An AI model now sits underneath a JORC-compliant mineral resource estimate at an operating gold mine. That’s the headline. Here’s what it really means — and what it doesn’t.
Stratum AI, a Toronto-based deep-tech outfit, spent four-plus years working with Perth’s St Barbara at the Simberi gold operation in Papua New Guinea. The result: St Barbara has incorporated Stratum’s deep-learning models — classifying weathering and material type (oxide, transitional, sulphide) — into the mineral resource estimate for its Simberi feasibility study. The estimate is reported under the JORC Code, which means it carries the disclosure requirements attached to an ASX listing. That’s not marketing fluff; a JORC-compliant resource has to survive scrutiny from geologists, auditors, and regulators who’ve seen every trick in the book.
St Barbara released the study on December 10, 2025 — the same day it announced a strategic transaction with Lingbao Gold Group valuing 100% of the New Simberi expansion at roughly A$800 million ($562 million). Take that timing as you will. Nothing quiets investor skepticism about AI-powered resource models like a half-billion-dollar transaction priced off the back of them.
The backstory matters more than the milestone. This didn’t start as a pitch deck. It started as pilots analyzing exploration data, then moved to operational data — determining which ore could run through the existing carbon-in-leach plant treating oxide material, and which needed to be stockpiled for future sulphide processing. That’s an unglamorous, practical application: sorting rock correctly before you crush it. Getting it wrong means feeding sulphides into a circuit built for oxides, chewing up the mill, and generating waste you paid to move for nothing.
Once the models were validated against grade-control reconciliation — the real test, where the model’s predictions get compared against what actually came out of the ground — St Barbara made them its preferred method for material-type classification. From there the models fed into the 2024 resource and reserve update and the 2025 feasibility study.
The numbers tell the story. In May 2024, St Barbara said the modelling identified an additional 143,000 ounces of treatable material, extending oxide processing life by 17 months. A month later, the company credited the same modelling with a 40% increase in proved and probable ore reserves. A 40% reserve bump without drilling a single new hole — that’s the value of re-reading data you already own.
And that’s the broader point worth sitting with. Mining companies generate enormous volumes of data across a mine’s life — drill core logs, assay results, grade-control samples, production reconciliations. Most of it gets filed away once the immediate decision is made. The marginal cost of re-analyzing that data with modern machine learning is trivial compared to the cost of a new drilling campaign. For any mine operator sitting on decades of legacy data, this is low-hanging fruit. It’s also, frankly, where AI in mining belongs — not in the pitch-deck fantasies of autonomous everything, but in the grinding work of classifying rock a little better than a tired geologist logging samples at 2 a.m.
Now, the skepticism. This isn’t the first time we’ve seen this circus — new technology rides to the rescue of a marginal asset, gets credited with miraculous reserves, and the promotional machine kicks into gear. I’ve watched geostatistics, then “big data,” then blockchain get sold to this industry with the same breathless framing. Stratum’s CEO calls this a significant milestone in AI adoption, and she’s not wrong, but remember who’s talking their own book.
The mitigating factor here is the JORC framework. A resource estimate that underpins a feasibility study and an A$800 million transaction isn’t a press release — it’s a legally accountable disclosure signed off by competent persons who stake their professional reputations on it. If the AI classification were garbage, it would surface in reconciliation within quarters. That’s the discipline that separates this from the usual tech hype cycle. The model has to eat its own cooking.
The risks, for anyone allocating capital here: AI classification models are only as good as their training data, and legacy geological data is notoriously inconsistent — different logging conventions, different assaying methods, decades of drift. A model trained on messy inputs can be confidently wrong, and “validated against reconciliation” is a claim that deserves scrutiny over multiple reporting periods, not one good update. Second, material-type classification is the easy version of this problem. Grade estimation — where the money actually is — is a harder, noisier task, and nobody’s claiming AI has cracked that yet at JORC standard.
Still, the direction is correct and, more importantly, economically boring in the best way. This is technology that reduces waste haulage, saves mill wear, extends oxide life by 17 months, and converts previously write-off material into reserves. That’s not a story — that’s margin. In a sector where all-in sustaining costs decide who survives the cycle, a few dollars per tonne of correctly classified ore compounds into real money.
Watch how quickly competitors copy this. When something works in mining, it gets cloned within two or three budget cycles. The durable question isn’t whether AI classification becomes standard — it will — it’s who owns the data pipelines and the reconciliation track record when it does. Stratum has a four-year head start and a feasibility study to point at. In this industry, that’s worth more than any amount of narrative.
