling), and the MIT Sloan Management Review argument that open weights are worth having for optionality. You leave with a lock-in audit to run on your top three AI use cases.

Key takeaways

• Build with open weights only if you have proprietary data, an existing platform team that already operates services, and high predictable usage; missing one means don't build.
• Check your own inference bills against the DBT Labs crossover numbers before assuming self-hosting is cheaper than paying per call.
• Treat open-weight models as insurance and switching optionality, using them as leverage in vendor negotiations rather than running them yourself.
• Audit your top three AI use cases and ask whether you could swap the underlying model in under a month without a rewrite.
• Fix the data plumbing and retrieval layer first, since the model is rarely the bottleneck.

Chapters
0:00 Mistral's $3.5 billion raise in context
0:29 Open weights versus open source explained
0:56 The hidden cost of self-hosting models
2:07 Three conditions for building in-house
2:41 Token volume crossover: API versus owning hardware
3:19 Optionality as insurance and Monday's audit

Go deeper, free lessons

• Build vs buy: RAG vs fine-tuning
• CDO AI strategy: prioritization, build/buy & value chain
• Generative AI in the enterprise: RAG, risks & governance
• LLMOps & evaluation
• Measuring AI ROI

Full article and transcript: https://www.mba-training.com/blog/enterprise-generative-ai-build-vs-buy

MBA Training, mba-training.com