A Berkeley startup is using machine learning to do what generations of human prospectors could not: reliably predict where the Earth is hiding its most critical metals.
In the summer of 2022, a team at KoBold Metals sat in an office in Berkeley, California, and pointed a machine at a map of Zambia. No helicopter, no drill rig, no geologist standing in the mud of the Copperbelt province squinting at a rock face. What they had was a software platform called TerraShed, loaded with over a century of geological records, satellite imagery, geochemical surveys, and gravity readings. A second platform, Machine Prospector, ran statistical inference across the data and suggested where high-grade copper mineralization was most likely to be hiding underground. Within roughly fourteen months of drilling on the site their AI had flagged, KoBold confirmed the Mingomba copper deposit in Zambia's northern Copperbelt, now described as potentially the largest high-grade copper discovery in at least a decade, and possibly the most significant Zambian find in a century.
KoBold Metals was founded in 2018 by Kurt House, Josh Goldman, and Jeff Jurinak. House is an unusual figure for the mining world: he holds a Ph.D. in applied mathematics and earth science from Harvard, taught as an adjunct professor at Stanford, and was a research fellow at MIT before building two prior energy ventures. His co-founder and president Josh Goldman brought deep field experience, having been involved in some of the largest mineral discoveries of the prior generation, including the largest copper and zinc finds of recent decades.
The company's thesis was that mineral exploration had become a data problem masquerading as a geology problem. For decades, the industry had drilled based on human interpretation of scattered, poorly digitized datasets. KoBold's argument was that the real information was already there, buried in old core samples, government surveys, and satellite passes that no human team could synthesize fast enough to be useful. TerraShed acts as the aggregation layer, ingesting and cross-correlating more than a century's worth of public and private geoscience data into a unified model of the Earth's crust. Machine Prospector then applies machine learning to that corpus, building subsurface predictions that flag where compositional anomalies are statistically most likely.
The team KoBold assembled reflects this hybrid vision. Roughly a third of the company consists of veteran exploration geologists, collectively carrying what House has described as more than 200 years of field experience. The other two-thirds are software engineers and data scientists, almost all of whom hold degrees in the physical sciences: astrophysics, condensed matter physics, geophysics, nuclear chemistry. The intent is not to replace geological judgment but to amplify it with far more signal than any individual mind could process unaided. In House's words, they are not trying to optimize consumer behavior online. They are making statistically valid predictions about compositional anomalies within the Earth's crust.
The Mingomba deposit sits in Zambia's northern Copperbelt, one of the most productive copper-producing regions on the planet, and yet its scale had been missed. KoBold entered the project through a joint venture with EMR Capital, an Australian private equity firm, and ZCCM-IH, Zambia's state-owned mining investment vehicle.
When drilling began confirming KoBold's models, what came up was striking. The deposit carries copper ore grades of approximately 5%, placing it on par with Ivanhoe Mines' Kakula deposit in the neighboring Democratic Republic of Congo, described by its developers as the world's fastest-growing, highest-grade, lowest-carbon major copper operation. KoBold is now planning a $2 billion to $2.3 billion underground mine projected to produce 300,000 tonnes of copper per year, with shaft construction targeted to begin around 2027 and first production in the early 2030s. Zambia's president Hakainde Hichilema has stated that at full production rates, the mine could yield over 500,000 to 600,000 tonnes annually, a figure that would put it in the tier of Escondida in Chile, currently the world's largest copper operation.
KoBold Africa CEO Mfikeyi Makayi announced the discovery publicly at Mining Indaba in Cape Town in February 2024. Makayi holds the distinction of being the first Zambian woman in history to lead a major mineral exploration company. Jito Kayumba, special assistant to the Zambian president, described what had been found as "quite phenomenal." The find validated not just the scale of the deposit, but the speed of confirmation: less than fourteen months from initial drilling to announcement, a compression of the usual exploration timeline that even KoBold's own backers found remarkable.
