Energy · Deep dive

KoBold Metals

Berkeley AI-first exploration company that mines a century of geochemical, geophysical and satellite data with machine learning to find undrilled copper, cobalt, nickel and lithium deposits — now a $2.96B-valuation, $1B+-raised outfit betting its Zambian Mingomba copper find can prove the model works end to end, from prediction to a producing mine.

emerging

The question that decides it: Does AI-guided exploration convert into a durable, >2x discovery hit-rate advantage over traditional greenfield exploration, and can KoBold retain enough equity in the deposits it finds — rather than being bought out or diluted down by majors with deeper construction capital — to compound its data moat into mine-level economics?

My take ★★ Very interesting ★ Interesting − Maybe ✗ Not interesting

Snapshot

KoBold Metals is a Berkeley, California AI exploration company that applies machine learning to geochemical, geophysical, remote-sensing and historical drilling data to predict where undrilled deposits of copper, cobalt, nickel and lithium sit — the minerals the energy transition needs at multiples of current supply. Founded in 2018, KoBold raised $537M in a Series C priced at a $2.96B post-money valuation in January 2025, led by T. Rowe Price and Durable Capital Partners, taking cumulative funding past $1B (Bloomberg, 1 January 2025; Africanminingmarket, January 2025). Its flagship result is Mingomba, a copper-cobalt deposit on Zambia’s Copperbelt that KoBold and the Zambian state investment vehicle ZCCM-IH are now building into a ~$2.3-2.5B underground mine targeting 300,000-500,000 tonnes of copper annually from the early 2030s — reported by Zambia’s president as potentially one of the largest copper mines in the world (Mining Indaba; Lusaka Times, September 2025). KoBold runs roughly 60 exploration projects across four continents, backed by Bill Gates, Jeff Bezos, Andreessen Horowitz, BHP and Equinor. The company matters now because it is the most capitalized test of whether a software-style, data-driven approach can actually beat the mining industry’s brutal hit-rate economics — where more than 99% of exploration projects never become mines (Praefidi, 2025) — and whether discovery alone, without downstream refining capacity, captures value in a critical-minerals supply chain China still dominates on processing.

Founding story

KoBold was founded in 2018 in Berkeley by Kurt Zenz House, Josh Goldman and Jeff Jurinak, three people who came to mineral exploration from adjacent industries rather than from mining itself. House holds a Harvard PhD in applied mathematics and Earth and planetary science, did private-equity and strategy work at Bain, then founded two energy start-ups before KoBold: C12 Energy in 2008, pursuing CO2-enhanced oil recovery, and Phase Change Resources in 2015, a natural-gas asset acquisition platform that built a physics-based model for forecasting production from fracked wells. That last project is the direct intellectual ancestor of KoBold’s pitch: House had already proven, in oil and gas, that a rigorous statistical model beats engineer intuition at predicting subsurface performance, and asked why hard-rock mineral exploration — an industry that still leans heavily on geologist judgment and sparse drilling — hadn’t done the same.

Goldman, a Harvard physics PhD whose research was in quantum computing, arrived via McKinsey, where he advised energy and materials companies on strategy and corporate finance and reportedly saw firsthand how capital-inefficient exploration economics were across the industry. Jurinak supplied the domain credibility the other two lacked: 35+ years in petroleum exploration and production, the core of it at Conoco and ConocoPhillips, where he rose to Chief Reservoir Engineer with sign-off authority on reservoir engineering for every major E&P project company-wide. The founding thesis was explicitly a transplant: take the subsurface statistical-inference discipline that oil and gas had already industrialized and apply it to copper, cobalt, nickel and lithium — commodities where, unlike oil, exploration had not yet been touched by modern data science at scale. Breakthrough Energy Ventures made its first KoBold investment on 5 March 2019, an early signal that Bill Gates’ climate-tech vehicle saw mineral supply, not just clean generation, as an energy-transition bottleneck.

