Predicting the Earth with Josh Goldman: How KoBold Uses AI to Find Critical Minerals
Summary
Mineral exploration offers venture-scale payoff—a few million dollars can create a 100–1,000x return—but industry productivity has fallen roughly 10x in 30 years, from about eight quality discoveries per $1 billion invested to less than one. Josh Goldman says KoBold targets $50–100 million per discovery, treating mineral supply as an information problem: although deposits occur in many places, they are rare in the Earth as a whole, and “the scarce resource is the information about where the ore deposits are located.”
KoBold’s Mingomba deposit in Zambia demonstrates how ore grade can transform both economics and environmental impact. Its core exceeds 5% copper versus roughly 0.6% across operating mines; compared with a 0.5% deposit, 5% grade can mean “ten times less stuff to haul out of the ground,” a smaller plant, less waste and lower capital intensity.
The data advantage comes from assembling irreplaceable but fragmented evidence, not possessing one magical data set. Tens of thousands of public repositories include satellite imagery, geophysics, samples, regulatory filings and nearly century-old Zambian maps hand-painted on linen: “The rocks haven’t moved. So there’s no expiration date on the data.”
KoBold combines sensors, a unified data system and dozens of models in an active-learning loop that deliberately investigates uncertainty. Field teams collect ground truth where models are least certain, retrain them daily and receive updated predictions; a proprietary airborne system captures 600 colors. Goldman rejects the “silver bullet” narrative: “There is no way to isolate the AI from the HI.”
High-quality deposits, rather than financing, are the binding constraint on new mining supply. Goldman pushes back on Elad Gil’s ESG-capital hypothesis: “Great projects don’t have problems getting funded,” while insecure property rights, inconsistent tax or royalty terms, or a missing social license can still prevent a technically successful discovery from becoming a successful mine.
Rare earths are “not that rare”; the strategic issue is China’s concentration of refining and downstream capacity. Chinese processors can accept lower margins when competing for feedstock, making new private facilities difficult to finance without guarantees or subsidies and pulling subsequent manufacturing stages toward the same geography.
KoBold institutionalizes falsifiability and competing hypotheses because sparse geological evidence permits many possible underground worlds. Its “Epistemology of Exploration” requires disciplined uncertainty reduction. Goldman separately describes a demand case of historic scale: by mid-century, humanity will need, over the next 25 years, to mine more copper than has been mined in all human history and increase lithium production roughly tenfold relative to today.
Deep dive
1. Exploration’s scarce input is information, not metal
Goldman defines KoBold as an exploration-first company because that is where technology differentiates and “way more value” can be created. A few million dollars might produce a 100–1,000x return, but only if the company improves a success rate made punishing by increasingly concealed deposits.
Copper or lithium may occur at only tens of parts per million across ordinary crust. Ore forms where geological processes gather material from huge rock volumes and redeposit it at perhaps 1% or more. There are many such deposits, even though they are rare in the Earth as a whole; hence Goldman’s inversion: reliable information about their locations is the scarce resource.
Elad’s pushback—could regulation, rather than geology, explain apparent US scarcity?—draws a qualified agreement. KoBold considers permitting, property rights and whether tax and royalty terms can remain dependable for decades, yet cannot filter too narrowly before starting with “the best prior” for geological success.
Development is ultimately hyper-local. Goldman points to state regulators, individual communities and Indigenous groups in the US, plus roughly 50 Zambian chiefdoms: “Technical success is not very helpful” unless a deposit is economic and the operator has invested in the relationships required for a social license.
2. KoBold compounds messy data through active learning
Exploration moves across scales: continental collisions and satellite imagery narrow the map; airborne sensors measure magnetism, density and conductivity; field teams collect and measure rock and soil samples. Much of this evidence is public, but scattered across tens of thousands of repositories, with structured and unstructured data requiring scientific judgment about what it means and whether it is fit for purpose.
Goldman’s favorite specimen is a set of nearly century-old Zambian maps, their originals hand-painted on linen and located after an elderly geologist identified the correct archive drawer. Those observations cannot be recreated across today’s farms and settlements, yet provide ground truth because “the rocks haven’t moved.”
