Discovery rates collapsed tenfold. KoBold still found a 5% copper deposit

On No Priors, Sarah Guo and Elad Gil introduce Josh Goldman as co-founder of KoBold Metals, a company building what they call the world’s largest geoscience data collection and using AI to find lithium, copper, and the other metals behind batteries and AI infrastructure. The hosts note that KoBold invests more than $100 million a year across roughly 70 projects on four continents. Goldman’s first correction is about the business itself: KoBold explores. It combines AI with human intelligence to become a better finder of the mineral sources those industries will need.

Asked whether the company also mines, Goldman calls that the central question and draws a hard line. Exploration is where far more value can be created, and it is where technology can actually differentiate. A few million dollars of capital, he says, can produce a hundred- to a thousand-times return if you get the claims right — the same logic nineteenth-century gold miners understood. The catch is that the success rate is terrible and the problem keeps getting harder, which is exactly why a technical edge matters.

It used to be possible to find minerals with your eyeballs. Many copper ore minerals that form at the surface are altered by air and water into blue and green, like the patina on the Statue of Liberty. Goldman’s punchline is that anything you can find by traipsing across the ground with your eyes has already been found. What remains is concealed underground by rock. Technology is how you look there. Once a discovery is made, the work sits on a continuum from rock samples to defining a continuous body, designing a mine, and building it, and KoBold stays involved as projects move along that spectrum.

Humans have been collecting data about rocks for as long as they have been looking at them, and a great deal of it is already in the public domain. At the planetary scale, Goldman points to satellite imagery in different wavelengths and datasets on ancient continental collisions that built mountain ranges. Zoom in, and airborne surveys measure magnetism, density, and electrical conductivity; on the ground, rock and soil samples give elemental concentrations. All of that is standard. The catch is that even structured data is scattered across tens of thousands of repositories, with messy quality problems, and much of the corpus is unstructured: a rich geological lexicon, company reports filed with regulators, and maps of every kind.

One of Goldman’s favorite datasets is a set of Zambian maps from almost a hundred years ago, the originals hand-painted on linen. An elderly geologist tipped KoBold off to the right drawer in the state archives. You could never collect that data again: the work was labor-intensive, and the land now holds farms and people. The observations were made by skilled geologists, and “the rocks haven’t moved,” so there is no expiration date. Those maps become ground truth for training machine-learning models on modern airborne geophysics and satellite imagery. The power, Goldman says, is the systematic combination of structured and unstructured data, old and new.

That is the setup for how you look for lithium — or anything else. There is an enormous amount of lithium in the crust, including in the granite outside your window and the soil in your driveway, but it sits at a few tens of parts per million and is not economical to extract. An ore deposit is a place where geological processes gathered metal from a huge volume of rock, moved it, and redeposited it at something more like 1% copper or 1% lithium, after which industry takes it the rest of the way to 100%. Those concentration events are rare in the Earth as a whole, yet they have happened in many places. Goldman’s formulation is that exploration is an information problem: the scarce resource is not the metal, and not even the deposits, but information about where they are. You learn the recipe, guess where those processes left a signature in the data, narrow the map, and then you go acquire the land.

One of the hosts asks whether the real constraint is geology or regulation — the sense that the United States has deposits but no longer wants to mine certain places. Goldman agrees that regulatory constraints matter, then warns against filtering too narrowly at the start. Good deposits are rare enough that you want the best prior you can get and the highest probability of success. It is not only a function of permits. KoBold also looks at security of property rights: if they find something, they must be able to develop a mine that produces for decades, or sell it to someone who will, which means they have to trust that ownership, tax rates, and royalty rates will hold.

Development is harder still because regulators are only one constituency. Community interests are extremely local. The United States is not monolithic: state regulators sit alongside many communities and indigenous groups. The same is true worldwide. In Zambia there are some 50 chiefdoms, and traditional leaders exist everywhere KoBold works. Technical success, Goldman says, is not very helpful. Success is finding something economic to develop, either for KoBold or for a buyer. Without the so-called social license to operate — without investing in community relationships, and without starting in a place where that is possible — the company will not succeed.

KoBold does two things: it finds prospective ground, then it tests hypotheses by sampling rocks, flying surveys, and drilling, while building technology to guide those decisions. The exploration portfolio is more than 60 projects on four continents — North America, Europe, Australia, and, critically, Africa — targeting copper, lithium, nickel, and cobalt. KoBold owns the exploration rights itself or with a joint-venture partner and operates the programs. Almost all of them are pre-discovery seeds. In Zambia, though, the company has Mingomba, which Goldman calls the highest-grade large copper deposit that is not yet a mine.