How it works

KoBold’s core asset is TerraShed, a data platform that ingests more than a century of geological maps, geophysical surveys (magnetotelluric and electromagnetic soundings, gravimetric surveys), geochemical assay data, historical drilling records, satellite and airborne imagery, topography and climate data from public agencies, universities and prior exploration companies’ abandoned datasets. TerraShed standardizes these wildly heterogeneous formats — different coordinate systems, sampling densities, units, and vintages spanning decades — and aligns them spatially and temporally into a unified representation of the subsurface across a target region.

On top of that data lake sits Machine Prospector, KoBold’s suite of algorithms: ensemble machine-learning models trained on historic deposit and non-deposit sites, full-physics joint inversions that reconcile multiple geophysical signal types against a single 3D subsurface model, and computer-vision models that read satellite and drone imagery for surface alteration signatures invisible to the human eye. The output is a probabilistic map of where an economic ore body is most likely to sit — effectively a Bayesian update over ore-body location and grade as each new dataset (a new geophysical survey line, a new drill hole) arrives. Each drill result is fed back into the model, refining the probability surface for the next hole; KoBold’s public materials and outside reporting (Chief AI Officer, 2025) describe this iterative loop as the mechanism that let KoBold’s models flag ground at Mingomba that decades of prior exploration by traditional operators had passed over. Because a private company faces no continuous-disclosure obligation, KoBold has not published a JORC, NI 43-101 or SAMREC-compliant independent mineral resource statement for Mingomba as of 2026 — its own internal estimate cited in press coverage is 247Mt at an average grade of 3.64% copper, with the richest 100Mt running roughly 8% copper (Discoveryalert.com.au, 2026) — figures that carry real weight given Kobold’s track record but have not been through the independent audit process public mining codes require.

Product and business overview

KoBold’s “product” is not sold in the conventional sense; it is deployed internally and through joint ventures. The three components:

TerraShed + Machine Prospector, the proprietary data-and-modeling stack, is not licensed externally as of 2026 — KoBold uses it exclusively to generate its own exploration targets and to evaluate targets brought to it by JV partners and host governments.

Owned and majority-owned exploration/development projects, the highest-profile being Mingomba in Zambia (80% KoBold, 20% rising toward 25% for state investment vehicle ZCCM-IH) and the Disko-Nuussuaq nickel-copper-cobalt-platinum project in Greenland, where KoBold earned a 51% stake through a two-stage earn-in joint venture with London-listed Bluejay Mining across a roughly 2,776 km² license area.

Minority joint ventures with majors and host governments, including exploration partnerships with BHP and Rio Tinto across Australia and Canada, and a July 2025 memorandum of understanding with the Democratic Republic of Congo government followed by seven exploration permits granted in September 2025, with KoBold committing to deploy more than $50M into DRC lithium exploration by 2027 (Africanews, September 2025). Roughly 60 active projects span Zambia, Namibia, the DRC, Quebec, Saskatchewan, Ontario, Western Australia and South Korea.

Business model and pricing

KoBold has no conventional pricing page or per-unit revenue model because it is not, today, selling either a SaaS product or refined metal — it is pre-revenue in the traditional sense. Its economics run on three tracks. First, owned equity in JV mining projects: the 80% Mingomba stake and 51% Disko-Nuussuaq stake are the value KoBold is actually accumulating — worth nothing until a bankable feasibility study, project financing and eventual production convert exploration success into cash flow, but already implicitly underwriting a large share of the $2.96B Series C valuation. Second, exploration-as-a-service-adjacent joint ventures with majors (BHP, Rio Tinto): KoBold contributes targeting and technical work, the major contributes capital and, often, existing tenements, and the two split resulting discoveries by pre-negotiated JV terms rather than a cash service fee. Third, a discovery-and-partner model: KoBold has stated intent — and Kurt House has said this explicitly in interviews — to move beyond pure discovery into actually developing mines, a strategic shift from the historic junior-explorer playbook of “find it, then sell or option it to a major for cash and a royalty.” The $537M Series C explicitly earmarked roughly 40% of proceeds to Mingomba’s development (per reporting on the raise), which is capital spent building a mine, not licensing software. Until Mingomba or another asset reaches production in the early 2030s, KoBold’s cash-flow model is entirely venture-funded equity plus, potentially, project-level debt and offtake financing as feasibility studies mature.