Not everyone has accepted the story at face value. Some industry geologists have noted that Mingomba sits in a historically mineralized region, prompting skeptics to describe the discovery as finding an elephant in elephant country rather than a paradigm shift in exploration science. Critics have also pointed to KoBold's limited public disclosure of its resource estimate and the fact that the deposit lies more than a mile underground, adding substantial complexity and cost to development. A third-party review seen by the Wall Street Journal broadly supported KoBold's findings, but questions about the AI's precise contribution have persisted in trade publications.
What makes KoBold unusual is not just the technology. It is the quality and diversity of investors willing to back a mining startup before a single tonne of ore had been sold.
Breakthrough Energy Ventures, the climate fund backed by Bill Gates and seeded by investors including Jeff Bezos and Ray Dalio, was an early anchor. Jack Ma's investment office joined. So did Andreessen Horowitz, which rarely takes positions in hard assets. Equinor and Mitsubishi Corporation added strategic industrial credibility. In January 2025, KoBold closed a $537 million Series C round co-led by Durable Capital Partners and two T. Rowe Price funds, pushing the company's valuation to $2.96 billion and its total funding past $1 billion. The final close came in $10 million above the initial target.
The allocation plan was direct: approximately 40% of the new capital would advance Mingomba from discovery toward production, with the rest funding exploration expansion across five continents. KoBold spent $100 million on exploration in 2023 alone, a figure comparable to programs run by BHP and Rio Tinto, both of which have formed active partnerships with the company. As of 2025, KoBold reported approximately 60 active exploration projects, with lithium targets in Finland, Botswana, and Canada, and newly commenced operations at the Manono lithium project in the DRC. House has publicly stated he wants to add at least three new jurisdictions in the near term.
The investment thesis behind KoBold is inseparable from a single structural fact: the world needs far more copper than it is currently producing, and the shortfall is getting harder to close.
The International Energy Agency has warned that current mine discovery pipelines could produce a 30% supply shortfall by 2035. Each electric vehicle requires roughly 80 kilograms of copper for its wiring and drivetrain components. Data centers, the physical infrastructure of the AI economy, require between 5,000 and 15,000 tonnes each. Wind turbines and solar installations are copper-intensive at scale. Meanwhile, the industry has been discovering fewer large deposits per dollar spent than it did a generation ago. The easy, near-surface orebodies have largely been found. What remains tends to be deeper, more remote, and more expensive to confirm.
The traditional industry's hit rate makes the scale of the problem concrete: roughly three in every 1,000 exploration targets develop into commercially viable mines. KoBold has not disclosed its own hit-rate metrics, a gap that critics note. But the Mingomba find is a hard counter-argument: one confirmed deposit of world-class scale, located by a team that had been operational for just four years when drilling began.
House has been candid about the competitive context. As China has tightened export controls on antimony, gallium, and germanium and proposed new restrictions on lithium processing technology, the geopolitics of critical minerals have shifted from background concern to front-page urgency. For Western governments and institutional investors, a company that can find the next Mingomba before a competitor does, and find it at speed, is no longer merely interesting. It is strategically necessary.
KoBold is also building the human infrastructure to support its long-term ambitions. The company has contributed over $200 million to the Zambian economy to date and sources goods and services primarily from local Zambian providers. It has partnered with Stanford University, Copperbelt University, and the University of Zambia to offer master's degrees in data science and exploration geology. In July 2025, it signed an agreement with the DRC government to digitize and publicly release geological records held in Belgium, a move that expands the training data available to its own models and to the broader scientific community. These commitments reflect a calculated reality: a company trying to operate mines in sovereign African nations over a multi-decade horizon has to be trusted as a partner, not just treated as a capital allocator.
The Mingomba mine, if it reaches full production as projected, will not begin shipping copper until the early 2030s. A great deal will change between now and then. But what KoBold has already demonstrated is the more immediate point: that the combination of historical data at scale, machine learning, and a team of geologists and physicists working from the same platform can compress the most uncertain phase of mining into something that begins to look less like prospecting and more like a repeatable science. In a world that will need far more copper than the ground has yet given up, that compression may matter more than any single deposit, however extraordinary.