The average operating copper mine today runs about 0.6% copper: a thousand kilograms of ore yields six kilograms of metal. The core of Mingomba is over 5% copper, and it is very large. Compare a 5% deposit with a 0.5% deposit producing the same copper, and revenue is the same, but ten times the grade means ten times less rock to haul, ten times less waste, and a plant an order of magnitude smaller. Capital intensity falls, operating costs fall, and so does the environmental footprint. Copper is a global commodity; everyone sells at the same price, so money is made by sitting on the low-cost, low-capital-intensity part of the curve. That, Goldman says, is the definition of the exploration problem: find the highest-quality assets.

The technology is a full-stack system of dozens of products on three themes. First, sensors: hardware KoBold has built to collect new kinds of Earth data. Second, a data system that puts proprietary measurements, historic structured files, and a huge unstructured corpus in one place so the team can query the whole thing at once — LLMs help here — instead of hunting and pecking. Third, dozens of models, operating at many length scales, that predict where and how to look: satellite imagery, proprietary hyperspectral airborne imagery, and rock samples used to forecast surface rock types and properties. The interesting loop is active learning. Field teams do not go where the model is already confident, because they would learn nothing. They go where it is highly uncertain. A small amount of new ground truth can dramatically improve predictive power, and models are retrained every day, then served back to the geologists — a duet of technologists and field scientists.

The predictions are three-dimensional and probabilistic. One example: pegmatites, the container rocks for lithium, predicted at a surface location, then sampled on landing. Another: a conductive layer between 200 and 300 meters, prospective for nickel, copper, and cobalt, with a distribution over sulfur and metal content, plus the best collar location and drill direction. A third: a known layer continuing in a given direction, with depth, thickness, and copper grade drawn as a probability distribution. After each new observation, the model is conditioned and a fresh prediction goes out. On sensors, KoBold uses whatever service providers offer, but mining companies adopt new data types slowly, so the team built its own hyperspectral imaging system in less than a year, flying a light aircraft and collecting 600 colors at much lower acquisition cost and faster processing. Asked whether any single dataset unlocked Mingomba, Goldman says no. The industry keeps inventing supposed silver bullets — airborne gravity gradiometry in the 1990s was going to find all the ore deposits — and they never do. Data is high-dimensional; adding dimensions helps. New analytical methods, quantified uncertainty, and scientific insight about how ore systems form all have to work together. “There is no way to isolate the AI from the HI.”

Elad Gil walks through a simple cash-flow picture — commodity price, grade, size, operating cost — and Goldman says that is exactly right, and actually easier than software. Natural-resource assets trade on the present value of future production, which is knowable. It is much easier, he argues, to know what a mine will produce 20 years from now than to know a SaaS company’s sales volume and pricing that far out. Gil jokes that he feels attacked. Goldman’s worked example is a plant that moves 10 million tons of ore a year, highest grade first: revenue is price times volume; costs are trucks, water, and power; capital is crushers, mills, and steel balls. The sensitive uncertainty is recovery — 90% of the copper versus 92% or 88% — then you discount for stage and jurisdiction. KoBold can say with high confidence what Mingomba’s sales volume will be in 20 years. Mines are often underwritten on the first 20 years and then run 50, 60, or 70, because going underground teaches you more geology and the resource keeps extending.

Gil then asks the hit-rate question: if you sample a hundred sites, how many discoveries do you get, and how much better can KoBold be? Goldman says industry success has gotten 10x worse in 30 years, because the problem got harder and the industry is slow to innovate. Counting attempts is ambiguous — some studies quote a half-percent success rate — so he prefers to think in capital. Put a billion dollars of exploration spend through the industry and ask how many successes you get. Thirty years ago that was about eight discoveries. Today it is fewer than one high-quality economic deposit, which is why exploration in the aggregate is not a great business. KoBold’s target is $50 million to $100 million per discovery. Mingomba counts as a success. Now they have to do it again and again.

