KoBold Metals.

An office full of geologists writing Python.

On a Tuesday afternoon in Berkeley, a room of geoscientists is arguing about a rock they have never touched. The rock sits about 1,700 meters beneath the soil of Zambia's Copperbelt. The argument, projected on a screen, looks more like a Kaggle leaderboard than a mining map - a cloud of probability surfaces, drill-hole vectors, and Bayesian updates running over Jupyter notebooks. Down the hall, a model retrains. Across the floor, somebody fixes a bug in a query that pulls forty years of government drill records into a single schema.

This is what mineral exploration looks like at KoBold Metals. Less hammer, more histogram. Less luck, more linear algebra. The company calls its operating principle, with characteristic modesty, "Predict Everything."

Pull quote sourced from approximately every public talk Kurt House has given since 2019.

The world wants batteries. The Earth keeps its secrets.

Here is the awkward arithmetic of the energy transition. Every wind turbine, electric motor, and grid-scale battery is a stack of metals - copper for wiring, cobalt and nickel for cathodes, lithium for cells. The International Energy Agency estimates the world will need several times more of those metals by 2040 than it produces today.

The trouble is that the mining industry, for all its size, has been getting worse at finding them. Discovery rates per dollar of exploration spend have fallen for thirty years. The easy deposits, the ones that outcropped at the surface and announced themselves with green-stained rocks, were mostly found by 1980. What remains is deeper, more subtle, and harder to see. Exploration geologists call this the "tyranny of depth." It is also a tyranny of dispersion - the signals are weaker, the data noisier, the chance of a wrong bet higher.

Traditional exploration responds to this by spending more money to find less. KoBold's founders looked at the same problem and noticed something different. The industry was sitting on a century of geoscience data - drill logs, magnetic surveys, geochemistry, satellite imagery - and almost none of it was being treated as a dataset. It lived in filing cabinets, PDFs, and the heads of retiring geologists.

Three founders, one unfashionable idea.

The company was started in 2018 by Kurt House, Josh Goldman, and Jeff Jurinak. House, a Harvard-trained geochemist, had spent years thinking about how to apply Bayesian inference to the subsurface. Goldman came from energy investing. Jurinak brought reservoir engineering experience. They had worked together before, on a data-science platform for oil and gas, and had watched data-driven methods slowly transform that adjacent industry.

Their bet was unfashionable in two specific ways. First, that the most expensive, slowest, most risk-averse corner of heavy industry could be rebuilt around modern machine learning. Second, that you could not just sell software to mining companies to do it - you had to be a mining company yourself. KoBold, in other words, decided to use AI and also drill the holes.

A prediction engine for the subsurface.

The technical core of KoBold is what employees call the Machine Prospector, supported by an internal data system named TerraShed. The prospector ingests every kind of geoscience evidence the company can find - public surveys, proprietary partner data, drill cores, gravity and magnetic readings, satellite hyperspectral imagery - and learns where economic deposits are most likely to be hiding.

The output is not a yes-or-no map. It is a distribution. Each square kilometer of ground gets a probability, an uncertainty band, and a value-of-information score that tells the exploration team where a single new drill hole would most reduce ambiguity. It is, in effect, active learning applied to dirt.

The team then goes and tests the prediction. Drills are still drills. Helicopters still fly transects. Geologists still look at chips under a hand lens. The difference is that the model updates with every new data point, and the next decision - drill here, abandon here, move 80 meters east - is rarely made on intuition alone.

A chart in which the most boring bar is doing the most interesting work.

Customers, capital, and a very large hole in Zambia.

The fastest way to evaluate a company that promises to revolutionize an old industry is to ask who has actually paid them. KoBold's answer is unusual. Its "customers" are partners - BHP, Rio Tinto, Stanmore, and the Zambian state mining company ZCCM-IH - and they pay not in subscription dollars but in joint-venture economics: shared exploration costs, shared upside, sometimes a percentage of resulting mines.

The flagship is Mingomba. The deposit sits in Zambia's Copperbelt, the same geological province that produced Kakula in the neighboring DRC. KoBold's analysis identified the asset as undervalued and underdrilled. After drilling campaigns, the company described it as the largest copper discovery in Zambia in a century, with ore grades around 5 percent - extraordinarily high for a copper deposit. The project is now under construction at an estimated $2.3 billion in capital, targeting production in excess of 300,000 tonnes per year.

This is the kind of number that matters for the energy transition. 300,000 tonnes of copper per year is roughly the metal content of three million electric vehicles.

Statistics rendered in yellow because the alternative was Excel.

Mining, but with a different vocabulary.

It is fashionable, particularly among technology companies adjacent to extractive industry, to talk about doing mining "responsibly" and then to leave the sentence at that. KoBold's public position is slightly more specific. The company argues that better prediction is itself an environmental tool - that if you can find good deposits with fewer holes, less helicopter time, and smaller footprints, you reduce the surface impact of exploration, and you make mining viable in jurisdictions with stricter standards.

None of this is to say that mining at depth, even very well-modeled mining at depth, becomes simple. Mingomba will be one of the deepest copper mines on Earth. Water management, tailings, community relations, royalty negotiation, and the politics of resource sovereignty are all real, and all sit outside the reach of any model. But the bet is that fewer wasted holes upstream means a better-justified pit downstream. It is a defensible position, and a useful one to interrogate as the company scales.

The metals you'll never see, in everything you own.

If KoBold is right about exploration, the consequences are not just commercial. They are systemic. Every projection of decarbonization assumes the global supply of certain metals will grow several-fold in the next two decades. Where those metals come from, who finds them, and how quickly determines whether the transition arrives on time, on budget, and with any credibility about its environmental footprint.

It is also, frankly, the kind of work that does not produce a viral moment. Nobody will demo Mingomba on a stage. The output is a refrigerator's worth of copper in a car you will buy in 2032. The product, in the end, is invisible. That may turn out to be the most modern thing about the company - it is industrial in the most concrete sense, and it has no interest in pretending otherwise.

Back in the Berkeley room, the geoscientists finish their argument. Somebody commits a change. The model retrains overnight. Tomorrow the probability surface over the Copperbelt will look slightly different. A drill rig in Zambia will move 80 meters. The kobolds, after a few hundred years of running the place, may finally be losing the argument.