On April 29, 2026, in Chililabombwe, Zambia, a $2.3 billion underground mine targeting 300,000 tons of annual copper production broke ground—the largest copper deposit discovered in the country in a century, found by KoBold Metals, an eight-year-old company backed by Bill Gates and Jeff Bezos, using two and a half years and 120,000 meters of drilling. At the same moment, the global mineral discovery rate has fallen 75% over the past decade, and the success rate for greenfield exploration is just one in five thousand. Geological exploration—one of humanity's oldest industries—is being reshaped by algorithms, satellites, and national security agendas alike.
The geological exploration industry faces a brutal statistical reality: over the past decade, the global major mineral discovery rate has fallen by roughly 75%; the easy-to-find "outcrop deposits" have largely been exhausted; the success rate for greenfield exploration projects (starting from scratch in areas with no known mineralization signs) from立项 to forming a commercial mine is only about 1/5,000; and the overall success rate of traditional drilling has long hovered around 0.5%—for every 200 boreholes drilled, perhaps only 1 actually hits an economic ore body.
Meanwhile, the energy transition and AI compute expansion are creating unprecedented demand: EV batteries need lithium, cobalt, and nickel; wind and solar power need copper and rare earths; data centers are themselves a "power + copper" heavy-asset game. The IEA projects that by 2035, the global copper supply could face a shortfall of up to 30%. Geological exploration is, for the first time, confronting the dual pressure of "resources getting harder to find" and "demand growing ever more urgent."
The core bottleneck AI is trying to solve, in the words of KoBold engineers: "The bottleneck in mineral exploration has never been a lack of data, but rather humans' inability to process and recognize patterns in massive, multi-dimensional datasets." Fusing geochemical, geophysical, satellite remote sensing, historical drilling, and other multi-source heterogeneous data into a unified model to identify anomalous patterns that human eyes would almost inevitably miss—this logic applies equally to metallic mineral exploration and oil & gas seismic data interpretation, and is beginning to extend to geothermal, groundwater, and carbon sequestration site selection.
KoBold Metals (founded 2018, headquartered in Berkeley, CA) is currently the best-funded and most fully storied company in AI prospecting. The core is not a single algorithm, but a combination of two platforms—TerraShed integrates geological maps, geophysical surveys, geochemical analyses, drilling records, and satellite imagery; Machine Prospector uses machine learning on this foundation to identify mineralization patterns. The company operates roughly 60 exploration projects across four continents, focused on copper, lithium, nickel, and other battery and grid metals.
The funding trajectory itself is a barometer of this boom: Series B $192M (2022); Series C $537M (January 2025, oversubscribed by $10M), led by Durable Capital Partners and a T. Rowe Price fund, valuing the company at roughly $2.96B, with cumulative funding exceeding $1B. The investor list is an intersection of Silicon Valley and energy capital: Bill Gates's Breakthrough Energy Ventures, Jeff Bezos, Andreessen Horowitz Growth, Norway's state oil company Equinor, and Mitsubishi Corporation.Discovery Alert
Noteworthy comparative data: the global average cycle from discovery to production for mining projects is about 17 years, and the US domestic average is even longer at 32 years—KoBold went from acquisition to ground-breaking in under 4 years, demonstrating AI's practical effect in compressing the "discovery" phase. But breaking ground is only the beginning; whether this speed advantage can extend into "mine construction"—an engineering phase KoBold has never independently completed—remains an open question (see § 06).Forbes
The bottleneck in mineral exploration has never been a lack of data, but rather humans' inability to process and recognize patterns in massive, multi-dimensional datasets.
