Editor's note: This is the first of a new five-part series by Colin Masson of ARC Advisory Group where he'll take a no-nonsense look at how the latest wave of AI collides with the physical laws of the plant floor.

He’ll strip away the hyperbole about AI, deconstruct the vendor marketing narratives, and provide plant managers, operations executives, and enterprise architects—part of our core audience—with actionable blueprints for governed and profitable implementations.

If you have followed my research at ARC Advisory Group over the years, you know I maintain a strict personal rule: I do not cover incremental frontier model releases.

Let consumer tech pundits breathlessly debate whether the latest hundred-billion-parameter foundation model writes slightly better poetry, passes a high school history exam, or shaves milliseconds off a conversational chatbot response.

See also: Smart Industry's annual State of Initiative Survey is open!

In the uncarpeted OT and cyber-physical engineering worlds where we live, in the plant, those benchmarks have historically mattered very little.

On the factory floor, in the chemical processing plant, and across multi-tier industrial supply chains, we don’t care if an AI can compose a sonnet. We care whether it understands the non-linear physics of a chemical reactor, respects the thermal limits of a motor drive, and won’t violate safety interlocks at 2 a.m. on a Sunday.

Huang pointed out that while a single one-gigawatt facility requires upward of $60 billion in capital expenditure, it has the potential to generate $50 billion in annual rental revenue as global token generation scales. Let that sink in.

See also: As AI adoption matures, data center controversies move to the fore

When hardware providers and cloud hyperscalers pour tens of billions of dollars into multi-gigawatt facilities drawing enough power to run a major metropolitan city, they are trapped in an uncompromising financial imperative: They must generate and sell astronomical volumes of tokens to service their debt, justify their CapEx, and hit Wall Street revenue projections.

The cloud hyperscaler business model requires continuous, high-frequency token consumption. And desktop-actuating agents are the ultimate token-consumption engine.

When a manufacturer hooks up continuous reasoning agents to cloud endpoints to monitor operational lines, your plant operating budget becomes the variable revenue stream funding the amortization of Silicon Valley’s infrastructure debt.

The “scaffolding discovery”: The harness dominates the weights

The most pivotal insight from independent audits of frontier models directly deconstructs the claim that raw neural weights represent an all-knowing artificial brain.

See also: Navigating financial anxiety around paying the bill for industrial AI

Consider standard benchmark evaluations designed by cognitive scientists to measure an AI’s ability to adapt to novel, unfamiliar environments without relying on memorized training data. In promotional launch briefs, frontier models achieved stunning, near-saturation scores of 98.6% to 99.9%.

However, independent technical replications revealed an immense empirical divide:

• When evaluated through standard, unassisted, stateless API calls, model scores plummeted to 62.7%.
• Over 36 percentage points of benchmark performance were generated entirely by an external software scaffolding harness—a deterministic software wrapper that actively pruned context windows, managed structured scratchpads, and orchestrated stateful tool loops.

In industrial engineering terms, that is like placing an electric motor on a dynamometer, discovering that it only delivers full-rated torque when coupled to a massive external gearbox, and then crediting the motor alone for the mechanical breakthrough.

See also: What to do when ‘rip-and-replace’ is financially impossible

The cognitive breakthrough belongs to the architectural scaffolding, not the latent weights of the neural model. For industrial enterprises, competitive advantage and operational reliability do not come from renting generic foundation model weights from hyperscalers.

They come from your proprietary scaffolding: your industrial data fabric, your semantic knowledge graphs, and open protocol connectors like CESMII’s i3X and the model context protocol.

Geometry is not physics

This brings us to the much-lauded demonstrations showing frontier models opening professional electrical CAD suites (like KiCad), placing components, and routing printed circuit board traces from scratch.

Watching an AI model route a circuit board or reconstruct a 3-D CAD mesh is impressive software automation. But this is precisely where plant and automation leaders must draw a hard line between software dexterity and cyber-physical truth.

See also: How industrial software operationalizes physics-informed digital twins

Drawing a copper trace on a computer screen does not mean the model understands electromagnetism. In high-speed hardware engineering, physical reality is governed by Maxwell’s equations, transmission line physics, and thermal dynamics:

• Trace geometry dictates characteristic impedance, high-frequency skin-effect losses, dielectric dissipation, and crosstalk coupling.
• Return-path discontinuities, via transition inductances, and ground-plane slot resonances can transform an otherwise clean visual layout into an unintentional radio transmitter that fails electromagnetic compatibility certification.
• Power distribution networks require precise low-inductance decoupling, copper balancing, and thermal dissipation modeling to prevent board warpage during high-temperature reflow soldering.

Frontier models approach trace routing as a spatial puzzle, moving the mouse to satisfy visual clearances dictated by the software’s design rule check. When an error appears, the model reads the screen text and tries a different coordinate until the visual warning clears. That is heuristic trial and error within a graphical interface—it’s not physical foresight.

Without tight integration into deterministic finite-element solvers, computational fluid dynamics, and SPICE simulators, geometric layout automation remains an assistive draft, never an autonomous engineering sign-off.

The Big Tech consensus: When model-builders beg for the brakes

If industrial IT and OT leaders think analysts are being overly conservative regarding uncontained AI agency, look at what the creators of frontier AI declared during their historic public alignment:

• Dario Amodei, CEO of Anthropic, published his urgent manifesto: “We Must Pace the Frontier,” calling on the industry to slow scaling until safety evaluation and containment frameworks catch up.
• Sam Altman, CEO of OpenAI, immediately backed Amodei's call, stating: “I agree with Dario that we need to pace the frontier.”
• Satya Nadella, chairman and CEO of Microsoft, demanded deliberate pacing, emphasizing that advanced systems must never resist human shutdown or correction, and that enterprises must retain sovereign control over their own models rather than yielding governance to hyperscalers.
• OpenAI’s Chief Scientist Jakub Pachocki warned that machine intelligence is expanding into an "alien mind" whose internal latent depth humanity is ill-prepared to audit.

When the CEOs and chief scientists of the world's most powerful AI labs publicly concede that autonomous machine intelligence is advancing faster than human ability to safely control it, industrial operators cannot look the other way.

See also: Roadmap to physically intelligent industrial operations

Waiting for a Silicon Valley truce or a government regulatory mandate is an operational fantasy. Market competition will not permit an unforced pause. If you grant an uncontained, cloud-tethered agent write access to physical control networks, you are taking on operational liabilities that the model creators themselves refuse to underwrite.