Physical AI: The Analog-Digital Boundary

Paradoxical Truth: The market assumes Physical AI will be won by better robot brains, larger models, and more GPUs. We believe the more investible bottleneck may sit one layer lower: the analog-digital boundary that allows intelligence to enter the physical world. As AI moves from generating tokens to controlling machines, the scarce resource becomes the ability to reliably sense, measure, and control physical systems – conditioning messy real-world analog signals, converting them into digital inputs, and turning digital decisions back into physical action. These are not commodity functions. They require scarce analog engineering talent, long product qualification cycles, and decades of accumulated design knowledge. If AI is moving from tokens to machines, then the companies that own the sense-measure-control layer may become the quiet toll roads of Physical AI.

AI is Moving From Data to Reality

The first wave of the AI super-cycle was largely digital. Models operated on information that had already been converted into data: text, documents, code, images and videos. The hard work of converting reality into data had already happened before the model ever saw the input. Physical AI changes that architecture. A robot, autonomous vehicle, drone, factory machine, battery system, medical device, or energy system does not begin with clean digital abstractions. It begins with physical phenomena: light, sound, pressure, current, voltage, temperature, acceleration, torque, vibration, and motion.

That shift changes the bottleneck. Digital AI turns data into intelligence. Physical AI must first turn reality into data, then turn intelligence back into action. With Physical AI, the scarce layer is the analog-digital boundary: the mixed-signal circuitry that allows machines to sense, measure, condition, convert, power, isolate, connect, and control physical systems. This is where continuous, noisy, real-world signals are preserved, cleaned, translated, and synchronized before they become machine-readable digital inputs. A model can only reason from the signal it receives; it cannot recover information lost to noise, poor conditioning, timing errors, power constraints, or thermal drift. Physical AI is therefore not merely a sensor problem or a compute problem. It is the interface layer where messy analog reality becomes trusted digital intelligence.