KoBold Metals EngineersBeyond KoBold, 2024–2026 has seen a crop of companies with diverse technical approaches, and 2025 was the most densely funded year for this niche: Earth AI (Australia, AI target-zone identification + vertically integrated core assay lab, aiming to compress assay turnaround from 5 months to 5 days), VerAI Discoveries (directly holding mineral rights portfolios rather than just selling software), GeologicAI (high-resolution core scanner; in December 2025 acquired LIBS scanner manufacturer Lumo Analytics), VRIFY (AI-assisted discovery SaaS; a 2025 survey showed 56% of 135 practitioners already using AI tools daily), SensOre (Australia, "data cube" containing 2,500+ data layers), Atomionics (Singapore, quantum gravimeter/magnetometer), Plotlogic / TerraEye (hyperspectral remote sensing mineral identification).TechCrunch Mining.com.au
This landscape reveals two industry logics: first, from "selling software" to "holding assets"—VerAI and KoBold are not satisfied with licensing AI tools; they directly hold exploration mineral rights, betting that AI advantages can translate into actual asset ownership; second, from "algorithms" to "full data-pipeline chains"—Earth AI building its own assay lab and GeologicAI acquiring rock scanner manufacturers reflect an emerging consensus shifting from "competing on model accuracy" to "competing on who can fastest convert field data into clean, training-ready data."
Hyperspectral remote sensing: Satellites or drones equipped with hyperspectral sensors scan the surface across hundreds of bands, capturing minerals' unique "spectral fingerprints" and completing initial mineral screening at aircraft or satellite altitude. Quantum sensing: Companies like Atomionics are developing quantum gravimeters and magnetometers that can theoretically "see" subtle anomalies in subsurface density and magnetic fields without drilling—the most cutting-edge and earliest-stage technology pathway on the exploration hardware side.
Although there is currently no "Earth foundation model" specifically designed for mineral exploration scenarios, Google DeepMind's July 2025 release of AlphaEarth Foundations provides a noteworthy reference—integrating optical satellite imagery, radar, LiDAR 3D mapping, and climate simulation data to generate a unified Earth representation at 10×10 m resolution, with storage requirements only 1/16 of comparable solutions, already partnering with over 50 institutions including the UN FAO and Stanford. Current applications focus on food security and deforestation monitoring, not yet extending to mineral exploration, but this kind of "unified Earth representation" infrastructure is exactly what SensOre and KoBold may leverage in the future.Google DeepMind
The technical consensus within the industry is: the vast majority of currently deployed solutions are pragmatic machine learning approaches like "feature extraction and pattern recognition," rather than large language models or generative image models—the latter are mostly used for辅助文案 work like summarizing reports, not for the core mineralization prediction环节. This represents a notable gap from many people's generative-AI associations with "AI prospecting."
Beyond metallic minerals, oil & gas exploration is the earliest and most mature scenario for AI application in geology. On April 4, 2025, oilfield services giant SLB announced an expanded partnership with Shell to deploy the Petrel™ subsurface software platform across its global assets, using AI to support seismic interpretation workflows, covering both traditional oil & gas exploration and energy-transition-related workflows.SLB
Compared to the "new challengers vs. old giants" narrative in metallic minerals, AI adoption in oil & gas mostly follows a "legacy oilfield services companies self-upgrading digitally" path—SLB and Halliburton themselves control the largest seismic data assets and distribution channels, and prefer embedding AI capabilities into their existing software platforms rather than ceding the market to startups. Multiple industry analyses note that AI-driven seismic reinterpretation can theoretically unearth previously overlooked stratigraphic traps from "存量" data, reducing dry-hole rates—but such cost-saving cases are mostly presented as "typical scenarios" rather than specific, publicly listed company project data, and the independent verifiability of quantitative evidence remains generally weak.
In 2025–2026, China's geological survey system密集ly rolled out multiple AI projects. In March 2026, two domestically developed software systems entered trial use: the Intelligent Geological Mapping System has been demonstrated in over 60 1:50,000 geological mapping projects; the Big Data Intelligent Prospecting Prediction System, based on the self-developed "Fuxi" prospecting large model, has delineated gold, copper, iron, and chromium prospecting target zones in the West Qinling Mountains, Jiaolai Basin, and Xinjiang's Saltohai—the West Qinling gold project "completes data input and calculation for 32 1:50,000 map sheets in just a few minutes," whereas traditional workflows typically require weeks or even months.China Daily
The "Future Exploration System" showcased in October 2025 leans more toward the hardware end: a treasure-hunting large-model integrated command platform, an intelligent drilling system, a heavy-lift electric drone capable of carrying 500 kg, and a field reconnaissance robot with intelligent rock-sample recognition capabilities. In July 2026, Yunnan Province launched an AI platform integrating 90 million data records, reportedly capable of shortening rare-metal development time by up to 95% (officially disclosed data; verification methodology未详述, and should be viewed cautiously).
