Chapter 11

How the frontier operates

Who is actually doing this, what they have demonstrated, and — importantly — how to tell the demonstrated part from the marketing. Then where our own products sit in the stack.

KoBold Metals

The most visible AI-native exploration company. Two internal systems are described publicly: TerraShed, a data platform that assembles heterogeneous historical and modern data — old drill logs, geophysical surveys, satellite imagery — into a queryable substrate; and Machine Prospector, the analysis layer that runs over it. The methods the company describes are ensemble machine learning, full-physics joint inversion (Ch. 5) and computer vision.

Their flagship project is Mingomba in Zambia, developed in partnership with ZCCM-IH, led by KoBold Africa CEO Mfikeyi Makayi, with 200+ Zambian employees and contractors. KoBold describes it as a "world class deposit," states it has contributed over $200 million to the Zambian economy, and calls itself the largest American investor in the country. It also funds MSc scholarships in data science and exploration geology with Stanford, Copperbelt University and the University of Zambia.

Note what is and is not published

KoBold's own Mingomba page states no grade or tonnage figures. Plenty of secondary coverage quotes specific numbers; treat those as journalism rather than disclosure until they appear in a technical report. The habit of checking what the primary source actually says — and noticing when it says less than the headline — is worth building now.

The strongest publicly documented technical link runs through the academic literature rather than the press. Intelligent Prospector v2.0 (Ch. 10) lists Kurt House (KoBold's CEO) and Damian Burch among its authors alongside Mern, Corso and Caers, tests on a sediment-hosted copper deposit, and states in its abstract that the algorithm "aided in the characterization of an ultra high-grade deposit in Zambia in 2023." That is a peer-reviewed claim of a decision-theoretic agent contributing to a real tier-one copper story — considerably more informative than any press release.

Stanford Mineral-X

The academic engine room, founded by Jef Caers in the Doerr School of Sustainability. The core thesis is a transplant: decision-making methodology developed over decades in oil and gas — uncertainty quantification, value of information, Bayesian optimal design — applied to mineral exploration, which historically has not used it.

The practical version of that thesis, and the sentence worth memorising from this whole primer: instead of drilling on a fixed grid to estimate grades, plan drilling to falsify human-generated geological hypotheses and strategically reduce uncertainty until the company can make an informed go-ahead or walk-away decision. Caers has argued this could reduce exploration drilling by roughly a factor of five.

Mineral-X focuses on critical minerals — copper, cobalt, nickel, lithium, rare earths — collaborates with industry (Ero Copper in Brazil, among others), and runs education programmes; in May 2025 Caers led a five-day course at Copperbelt University in Zambia billed as Africa's first programme on data science and AI for mineral exploration. Their publication list is the single best reading queue in this field.

How to read AI mining claims

You will encounter a great deal of enthusiasm. Some of it is earned. A short checklist for keeping your footing:

Claim patternWhat to ask
"AI discovered a deposit"Discovery is a multi-year effort by dozens of people. What specifically did the model contribute — target generation, hole placement, an inversion? Would a good geologist have gotten there anyway?
"Our model achieves 0.97 AUC"At what prevalence, and under what cross-validation? Random-split AUC on spatially autocorrelated data is close to meaningless (Ch. 9).
"Validated on known deposits"Retrospective validation on deposits already in the training distribution. Was there a temporal or region hold-out?
"Reduces drilling by N×"Against what baseline — a naive grid, or an experienced geologist's programme? In simulation or in the field?
"Processes petabytes of data"Volume is not skill. What decision changed?

And the structural difficulty behind all of it: with a discovery base rate near 10⁻³ and a decade-long feedback loop, a single success is not statistical evidence. The field cannot yet run the experiment that would settle whether AI-led exploration outperforms, and honest practitioners say so. What can be evaluated now is the intermediate quantity: information gained per dollar spent. That is exactly why Chapter 10's framing matters — it is measurable on a timescale shorter than a career.

The commercial stack

For orientation, the software a working exploration team actually touches. Know the names; you will be asked to integrate with them, not replace them.

LayerTypical tools
Drillhole databaseacQuire, DataShed, MX Deposit, and an enormous quantity of Excel
3-D geological modellingLeapfrog Geo (Seequent), Micromine, Datamine, Vulcan
Resource estimationIsatis.neo, Supervisor, Datamine, Vulcan
Geophysical inversionOasis montaj (Seequent), VOXI, UBC-GIF codes, SimPEG (open source)
GISArcGIS, QGIS
The integration lesson

Exploration teams will not migrate off this stack for a new tool, and asking them to is the fastest way to lose a deal. Anything you build should attach to what they already run — read-only, no migration, minimal configuration. That constraint is not a limitation to work around; it is the product requirement.

Where Aethermine sits

Our thesis is that general foundation models are not spatially competent: they cannot ingest the volume or the geometry of exploration data, and they reason about a drill hole as text rather than as a trajectory through a 3-D field. Geo-spatially aware models — models that hold real spatial structure and can cite the record they are reasoning over — are a different product category. Four products, mapped onto the chapters you have just read:

Coreroom

The AI terminal for your projects

Agentic interface over the whole project record — the retrieval, memory and tool layer that makes Ch. 4–3 data conversational and auditable.

Audit

Automatic QA/QC for drilling data

Chapter 6 as a product: connector-first, read-only, continuous deterministic checks with cited evidence. Flags and explains; a QP decides.

Data Synergy

Custom predictive & probabilistic models

Chapters 8–10: heat maps, prospectivity, uncertainty quantification and decision support built against a specific project's data.

HyperSpec

Multispectral & hyperspectral

Chapter 5's spectral layer: tasking and processing imagery, paired with context-aware agents that tie mineralogy back to the project record.

Aethermine Solutions wraps these as consulting and technical projects for teams that want the outcome rather than the software.

The ordering matters and it is deliberate. Audit sits underneath the rest because — per Chapter 6 — a prospectivity model, a resource estimate and a drill planner are all downstream of whether the assay database means what it says. Fixing the substrate is less exciting than the planner in Chapter 10 and strictly more valuable, because it is the precondition for the planner being worth running at all.

Sources for this chapter

Company claims in this chapter are attributed to their public sources and linked in Further reading. Where a figure appears only in secondary coverage it is identified as such.