liveBook · Manning

Vendor lock-in

As explained in Rearchitecting LLMs

Vendor lock-in is a condition in which dependence on a provider makes switching to an alternative prohibitively costly or difficult.

Vendor lock-in arises when dependence on a provider makes switching prohibitively expensive or technically difficult. Fine-tuning a closed model creates a hosted artifact tied to the provider’s base model: weights cannot be exported, and upgrades or retirement can force the workflow to be repeated. section 1.1.2

Common lock-in pressures include:

  • Provider changes can eliminate distinctions between standard and fine-tuned models. section 1.1.2
  • Pricing can change unpredictably, while discounts may depend on sharing training data. section 1.1.2
  • A base-model upgrade or retirement may require rerunning fine-tuning with the same dataset and evaluations. section 1.1.2
  • Open-source models can also create dependence when their oversized, benchmark-oriented capabilities do not fit domain-specific workloads. section 1.1.2

How Rearchitecting LLMs covers it

How deep it goes
Explains financial and technical switching costs through hosted weights, provider lifecycle changes, pricing, and repeated fine-tuning.
Examples
Uses fine-tuned closed models, provider upgrades or retirements, pricing discounts, and the leased-house analogy.
What it assumes
Assumes familiarity with model providers, fine-tuning, base models, datasets, and evaluations.
What you can do after
Operate/configure version pinning and repeat fine-tuning with preserved datasets and evaluations when providers change models.
How it connects
Connects vendor lock-in with fine-tuning, closed models, open-source models, foundation models, and model versioning.

Also explained in

Compare how 8 books explain vendor lock-in →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Rearchitecting LLMs. The section links point to the parts of the book it draws on.