Gil has heard that Western buyers of mining assets are capital-constrained, in part because of ESG. Goldman does not buy the scarcity-of-capital story. The real scarcity is good-quality ore deposits. Great projects, whoever owns them, do not have trouble getting funded; they have lots of suitors. There just are not very many of them, which is why KoBold’s job is to find more. Where to look depends on the commodity. Copper has been a target for a long time in South America and Central Africa, yet parts of those regions remain underexplored. Mingomba sits in a deeper part of a Zambian basin with no surface expression. Congo has had challenges, Goldman says, but the exploration potential remains great. Lithium is more extreme: much of the world is underexplored because lithium was not a primary target until batteries for EVs, drones, and other large devices. Many of the big hard-rock lithium mines in production today were found in the 1980s by people looking for tantalum for capacitors. The science of how lithium deposits form is still incipient, which Goldman finds exciting — a little more scientific understanding can be a potent differentiator.

Asked which commodities are overstated as scarce, Goldman puts rare earths at the top of the list. A lot of the noise, he says, is because the name contains the word rare. Lithium, copper, nickel, and cobalt are not rare-earth elements. The actual rare earths that matter for electric motors are materials such as neodymium and dysprosium, used in permanent magnets. The reason they stay in the news, besides the branding, is a concentration of downstream processing in China, spurred by incentives, not only for rare earths but for lithium and now copper smelters as well. That makes it hard for anyone else to underwrite a processing plant: if a Chinese buyer will pay more for feedstock and accept a thinner margin, private capital elsewhere struggles without guarantees or subsidies, which KoBold does not rely on. The second effect is that raw materials flow to China, downstream manufacturing clusters there, and that becomes a structural strength of Chinese industry. On the other side of the ledger, Goldman is not currently chasing some other industrial metal he thinks cannot meet demand. The company’s list already has strong demand tailwinds and market depth, and the goal is to be the low-cost producer rather than to have a precise price view. Gil jokes about buying Greenland; Goldman will go wherever the algorithms and the first rock samples say extraction will be efficient.

The hosts recall an earlier conversation in which Goldman talked about how important philosophy was to how the company operates. He does not treat that as a flourish. KoBold, he says, is an epistemic project. The business is making better predictions: form a hypothesis, deploy capital and time to test it, and succeed or fail based on prediction quality. Models forecast what rocks are at the surface and below it — density, copper, nickel, and the rest. With sparse data, many possible geologies are consistent with the same surface or airborne observations. Standard industry practice is to pick one “best” model, because you cannot keep 10,000 inconsistent hypotheses in your head. Goldman says that is exactly what they have to do. The job is to embrace uncertainty and reduce it judiciously. Data is useful inasmuch as it reduces uncertainty.

That view is written into a company document called KoBold’s epistemology of exploration, a short list of core ideas. Predictions must be definite enough to be falsifiable: before you collect the data, you go on record about what observation would cause you to abandon the hypothesis. That is how they fight confirmation bias, which this business is extremely susceptible to — you have an idea, collect some data, modify the story to accommodate it, and justify spending more time and money. You also have to work with multiple alternative hypotheses, not one, and the point of data collection is to distinguish among them. At least one of those hypotheses has to be economically relevant. “We are a business, not a science project.” Oddly, they have a chief philosopher, the epistemologist Michael Strevens, author of *The Knowledge Machine*, a book about what science is and how it differs from other ways of knowing. The philosophy guides both field practice and technology: many of the tools exist to quantify uncertainty and then ask which measurement would cut it most effectively.

Goldman’s path does not look like a mining career. He studied physics to grapple with hard questions, did a PhD in quantum computing, then worked as a management consultant to power companies, oil and gas firms, and the industrials that make power and oil-field equipment. Co-founder Kurt House — physics and philosophy as an undergraduate, applied math and earth sciences as a graduate student — had become a friend in graduate school at Harvard. The two later did private-equity oil and gas investing together at a firm whose leaders had sponsored House’s previous startup. They read energy papers and visited power plants and coal mines. In 2018 they decided they no longer wanted to work on fossil fuels.

From first principles they asked what raw materials the future economy would need. Look around a desk or a house, Goldman says, and every product ultimately came from agriculture or from rocks. The big trends were batteries and AI. Cars, trucks, drones, aircraft, and robots that need long range and durability require lithium, which fuel-burning vehicles do not use at all. The data-center buildout, and the electricity to power it, requires an enormous amount of copper. The scale is gigantic: to build a future powered by batteries and AI by mid-century, the world will need to mine, over the next 25 years, more copper than has been mined in all of human history, and a tenfold increase in lithium production relative to today. Those materials have to be found, at a moment when the problem is getting much harder and innovation has been slowing. For Goldman, that is the perfect application — a differentiated technical business that also does something important for society, which is why people join KoBold.