No PriorsNo Priors Ep. 111 | With KoBold Metals Co-Founder and President Josh Goldman
CHAPTERS
- 0:00 – 3:11
KoBold’s mission: AI-driven mineral exploration (and why exploration, not mining)
Sarah and Elad open with what KoBold Metals is and what “intelligent mining” means in practice. Josh explains KoBold focuses on exploration because the value creation and technological differentiation are greatest there, especially as easy-to-find surface deposits are largely gone.
- •KoBold explores for battery/AI-critical metals like lithium and copper
- •Exploration can yield 100–1,000x returns but has very low baseline success rates
- •Modern exploration is harder: deposits are deeper and concealed
- •Technology + human judgment is the differentiator across the exploration lifecycle
- 3:11 – 5:16
The data advantage: assembling a planet-scale geoscience corpus
Josh details the diversity of geoscience data and why aggregation is itself a major moat. He walks through the spectrum from satellite imagery to airborne geophysics to geochemical sampling, plus the messy reality of fragmented repositories and unstructured reports.
- •Multi-scale data: satellites, geology/tectonics, airborne magnetic/density/conductivity surveys
- •Ground truth sources: rock/soil samples and historical field mapping
- •Most data is public but scattered across thousands of repositories
- •Unstructured text and maps are critical but difficult to use without modern tooling
- 5:16 – 6:18
Ground truth from the past: turning archival maps into training data
A vivid example illustrates KoBold’s approach to data: nearly century-old hand-painted maps in Zambia’s archives. Josh explains why such data doesn’t “expire” and how it can anchor model training against modern remote-sensing signals.
- •Archival datasets can be uniquely valuable and impossible to recreate
- •Old field observations remain valid because the rocks haven’t moved
- •Digitization and curation convert rare artifacts into usable training data
- •Combining historical ground truth with modern imagery boosts predictive models
- 6:18 – 8:47
How lithium exploration traditionally works: ore deposits as an information problem
Sarah asks how explorers look for lithium absent KoBold’s approach. Josh reframes exploration around understanding the ‘recipe’ for ore formation and locating where natural processes concentrated dilute crustal metals into economic grades.
- •Ore deposits form when geology concentrates metals from ppm levels to ~1%+
- •Exploration starts from formation models (“recipes”) and hypotheses
- •Key scarcity is information about locations, not total metal in the crust
- •Workflow: hypothesize → narrow regions → acquire land → test with new data
- 8:47 – 10:57
Regulatory constraints vs. geologic rarity: choosing where discovery can become a mine
Elad probes whether regulation is the primary barrier, especially in the U.S. Josh explains both matter: deposits are rare enough that early filters can’t be too narrow, but long-term property rights and community acceptance are essential for economic success.
- •Deposits are rare enough that global search breadth matters
- •Jurisdictional factors: property rights stability, taxes/royalties over decades
- •‘Social license to operate’ and community relationships are make-or-break
- •Constraints are hyper-local (states, communities, indigenous groups, chiefdoms)
- 10:57 – 13:46
KoBold’s portfolio at scale and the Zambia flagship discovery (Mingomba)
Josh outlines KoBold’s operating model and global footprint, then highlights Mingomba in Zambia as an exceptional, high-grade copper discovery. He explains why grade drives not just economics but also environmental footprint and capital intensity.
- •Portfolio: 60+ projects across four continents; mostly pre-discovery seeds
- •KoBold operates exploration programs and holds exploration rights (often via JVs)
- •Mingomba: exceptionally high-grade large copper deposit not yet a mine
- •High grade means less rock moved, smaller plants, less waste, lower costs/footprint
- 13:46 – 16:23
The full-stack technical approach: sensors, unified data systems, and predictive models
Elad asks what’s being built from an AI perspective. Josh describes KoBold’s three-part system—novel sensors, an integrated data platform, and many models operating at multiple scales—enabling systematic interaction with both structured and unstructured data.
- •Full-stack exploration decision system with dozens of products
- •Theme 1: new sensors/hardware to capture new Earth data types
- •Theme 2: unified data system to work across the entire corpus at once
- •Theme 3: multiple models across length scales; LLMs help interface with messy data
- 16:23 – 17:55
Closed-loop exploration: hypotheses, uncertainty, and daily model retraining in the field
Sarah drills into how predictions get validated. Josh explains how KoBold generates falsifiable hypotheses (surface and 3D subsurface), targets high-uncertainty areas to learn faster, and continuously updates models with new ground truth.
- •Predictions range from surface rock type to 3D conductive layers and ore grade distributions
- •Selecting drill locations/directions is framed as testing probabilistic hypotheses
- •Field teams collect new training data daily; models are retrained and re-served
- •Learning is accelerated by sampling where model uncertainty is highest
- 17:55 – 19:21
Building new sensors when the industry won’t: proprietary hyperspectral airborne imaging
Josh gives an example of moving beyond off-the-shelf tooling by building KoBold’s own hyperspectral imaging system. The goal is faster, cheaper acquisition of richer imagery (hundreds of spectral bands) to update plans while exploration is underway.
