IXSAR Insights · Investment Perspective · 2026-08-13 · 2 min read · By IXSAR Capital

From Software to Silicon to Steel: Understanding the Physical AI Stack

Physical AI is not a single technology. It is a stack, and understanding where value accretes at each layer is the foundation of how I think about investing in this category.

When I diligence a physical AI company, I am always mapping where in the stack they live, how defensible that position is, and whether the market structure of their layer creates a durable business or a component that gets commoditized by the layer above or below. The physical AI stack is the most interesting multi-layer investment landscape I have seen in my career.

The silicon layer

NVIDIA Alpamayo 2 Super sits at the top of the silicon story for physical AI. The architecture is purpose-built: a tensor processing pipeline optimized for spatiotemporal inference, which is the core computation in any system that needs to understand the physical world in real time. The critical insight is that physical AI requires a fundamentally different compute profile than language AI: lower latency, higher determinism, and tighter integration with sensor hardware. NVIDIA is building the full stack from the chip to the model to the simulation environment, which is a moat-building strategy that very few companies can execute.

The silicon layer in physical AI is winner-take-most. The model layer is not. That asymmetry shapes where I place my bets.
John Gabriel, IXSAR Capital

The model layer

Below the chip and above the machine, the model layer is where I expect the most interesting competitive dynamics to play out over the next five years. Alpamayo 2 Super has established a strong baseline for generalist robotic control, but the history of foundation models suggests that fine-tuned specialists will outperform generalists on specific tasks once the generalist capability floor is high enough. The companies I am most interested in are those building proprietary data loops: every deployment makes their model smarter for the next deployment, in a way that a third-party model cannot replicate without their operational data.

The steel layer

The steel layer is hardware: the actuators, sensors, power systems, and mechanical structures that make a robot or autonomous vehicle physically capable. This layer has historically been where the capital went and where the value accreted. That is changing. As foundation models commoditize control, the differentiation in hardware is shifting toward manufacturability, reliability, and cost per unit at scale rather than technical capability at the margin. The hardware companies I back are the ones that can ship, not just demo.

Where the stack intersects

The most interesting companies in physical AI sit at the intersection of two or more stack layers. A hardware company with a proprietary training data loop. A software company with custom silicon integration. An infrastructure company with a model fine-tuned on its operational data. These cross-layer positions are where durable competitive advantage concentrates, and they are the profile I look for when I am evaluating a physical AI company for the IXSAR portfolio.

Explore

This site requires JavaScript for the interactive version. For inquiries, email partners@ixsar.com (investors) or companies@ixsar.com (founders).