The macro backdrop: China has high external dependence for 12 strategic minerals including iron, copper, nickel, cobalt, and lithium, and AI prospecting is explicitly positioned as a key lever for transitioning from "experience-driven" to "data-driven" exploration. China Geological Survey researchers also emphasize the importance of data quality—the digital geological logging system incorporates over 320 validation rules, with a straightforward rationale: "Without真实 reliable data, AI is just talking nonsense."
China's "Fuxi" large model, the US CriticalMAAS program, and Silicon Valley-capital-backed KoBold are密集ly ramping up efforts within nearly the same time window—on the surface they are technology competitions, but at their core they are resource sovereignty competitions. Multiple industry analyses consistently note that the underlying driver of the 2024–2026 AI prospecting race is shifting from pure commercial logic toward national security agendas.
Economic geologist Nicholas Vafeas has提出 a widely cited three-stage thesis: AI's image in mineral exploration has evolved from initially being deemed "unreliable," to一度 being overhyped as a "万能 solution," to now gradually returning to a rational认知. He compares AI to the 1979 VisiCalc spreadsheet—"an amplifier of human capability, not a replacement."EGU Blogs
"Garbage in, garbage out" remains the most fundamental constraint—historical drilling records largely depend on the personal interpretations of different geologists, with enormous variations between them, which is precisely the hardest domain for algorithms to handle. Key metrics circulating in the industry—such as "drilling success rate improving from 0.5% to 75%," "CO₂ emissions reduced by 90%," and "rare-metal development time shortened by 95%"—are本质上 mostly vendor public-relations claims or results from specific optimization scenarios, rather than third-party-audited, industry-wide established facts.
KoBold's Mingomba project is pulling this debate from "lab/PPT" into "real engineering" territory. As of May 2026, the Zambian copper belt has abundant groundwater resources, and water control will span the mine's entire lifecycle; the ore body lies at 1,700 m depth, where high-temperature, high-pressure environments impose extreme engineering demands; and KoBold has never independently built a mine before, creating a stark contrast with its demonstrated technical优势 in the "discovery" phase. Copper prices also showed notable volatility in Q1 2026 (falling from over $13,000/ton in early January to the $10,000–$12,000/ton range in April), further amplifying the uncertainty inherent in such long-cycle, heavy-asset projects.
A structural issue worth continued monitoring: KoBold's technical优势 heavily relies on historical exploration data accumulated early on by legacy mining giants like Rio Tinto and later transferred out. If "who controls the highest-quality historical geological data" matters more than "whose algorithm is more advanced" in determining the competitive landscape, AI prospecting may ultimately not break the data barriers of old mining majors, but rather spawn new data oligopolies on top of them.
Taking the 2024–2026 progress together, AI's most solid value in geological exploration is concentrated in the "discovery"环节: fusing multi-source data, identifying anomalous patterns hard for human eyes to catch, and converting greenfield exploration's blind trial-and-error into more evidence-based target prioritization. But this boom also exposes clear boundaries: AI currently cannot replace mine construction, hydrological control, or high-temperature/high-pressure engineering—these capital- and time-intensive "hard" stages; data quality, not algorithmic sophistication, remains the primary variable determining project success or failure.
For readers focused on AI adoption, three questions warrant ongoing tracking: Can AI extend its "discovery speed" advantage into a "mine construction speed" advantage? KoBold's Mingomba project is one real-world stress test. Will ownership of historical geological data become the true barrier in the next round of competition? To what extent will the "critical minerals race" override commercial logic and shift toward national security agendas?
First published 2026-07-24