The U.S. Can’t Build AI Without These Materials
The Western mining industry is broken, hampered by a talent drain, slow technology adoption, and misaligned incentives. A new, vertically integrated, software-first approach leveraging Reinforcement Learning (RL) and LLMs can build and operate mines and refineries faster, cheaper, and more flexibly, addressing critical geopolitical supply chain risks.
The Geopolitical and Industrial Challenge
Critical minerals underpin our entire modern economy, from consumer electronics and renewable energy to AI data centers and defense applications. However, the Western world has largely lost the ability to build and operate the complex infrastructure required for mining and refining. The industry is characterized by:
- A Contracting Labor Pool: The mining industry is not seen as a “sexy” field, leading to a shrinking talent pool of engineers, metallurgists, and skilled trades.
- Geopolitical Vulnerability: Supply chains are heavily reliant on China, which has a strategic, top-down approach and a vast, experienced talent pool that enables them to build and commission refineries in months, compared to the 2-4 years it takes in the West.
- Incentive Misalignment: Unlike manufacturing where volume drives costs down, in mining, higher demand increases prices. This creates a fundamental conflict with customers who need to scale.
The Mining Value Chain: From Rock to Product
The process of turning ore into a usable product is incredibly complex, bespoke, and capital-intensive.
- Exploration & Mining: Finding and extracting ore, which often has a concentration of the target metal below 1%.
- Concentration: A mechanical, thermal, or chemical step to produce an intermediate product (e.g., 10-50% concentration).
- Refining: A series of chemical processes to create a high-purity metal.
- Manufacturing: Converting the pure metal into engineered materials, like cathodes for batteries or alloys for magnets.
A key challenge is that the “flowsheet”—the series of unit operations for processing—is custom-designed for each specific mineral deposit. This makes the system inflexible to the natural changes in ore grade and composition as a mine develops.
Why the Industry is Ripe for Disruption
The incumbent mining industry is “calcified” and slow to innovate.
- Risk Aversion: With billions of dollars on the line, operators are not incentivized to try new things. Even small changes risk shutting down a plant, a multi-million dollar event.
- The Startup “Death Spiral”: New technology vendors face a difficult sales cycle. They may get a pilot project, but the long timelines (a new mine is built perhaps every 5 years) and a culture of “don’t touch my cash register” make it incredibly hard to get new tech into commercial-scale applications.
- Outsourcing Core Competency: Large mining companies often outsource engineering and development to EPCs (Engineering, Procurement, and Construction), losing control over the final product—the refinery itself.
A Software-First, Vertically Integrated Approach
The solution is not a point solution but a full-stack, vertically integrated minerals company that treats the mine and refinery as a single, optimizable product. This is enabled by two core software platforms:
-
Capital Project OS (for Construction): Leverages LLMs to automate workflows in engineering, procurement, and construction management. The goal is to reduce churn and latency, running a construction project with the efficiency of a modern manufacturing facility. This is about democratizing access to data to shorten the 3-week lag that typically exists between field activity and executive decision-making. The aim is to enable a team of 200 to do what currently requires 10,000.
-
Plant OS (for Operations): Uses Reinforcement Learning (RL) to remove humans from the loop in controlling the complex, interconnected web of refinery operations.
These are large multivariable optimization problems that RL is perfectly poised to solve.
Refineries are effectively large robots with sensors and actuators, but the constantly changing ore feed makes manual optimization impossible. The current solution is to blend feedstock to minimize variability. Plant OS flips this by creating a hyperdynamic and flexible refining circuit.
Inspired by Google DeepMind’s success in reducing data center energy consumption by 30-40% using RL, Plant OS can find the global optimal operating conditions for a system with a thousand control variables and high latency. This dramatically reduces reagent and energy consumption, increases metal recovery, and slashes the time it takes to commission a new refinery from years to months.
Strategy and Path Forward
The initial strategy is to acquire and develop what the industry considers “subscale” assets, which major players ignore. By applying this tech-first approach, these assets become economically viable.
For this to succeed at a national level, several things are needed:
- Streamlined Permitting: The current permitting process in the U.S. is a major bottleneck for both exploration and development, deterring talent and capital.
- Demand-Side Support: To mobilize the trillions in private capital sitting on the sidelines, the government should provide offtake agreements with price floors for critical minerals. This de-risks investment in an otherwise volatile commodity market.
- Focus on Key Metals: The highest-need metals by mass are the big ones: copper (for electrification), aluminum, iron, and zinc. Lithium is also critical, needing to 4x its production capacity in the next decade to meet battery demand.