Traction over time

Because KoBold is pre-revenue, “traction” here is capital raised, deposits advanced, and headcount — proxies for progress rather than a P&L. The pattern is a company moving unusually fast from discovery (2024 Mingomba announcement) to construction (2026 groundbreaking) for an industry where discovery-to-production typically runs 10-20 years.

Market analysis

The global critical-minerals market was valued at roughly $391-410B in 2025 and is projected to reach $715-800B by 2035 (SNS Insider; DataM Intelligence, 2025-2026), driven by IEA’s Global Critical Minerals Outlook 2025 projections that under net-zero-aligned demand, lithium demand rises roughly 7x, nickel and cobalt roughly 3x, and copper roughly 2x by 2035. The IEA also flags a structural copper problem specific to KoBold’s core commodity: the current global mine-development pipeline points to a potential 30% supply shortfall by 2035 due to declining ore grades, rising capital costs and a shrinking pace of new discoveries — precisely the discovery-rate crisis KoBold’s thesis is built to solve. J.P. Morgan’s commodities research and the IEA both estimate roughly $800B of mining and refining investment is needed globally by 2040 to meet net-zero mineral demand.

The structural force most relevant to KoBold’s business model, though, is not exploration economics but refining concentration. China refines roughly 76% of the world’s cobalt, 91% of rare earths, and even in the more diversified copper-refining market holds a 44% share — and is the dominant refiner in 19 of 20 minerals the IEA tracks, controlling roughly 70% of global processing capacity overall (Visual Capitalist; Z2Data, 2025-2026). Most of the DRC’s cobalt, regardless of who owns the mine, is shipped to China to be refined into battery-grade material. This means KoBold can discover and even mine a deposit in Zambia or the DRC and still not control the most economically strategic link in the chain — the ore has to go somewhere to be processed, and China’s capacity dominance sets the terms.

Competitive intel

Earth AI is the closest peer on the AI-exploration thesis: Sydney-based, raised an oversubscribed $20M Series B in January 2025 (Tamarack Global, Cantos Ventures) and claims its algorithms have found critical minerals in ground competitors ignored (TechCrunch, March 2025). It is a fraction of KoBold’s scale but structurally similar — data-driven targeting sold as a capability rather than a mine.

VerAI Discoveries, an Israel-linked AI exploration start-up, and GoldSpot Discoveries, a TSX-V-listed, profitable Canadian AI-for-mining-data company, both validate the category without KoBold’s capital intensity — GoldSpot in particular proves a smaller, services-first version of this business can be a standalone public company rather than requiring venture-scale capital.

Ideon Technologies is a genuinely different technical bet: muon tomography that images subsurface rock density directly using cosmic-ray muons, already deployed at Rio Tinto’s Kennecott copper mine in Utah. If muon imaging scales down in cost, it becomes a physical-measurement alternative to KoBold’s probabilistic, data-inference approach — a real technology risk to KoBold’s core pitch that ML-plus-historical-data beats direct physical sensing.

Fleet Space Technologies, Australian, sells exploration-as-a-service using nanosatellite constellations and ambient-noise seismic tomography directly to major miners — the pure services model KoBold has largely chosen not to pursue, betting instead on equity ownership.

BHP, Rio Tinto, Freeport-McMoRan, Glencore, Anglo American and First Quantum Minerals are simultaneously the biggest threat and, in BHP’s and Rio Tinto’s case, JV partners and investors. Freeport and First Quantum in particular have decades of Copperbelt and broader African operating experience that KoBold — an exploration company two years into its first mine build — does not have, and could out-execute KoBold on the harder, more capital-intensive job of actually building and running Mingomba even though KoBold’s data found it first.

