The Model Was Never the Bottleneck. The Data Is.

Billions are flowing into agentic AI infrastructure for finance. The part the coverage skips is the one that determines whether any of it actually works.

Last week, the largest financial institutions on the planet stopped using AI and started funding the infrastructure that runs it. JPMorgan and Anthropic announced a $1.5 billion joint venture. Rogo, the AI platform for investment banking, raised $160 million at a $400 million valuation. OpenAI is reported to be raising at a $340 billion valuation, with Goldman and Citigroup as likely participants.

The coverage was breathless. It was also incomplete.

You can fund any model you like. You can embed engineers in any workflow you like. You can build agents capable of drafting an IC memo in three minutes flat. But if the underlying data they are running on is not structured, sourced, and validated at the point of extraction — not after — the agents will be confidently wrong. And in financial services, confident and wrong is worse than slow and right.

Recent industry analysis puts enterprise AI fabrication rates above 50% when retrieval runs on ungoverned data. Most enterprise data is ungoverned. Most of the AI infrastructure being built right now will hit this problem within eighteen months of deployment — not because the models are bad, but because the data layer was never fixed.

Most data in global mining and natural resources is unstructured, fragmented across thousands of PDFs, scattered across eight languages, and inconsistently reported under three or four different regulatory regimes depending on jurisdiction and era. The models everyone is funding cannot fix that. Only the data layer can.

Anthropic and OpenAI are building horizontal AI infrastructure for the whole economy. Rogo is doing it for sell-side investment banking. The equivalent for mining and natural resources — a domain-specific, source-verified, multi-language data infrastructure that agents can actually rely on — is what we have spent two years building.

The model was never the bottleneck. The validated data layer is what separates a defensible workflow from a confident hallucination. In mining finance, that distinction is the product.

Less searching. More strategising.™