Most fusion coverage oscillates between breathless futurism and dismissive skepticism. This episode does neither. Brandon Sorbom, CSO and co-founder of Commonwealth Fusion Systems, walks through the specific engineering decisions that took fusion from a multi-decade government science project to a venture-backed company that has raised over $3 billion and is building a reactor designed to put net-energy electrons onto the grid.
The core unlock is high-temperature superconducting tape—specifically, rare-earth barium copper oxide (REBCO) wound into magnets that can generate 20-tesla fields at a fraction of the size and cooling cost of the low-temperature superconductors used in ITER. Sorbom explains why this single materials advance changed the scaling math: magnetic confinement quality scales with the fourth power of the field strength, so doubling the field shrinks the reactor volume by a factor of 16. That turns a $20-billion international megaproject into something a startup can actually build.
The conversation then moves to where the real frontier now sits: computational control and simulation. CFS is collaborating with Google DeepMind to apply reinforcement learning to plasma shaping, treating the tokamak’s magnetic coils as an action space where an agent learns to maintain stability against edge-localized modes and other instabilities. Separately, they are building a full digital twin of the reactor on NVIDIA Omniverse, which lets them run hardware-in-the-loop tests on control systems before physical commissioning. Sorbom is candid about what is still unsolved—tritium breeding ratios, neutron damage to materials, and the sheer manufacturing logistics of mass-producing magnets with nanometer-scale tolerances—and what is simply a matter of execution. For anyone who wants to understand fusion as an engineering discipline rather than a science-fair project, this is the clearest 90 minutes you will find.
Key Insights
- High-temperature superconducting (HTS) magnets are the single enabling technology that allowed CFS to shrink a net-energy tokamak from a multi-government-scale project to something that fits in a suburban parking lot, reducing volume by a factor of 40 compared to ITER-class designs.
- Fusion is not just inherently safer than fission—it is physically incapable of a runaway reaction because the plasma is so fragile that any perturbation instantly quenches it; the engineering challenge is not containment of an explosion but preventing a sneeze from killing the reaction.
- CFS is using reinforcement learning from Google DeepMind to control plasma shaping in real time, treating the magnetic coil currents as an action space where the agent learns to maintain stability against instabilities that have no closed-form analytical solution.
- NVIDIA Omniverse is being deployed to build a full digital twin of the reactor, enabling hardware-in-the-loop simulation where control algorithms can be stress-tested against synthetic sensor data before a single magnet is energized.
- The tritium breeding blanket is the unsolved materials-science bottleneck: every neutron from the D-T reaction must breed its own tritium fuel in a lithium blanket, and the neutronics of that process are computationally brutal to simulate at full scale.
- Sorbom frames the entire fusion effort as a manufacturing problem, not a physics problem—the physics of net-energy fusion is settled; the remaining work is supply-chain scaling, magnet mass production, and reliability engineering for components that must withstand neutron bombardment for years.
Who should listen: Engineers and technical operators working on complex physical systems where simulation fidelity, control-loop latency, and supply-chain scaling are the binding constraints to deployment.
Why This Matters
Fusion is transitioning from a physics experiment to a compute-and-manufacturing scaling problem, which is exactly the kind of shift that turns science projects into engineering companies—and eventually into infrastructure assets. We track this because the same AI-driven simulation and control techniques that CFS is applying to plasma dynamics are becoming the standard playbook for any hard-tech venture trying to compress development timelines.