History and evolution

• 2018 — Kurt House, Josh Goldman and Jeff Jurinak found KoBold Metals in Berkeley, CA.
• 2019-03-05 — Breakthrough Energy Ventures makes its first investment in KoBold.
• 2021 — Equinor takes an equity stake via Equinor Ventures and pledges additional funding.
• 2022-02 — ~$192.5M round led by T. Rowe Price pushes valuation past $1B; BHP, CPPIB, Apollo Projects, Mitsubishi and others participate.
• 2022 (late) — KoBold acquires its 80% controlling interest in the Mingomba project on Zambia’s Copperbelt.
• 2023-06 — Additional ~$200M tranche closes with Bill Gates-, Jeff Bezos- and Jack Ma-linked capital participating.
• 2024-02 — KoBold publicly discloses the Mingomba copper discovery, characterized in press coverage as Zambia’s largest copper find in a century.
• 2024-10-07 — TechCrunch reports SEC filings showing ~$491M raised cumulatively.
• 2025-01-02 — $537M Series C closes at a $2.96B post-money valuation, led by T. Rowe Price and Durable Capital Partners.
• 2025-07-17 — DRC government signs a mining-cooperation MOU with KoBold.
• 2025-09-02 — KoBold receives seven DRC exploration permits, commits $50M+ to DRC lithium exploration by 2027.
• 2026-04-28 — ZCCM-IH inauguration ceremony marks the formal start of Mingomba shaft-sinking; ZCCM-IH signals intent to raise its Mingomba stake from 20% to 25%.
• 2026 (through the year) — Headcount grows to an estimated 350-450 employees; company opens a DRC office in Lubumbashi.

What people say

The case for. Investors and energy-transition commentators treat KoBold as the most credible attempt yet to apply modern data science to an industry that has historically resisted it — Breakthrough Energy, a16z, T. Rowe Price and BHP have all re-upped across multiple rounds, a signal of conviction rather than one-off speculation. Reporting on the Mingomba discovery (CNN, February 2024; Mining.com) frames it as validating the thesis empirically: KoBold’s models flagged ground that decades of prior, non-AI exploration in one of the world’s most heavily explored copper belts had missed. Podcast-length interviews with Kurt House (Volts, MCJ) and coverage of the Harvard Business School case built around the company treat KoBold as a genuine transformation story for mineral exploration economics, not a marketing exercise.

The complaints. Roland Gotthard, a Perth-based exploration geologist, publicly questioned the basis for KoBold’s $537M Series C, warning that AI models risk misinterpreting sparse or noisy geological data and producing confidently wrong predictions (reported around the Series C, 2025). Because KoBold is private, it faces no continuous-disclosure obligation and has never published a JORC, NI 43-101 or SAMREC-compliant independent resource statement for Mingomba — meaning the widely cited 247Mt-at-3.64%-copper figure is KoBold’s own internal estimate, not an audited one, a genuine transparency gap for a company this well capitalized. Glassdoor reviews are thin (8 total as of 2026) but sharply bimodal: some describe an “interdisciplinary, collaborative, transparent” culture, while others allege a “culture of scapegoating and blame,” legally questionable terminations of employees on leave, and concealment of misconduct — serious enough allegations, even from a small review sample, to flag as a governance risk at a company this valuable. Human-rights researchers (Wilson Center, NYU Stern Center for Business and Human Rights) note that KoBold’s DRC entry follows decades of well-documented child and forced labor in the country’s artisanal cobalt sector — roughly 40,000 of the 255,000 Congolese cobalt miners are children — and that affected communities in mineral-rich DRC regions have historically not been meaningfully consulted before new mining agreements are signed, a live reputational and operational risk regardless of KoBold’s own stated ethical standards.