- •Mining services can be slow to adopt new chip/sensor advances
- •KoBold built a hyperspectral system quickly and deployed it on light aircraft
- •600-color imagery at lower cost and faster turnaround
- •Tight integration of new sensing with other datasets improves decision cycles
- 19:21 – 20:51
No ‘silver bullet’: discovery comes from high-dimensional data + scientific insight + HI/AI
Elad asks if one dataset enabled the Zambia discovery. Josh rejects the ‘single killer dataset’ framing and emphasizes that predictive power comes from combining many data dimensions, quantifying uncertainty, and integrating scientific understanding with AI.
- •High-dimensional integration beats reliance on any one technology
- •Incremental gains (e.g., gravity gradiometry) matter but aren’t magic
- •Quantifying uncertainty is central to better decisions
- •AI cannot be separated from human scientific judgment in this domain
- 20:51 – 23:53
Valuing discoveries: mines as knowable cash-flow assets (and why deposits can last decades)
Sarah asks how projects are valued when sold or developed. Josh explains standard NPV-of-future-production valuation, noting operational variables are relatively concrete (throughput, recovery, capex/opex), and mines often extend beyond the initial 20-year plan.
- •Valuation: present value of future production, adjusted for stage and jurisdiction risk
- •Inputs are tangible: plant size, trucking, water/power costs, recovery rates
- •Commodity price is a key sensitivity, but volumes are highly engineerable
- •Mines frequently expand resources and operate 50–70+ years beyond initial underwriting
- 23:53 – 25:39
Exploration success rates are collapsing—and KoBold’s target economics per discovery
Sarah asks how often explorers succeed and how much better KoBold can be. Josh argues the industry’s success rate has worsened ~10x over 30 years as the problem got harder and innovation lagged; KoBold targets a much lower cost per high-quality discovery.
- •Measuring success is ambiguous; better metric is discoveries per capital deployed
- •Industry: per $1B, once yielded ~8 discoveries; now less than 1 high-quality deposit
- •Exploration is unattractive in aggregate due to many failures and few wins
- •KoBold target: ~$50–$100M per discovery; Mingomba is a proof point but must repeat
- 25:39 – 28:22
Scarcity isn’t capital—it’s great deposits: where the world is underexplored
Elad probes whether ESG and funding constraints impede mining investment. Josh says great projects get funded; the real scarcity is world-class deposits, and underexploration varies by commodity—especially for lithium, where science and targeting are still early.
- •Capital is available for high-quality, de-risked projects; weak projects struggle
- •Copper is mature but still underexplored in deeper/concealed basins (e.g., Zambia)
- •Some jurisdictions (e.g., Congo) have large potential but complex challenges
- •Lithium exploration is comparatively young; improved scientific understanding is a major edge
- 28:22 – 32:24
What’s actually ‘rare’: rare earths vs. processing chokepoints, and KoBold’s commodity focus
The conversation turns to perceived scarcity and geopolitics. Josh explains rare earths aren’t truly rare; the bottleneck is concentrated processing capacity (notably in China), which shapes supply chains and economics more than geology does.
- •Rare earths are often overstated due to naming; scarcity is more about processing
- •China’s downstream buildout creates competition for concentrate feedstock
- •Processing concentration reinforces manufacturing concentration and supply-chain leverage
- •KoBold focuses on large markets with strong demand tailwinds (Cu, Li, Ni, Co)
- 32:24 – 36:34
Exploration as an epistemic discipline: falsifiability, uncertainty, and avoiding confirmation bias
Sarah revisits a prior discussion about philosophy’s role at KoBold. Josh frames the company as an ‘epistemic project’ focused on better predictions—embracing uncertainty, working with multiple hypotheses, and defining in advance what evidence would falsify a model.
- •Core practice: quantify uncertainty rather than commit to a single ‘best model’
- •Predictions must be falsifiable; pre-register what would disprove the hypothesis
- •Multiple competing hypotheses guide what data to collect next
- •Company culture formalizes this via an ‘Epistemology of Exploration’ and a chief philosopher
- 36:34 – 40:53
Origin story: from physics and energy investing to mining the materials for batteries and AI
Josh closes by explaining how he and co-founder Kurt House arrived at KoBold. They reasoned from first principles about future raw material needs—especially copper for electrification/data centers and lithium for batteries—concluding society must find far more deposits with better tools.
- •Background: physics/quantum computing, energy consulting, oil & gas investing
- •2018 pivot away from fossil fuels toward critical materials
- •Scale thesis: next 25 years may require more copper than all prior human history; ~10x lithium growth
- •Technology-enabled exploration is both a differentiated business and societal necessity
