Physical AI and the Hidden Cost of a 1% Assumption

In a previous piece, I explored how physical systems require rehearsal before deployment. This raises a deeper question—what assumptions are we making when we model them?

We’ve become extraordinarily good at modeling infrastructure. We can simulate turbine performance, grid behavior, even orbital trajectories with remarkable precision. The level of technical sophistication across engineering systems is undeniable.

And yet, large-scale infrastructure continues to underperform its modeled expectations with surprising consistency. That gap is often attributed to execution, market conditions, or unforeseen variability. But there may be something else at play. Something quieter. Something embedded in how we model the system itself.

Precision isn’t the same as understanding.

What the Data Is Already Showing

This isn’t a new observation.

The International Energy Agency has documented persistent deviations between projected and actual energy output due to environmental variability and system inefficiencies¹. McKinsey reports that large capital projects routinely experience cost overruns and performance shortfalls in the range of 20–50%². Bent Flyvbjerg’s work on megaprojects points to a similar pattern—systematic bias and model limitations consistently outweigh technical capability or effort³.

None of this suggests a lack of expertise. If anything, it suggests something more structural.

A Subtle Assumption

Most infrastructure models appear to share a common assumption: That variables can be understood in isolation. Weather is modeled separately from placement. Placement is modeled separately from load. Load is modeled separately from time.

Each component becomes increasingly precise. But the system, as a whole, is rarely evaluated under interaction.

The 1% Problem

This is where small deviations become economically meaningful.

A 1% variance in expected output—driven by siting, environmental exposure, or timing—doesn’t remain isolated.

It propagates. Across a utility-scale asset, this can translate into millions in annual variance. Over the lifecycle of a project, the impact compounds into tens or hundreds of millions, depending on structure and pricing. Across portfolios, it becomes systemic. From an engineering perspective, this behavior is expected⁴. From a capital allocation perspective, it is often averaged out.

More Data, Same Question

Over the past decade, the response has been to increase data.

Sensors. Telemetry. Real-time monitoring. Visibility into system performance has improved dramatically. But it raises an interesting question: If we have more data than ever, why do these deviations persist?

It may be because most systems are designed to answer: What is happening? What has happened? And in some cases: What might happen under assumed conditions?

But rarely: How does a decision behave when conditions interact and change?

Representation vs Consequence

This starts to look less like a data problem, and more like a modeling limitation.

Digital twins have advanced representation. We can now mirror physical systems with high fidelity. But many implementations still operate within segmented or static assumptions.

They show us what exists. They don’t always allow us to explore how decisions perform under evolving, interacting conditions.

Which may explain why capital is still often deployed based on approximations, rather than evaluated consequences.

Rehearsal Before Reality

In a previous piece, I explored the idea that in complex systems, rehearsal often precedes reality.

Aerospace missions are simulated before launch. Defense operations are war-gamed before execution. Robotics systems are trained in simulation before deployment.

Yet in infrastructure—where capital commitments are large and difficult to reverse—decisions are still frequently made without this level of rehearsal.

Not because the data doesn’t exist. But because the interaction between variables is difficult to explore before deployment.

Toward a Different Framing

This suggests a shift in how we might think about modeling. Not just as representation of systems. But as evaluation of decisions.

What if the goal isn’t simply to model assets…

But to understand how decisions perform inside environments that evolve? That would require moving beyond isolated variables, toward systems where: Space, time, environmental data, operational context, and human decisions interact continuously. Not as separate layers. But as a single system.

A New Layer Emerging

If that framing holds, it points to a new layer in the infrastructure stack. Not physical assets. Not data pipelines. Not dashboards.

But something closer to a decision layer. A way to explore how assumptions behave under interaction. A way to test sensitivity before it becomes cost. A way to evaluate outcomes not as single projections, but as distributions shaped by changing conditions.

Where We Are Exploring

At Chateauz™ , this is the direction we’ve been exploring.

Not as a visualization problem. And not purely as a data problem. But as a question of how to evaluate decisions inside environments that don’t remain static.

How placement interacts with environment. How time changes exposure. How conditions reshape outcomes.

Not to predict a single future. But to better understand the range of possible ones.

An Open Question

We are still early in this. But across energy, infrastructure, and even space systems, a similar pattern seems to be emerging:

The challenge may no longer be modeling assets. It may be understanding interaction. And more specifically: How decisions behave inside systems that are continuously changing?

If that’s true, then the advantage may not come from better models of things.

But from better ways to evaluate decisions.

References

¹ International Energy Agency (IEA) – World Energy Outlook reports (various years) 

² McKinsey & Company – Capital project performance studies 

³ Bent Flyvbjerg – Megaprojects and Risk: An Anatomy of Ambition 

⁴ Steven Strogatz – Nonlinear Dynamics and Chaos; Nassim Nicholas Taleb – Antifragile 

⁵ Andy Clark – Surfing Uncertainty