Outlook: the open question

Does AI-guided exploration convert into a durable, >2x discovery hit-rate advantage over traditional greenfield exploration, and can KoBold retain enough equity in what it finds to compound that edge into mine-level economics — rather than becoming, functionally, a specialized service arm for the majors that end up building and owning the mines?

The yes case requires three things to hold simultaneously. Mingomba must prove out at commercial scale — an independently audited resource statement (JORC or NI 43-101 equivalent) confirming grade and tonnage broadly in line with KoBold’s internal 247Mt-at-3.64%-copper estimate, and the mine actually reaching its targeted 300,000-500,000 tonnes/year of copper production in the early 2030s without a Zambian political, permitting or financing derailment. KoBold’s next three flagship targets must also deliver — Disko-Nuussuaq in Greenland and at least two of the DRC or Australian/Canadian JV targets need to convert from “AI-flagged anomaly” into drill-confirmed, economically viable deposits, demonstrating the hit rate is systematic and not a single lucky Mingomba outcome. And KoBold must retain 30%+ operating equity in the resulting mines rather than selling down its stake to majors for cash at the feasibility stage — the current 80% Mingomba position is a good early sign, but the capital intensity of mine construction (Mingomba alone needs $2.3-2.5B) creates constant pressure to dilute down in exchange for a major’s balance sheet.

The no case is just as plausible. Majors bundle their own AI — BHP, Rio Tinto and Freeport all have in-house data-science teams and can license or build comparable ensemble-ML targeting without paying KoBold’s valuation premium, especially once KoBold’s public track record demonstrates the approach is replicable. The hit rate turns out comparable to well-funded traditional exploration once you control for KoBold’s unusually large, well-capitalized dataset access (itself a function of raising $1B+, which most junior explorers cannot do) rather than any unique algorithmic edge. And KoBold becomes a service provider by default — needing majors’ capital and construction expertise for every project past discovery, KoBold’s JV terms erode over successive rounds of dilution, and the company ends up capturing exploration-stage value (options, minority stakes, technical fees) rather than production-stage value, which is where the actual multi-billion-dollar economics of a Mingomba-scale deposit sit.

How to attack it

Do not attack KoBold on pure algorithmic targeting — it has a two-year, $1B+ head start on data acquisition and a validated discovery (Mingomba) that is hard to argue with. Attack the two things KoBold structurally cannot fix quickly: its distribution model and its downstream exposure.

Wedge 1: a sovereign-licensing platform, not a major-partnership platform. KoBold’s JV strategy runs through majors (BHP, Rio Tinto) and host-government equity partners (ZCCM-IH), but its commercial relationship is still fundamentally “we bring the data, you bring the capital and the mine-building expertise.” A new entrant could instead build an AI exploration platform explicitly licensed to sovereigns — Chile, Indonesia, the DRC, Zambia itself — that own their subsurface data and mineral rights outright, positioning the AI vendor as a state’s own in-house capability rather than a competing equity claimant. This inverts KoBold’s structure: the state keeps 100% of the discovered resource and pays a licensing or success fee, which is a far easier sale in resource-nationalist political environments (see: Indonesia’s nickel downstream mandates, Chile’s lithium nationalization moves) than KoBold’s equity-taking model.

Wedge 2: attack the refining chokepoint, not the discovery chokepoint. China’s 76% cobalt / 91% rare-earth / 44% copper refining share means discovery is not actually where the scarcity value sits — processing is. An AI-plus-process-engineering company targeting non-Chinese refining capacity (hydrometallurgical cobalt/nickel separation, copper smelting) captures a chokepoint KoBold’s business model does not touch at all.