KoBold’s full stack has three layers: proprietary sensors, one system spanning structured and unstructured evidence, and dozens of predictive models. LLMs help interrogate the corpus, while models predict surface rock types or a conductive underground layer’s depth, thickness and possible nickel, copper, cobalt and sulfur content.
Teams do not merely visit high-confidence locations. They sample where uncertainty is greatest, retrain models daily and send revised predictions back into the field—“this duet” of geologists and technologists. KoBold also built a light-aircraft hyperspectral system in less than a year, capturing 600 colors more cheaply and quickly than available services.
3. Mingomba shows why grade governs margin and footprint
The show introduces KoBold as investing more than $100 million annually across 70 projects; Goldman later describes a portfolio of more than 60 across North America, Europe, Australia and Africa. Most are pre-discovery “seeds,” targeting copper, lithium, nickel and cobalt under owned or joint-venture exploration rights.
Mingomba’s core exceeds 5% copper, versus about 0.6% for operating mines. At 5% rather than 0.5%, equal copper output requires at least ten times less ore movement, waste and plant size—lowering capital and operating costs while shrinking the footprint in a commodity market where every producer receives the same copper price.
Goldman calls mine valuation unusually straightforward: discount future production using knowable throughput, grade, commodity price, capital needs and operating costs. Recovery sensitivity might be 90%, with 92% “juicier” and 88% dilutive; projects can be underwritten on 20 years even though resource extensions may sustain operations for 50–70 years.
4. Good deposits—not capital—are the industry bottleneck
Exploration performance has deteriorated 10x in 30 years. Goldman’s preferred denominator is invested capital: $1 billion once yielded roughly eight high-quality discoveries amid hundreds of failures; today it yields fewer than one. KoBold’s objective is one discovery per $50–100 million, with Mingomba the first proof point to repeat.
Elad asks whether ESG pressure has deprived Western mine buyers of capital. Goldman rejects that diagnosis: “Great projects don’t have problems getting funded whoever owns them.” Attractive deposits draw many prospective buyers; the genuine scarcity is a pipeline of high-grade, low-cost, developable assets.
Copper’s established regions still contain underexplored areas, including deeper Zambian basins without surface expression. The big hard-rock lithium deposits in production today were found while prospectors sought tantalum for capacitors in the 1980s, and Goldman calls the science of lithium-deposit formation “incipient”—making modest insight potentially highly differentiating.
Rare earth anxiety confuses geology with industrial concentration. Neodymium and dysprosium matter for permanent magnets, but China’s processing build-out is the strategic issue: processors willing to accept thinner margins can outbid new refiners for feedstock, deterring unsubsidized entrants and concentrating the downstream manufacturing chain around landed raw materials.
5. KoBold makes disciplined doubt an operating system
Goldman describes the company as “kind of an epistemic project”: sparse observations permit many underground geologies, yet standard practice chooses one best model because working with 10,000 inconsistent models is difficult to manage. KoBold instead treats data as useful only insofar as it judiciously reduces uncertainty among those possibilities.
Its internal “Epistemology of Exploration” requires definite, falsifiable predictions recorded before new evidence arrives: what observation would force the team to abandon a hypothesis? That discipline counters confirmation bias—the temptation to modify a hypothesis to accommodate new data and justify more spending.
Teams must maintain multiple alternative hypotheses, with data collection designed to distinguish among them—and at least one must be economically relevant because “we are a business, not a science project.” Chief philosopher Michael Shrevens, author of The Knowledge Machine, helps connect that culture to technologies that quantify uncertainty.
Goldman and co-founder Kurt House decided in 2018 to stop working on fossil fuels after their work in energy and private equity, then reasoned from the materials required by batteries and AI. Their conclusion: building a future powered by batteries and AI will require, over the next 25 years, mining more copper than has been mined in all human history, while lithium production needs to increase about tenfold—a scale that makes better discovery both a business opportunity and an industrial necessity.