Industrial Foundation Models and the Return of Domain Expertise
General-purpose models plateau at the factory gate. The next tranche of value belongs to sector models trained on process data that exists only inside operating companies.
The plateau is real and it is specific
General models are remarkable at tasks whose knowledge is well represented on the public internet. They are markedly less useful the moment a task depends on how a particular kiln behaves at a particular humidity, or what a specific turbine sounds like three weeks before a bearing fails.
This is not a temporary weakness that scale will resolve. The information simply is not in the training corpus. It exists in maintenance logs, sensor histories, quality-control records and the working memory of experienced operators — none of which have ever been published.
Why this redistributes advantage
For most of the last decade, advantage in AI accrued to whoever had the most compute and the best research organisation. Industrial foundation models change the calculation, because the scarce input becomes proprietary process data.
That input is held by manufacturers, utilities, logistics operators and hospital systems — organisations that have generally been buyers of AI rather than builders of it. It gives them a genuine position, provided they recognise it before signing it away in a procurement contract.
The corollary is that data governance stops being a compliance function and becomes a strategic one. Who holds the rights to derived models, who can use the training data for other customers, and what happens at contract termination are now board-level questions.
What a sector model actually needs
The engineering requirements differ from general model work in ways that catch teams out. Industrial data is smaller, noisier, heavily imbalanced and frequently mislabelled — a fault class with forty examples is common. Physical constraints matter: a model that predicts an impossible state is worse than no model. And evaluation cannot be a benchmark score, because the cost of a false negative on a safety-critical prediction is not symmetric with a false positive.
The organisations succeeding here tend to do three things. They invest in labelling and instrumentation before modelling. They build in physical constraints rather than hoping the model infers them. And they define success as a decision improvement measured in the operation, not as an accuracy figure measured in a notebook.
The national dimension
For governments, industrial foundation models offer something a national language model rarely does: a direct line to industrial productivity. A shared model for a country's dominant export sector — trained on pooled, governed data from firms that individually lack the scale — is one of the more credible interventions available in industrial policy.
It requires a trusted intermediary to hold the data pool and a governance structure that competitors will accept. That is difficult institutional work, which is precisely why it is a legitimate role for the state.