Weaknesses to exploit. KoBold has never built or operated a mine — Mingomba is its first, and mine construction, not discovery, is where cost overruns and multi-year delays actually happen (see: Rio Tinto’s Oyu Tolgoi underground expansion, years late and billions over budget). Zambia carries real sovereign and political risk — a change in government mining policy or ZCCM-IH’s stake negotiations could reprice the entire Mingomba economics. Water and community-consent risk is unresolved: large open-pit and deep-underground copper mines carry heavy tailings and water footprints regardless of how the deposit was found, and DRC communities have historically not been meaningfully consulted ahead of mining agreements (Wilson Center, 2025). Discovery-to-production cycles run 10-20 years; an attacker targeting brownfield reprocessing or tailings retreatment, with 2-4 year cycles to cash flow, can compound capital and learning far faster than KoBold’s greenfield model ever will.

Adjacent-segment play

The most direct adjacent-segment play is tailings reprocessing — an estimated $40B of recoverable critical minerals sits in existing mine tailings worldwide, material that is already mined, crushed and characterized, meaning the “find it” problem KoBold solves with expensive greenfield exploration is already solved. KoBold’s TerraShed/Machine Prospector stack could in principle be repointed at historical tailings datasets (assay records, mill throughput logs) to identify which of thousands of legacy tailings dams hold the highest-grade recoverable copper, cobalt or nickel — a 2-4 year cycle to cash flow versus Mingomba’s 10+ years. Nth Cycle and other tailings/e-waste recovery companies are already working this angle from the processing side; nobody has yet paired KoBold-grade ML targeting with tailings reprocessing specifically.

A second adjacent segment is geothermal reservoir targeting — KoBold’s subsurface Bayesian-inference stack (originally built on House’s oil-and-gas forecasting background) maps closely onto the reservoir-characterization problem geothermal developers like Fervo Energy solve, just with heat and permeability as the target variable instead of ore grade. The data types differ (temperature gradients, permeability logs vs. geochemical assays) but the inference architecture — full-physics inversion plus ML on sparse subsurface data — transfers directly.

A third, more contrarian adjacent segment is deep-sea polymetallic nodule targeting, where Impossible Metals and The Metals Company are already extracting nickel-copper-cobalt-manganese nodules; an AI-first seabed-mapping company using satellite, sonar and prior survey data could play the KoBold role for ocean floor exploration rather than terrestrial. The wedge that does not generalize well is oil-and-gas basin remapping: it is technically the closest fit to KoBold’s own subsurface-inference DNA, but it is directly counter to KoBold’s climate-mission positioning and investor base (Breakthrough Energy would not fund it), so while the model transfers, the company and its capital base cannot follow it there.

Sources and further reading

• Gates-Backed KoBold Metals Raises $537 Million in Funding Round — Bloomberg, 1 January 2025.
• KoBold used AI to find copper — now investors are piling in to the tune of $537M — TechCrunch, 2 January 2025.
• AI-powered critical mineral startup KoBold Metals has raised $491M, filings reveal — TechCrunch, 7 October 2024.
• Gates, Bezos-backed KoBold Metals says Zambia copper find largest in a century — Mining.com.
• KoBold Metals breaks ground on $2.3 billion Mingomba mine — Mining Indaba.
• Zambia’s Stake in Mingomba Copper Mine: 2026 Update — Discovery Alert, 2026.
• Bluejay and KoBold Metals form Greenland critical metals JV — Mining Technology.
• KoBold Metals secures exploration permits in DRC amid U.S. mining push — Africanews, 4 September 2025.
• Report: KoBold Metals Business Breakdown & Founding Story — Contrary Research.
• Charted: China’s Grip on Critical Mineral Refining — Visual Capitalist, 2025-2026.
• Executive summary – Global Critical Minerals Outlook 2025 — IEA, 2025.
• How Kobold Metals Found 300,000 Tonnes of Copper Using AI — Chief AI Officer, 2025.
• Inside Berkeley’s $3 billion AI mining company — Berkeleyside, 23 June 2026.
• KoBold Metals Reviews — Glassdoor — Glassdoor, 2026.
• The DRC Mining Industry: Child Labor and Formalization of Small-Scale Mining — Wilson Center.