Erin Price-Wright hosts Turner Caldwell (Mariana Minerals) and Drew Baglino (former Tesla SVP) for a rare practitioner-level discussion on rebuilding the physical substrate of American industry. The conversation splits into two concrete domains: upstream mineral extraction and downstream grid infrastructure, with both founders arguing that software-centric approaches fail without vertical integration into hardware and permitting.
Caldwell walks through how Mariana Minerals is collapsing the exploration-to-production timeline by treating mining as a robotics problem. Rather than the industry-standard approach of separate contractors for sensing, drilling, and feasibility studies, his team fields autonomous drilling rigs guided by reinforcement learning models that improve with each borehole. The key insight: supervised learning fails in this domain because there is no labeled training set for subsurface mineralogy—the model must learn by interacting with the physical world. He makes the economic case that co-locating processing facilities at the mine mouth eliminates the dominant cost driver (transporting waste material), making US-based operations competitive without tariff protection.
Baglino addresses the grid side with equal specificity. He identifies the aging transformer fleet as the binding constraint on electrification, explaining why solid-state architectures using silicon carbide semiconductors can replace passive 50-year-old units with software-defined voltage control. This matters because solar arrays and battery storage natively produce DC power; solid-state transformers eliminate redundant AC-DC conversion stages while providing the granular control needed for bidirectional power flows. Both founders converge on permitting reform and federal investment frameworks as the non-negotiable policy prerequisites, noting that capital currently flows to jurisdictions where projects can actually break ground within a decade.
Key Insights
- Mineral exploration can be compressed from a 15-year to a 5-year timeline by vertically integrating AI-driven sensing, drilling, and refining, removing the latency of disjointed contractor handoffs.
- Reinforcement learning outperforms supervised learning in subsurface exploration because drilling data is sparse and unlabeled; the system must learn from physical trial and error rather than static datasets.
- Co-locating processing facilities next to mining sites eliminates the cost of transporting 99% waste rock, making domestic extraction economically viable even against low-cost overseas labor.
- The US grid's bottleneck is not just generation capacity but the 50-year-old transformer fleet; solid-state transformers can provide granular voltage control and DC-native interfaces for solar and storage.
- Permitting reform is the single highest-leverage policy change: current NEPA timelines force capital into jurisdictions with functional regulatory systems, undermining domestic supply chain goals.
- Federal offtake agreements and grid investment tax credits must shift from supporting only generation assets to including transmission and distribution modernization to unlock the next increment of capacity.
Who should listen: Hardware engineers and systems architects working on physical infrastructure problems where software alone cannot solve the bottleneck.
Why This Matters
This episode maps directly to the hardware-software co-design thesis we track: the same vertical integration pattern that made Tesla a manufacturing anomaly is now being applied to the pre-competitive layers of the energy transition stack—minerals and grid infrastructure.
