Schneider and Emerson commit $10.3 billion to the software layer above plant control

The two prices are not measuring the same thing. Together they come to $10.3 billion of disclosed value ^[16], which is the number that will land in somebody's strategy deck this quarter. Schneider Electric's $3.1 billion buys Cognite, a company it did not own ^[1]. Emerson's $7.2 billion buys the AspenTech shares it did not already hold ^[2], settling ownership of a capability it had already consolidated rather than adding one ^[2]. Only the first tells you what an independent industrial AI asset fetches from an outside buyer.

Layer is a precise word here rather than a marketing one. In the hierarchy Gregory Shahnovsky sets out, regulatory control holds the process steady and advanced process control coordinates the interacting variables, while optimization above them decides where the plant should operate at all ^[8]. Commercial model predictive control has been doing the middle job in process industries since the 1970s ^[7], roughly half a century of installed practice that predates these deals ^[17]. AI enters above that, asking whether the target is still the right target rather than how to hold it ^[15]. A recommendation about set points still has to pass through the systems underneath, so the new layer inherits every constraint they enforce.

The binding constraint is not model availability. Conventional optimization leans on mathematical process models, where a simple one runs fast but is not accurate enough and a rigorous one describes the physics but becomes hard to maintain or expensive to solve online ^[9]. Shahnovsky's crude distillation example lists crude composition, furnace performance, exchanger fouling, utility costs, product values and downstream constraints among the inputs, with relationships that are nonlinear and change over time ^[10]. Machine learning can read those relationships out of operating data, but the data are late and unreliable in specific ways: lab results arrive hours after the conditions that produced the sample, instruments drift, and analyzers fail or stop correlating with the lab ^[11]. A model can then find a statistically strong relationship that makes no engineering sense ^[12], which is why he argues the useful version needs process context, engineering constraints, data validation and preferably a link to physical models ^[13], and why industrial analytics work is moving toward hybrid modeling that combines first principles with data-driven methods ^[14]. The scarce input in that description is process engineering judgment, and no acquisition supplies it.

Shahnovsky runs Modcon Systems, which sells process analytics and AI-driven optimization ^[6], so his framing of AI as a layer needing process context also describes his own revenue line. The deal values are not his to shape, and they run in the direction he claims, with Forbes tracing a wider pattern of acquisitions and partnerships across Siemens, ABB, Honeywell and Rockwell Automation ^[3]. The record so far covers only the sell side, not the buy side: what a plant pays for the layer, how many installations are running closed loop, and whether either acquisition delivers what was underwritten all remain unreported.

For an operator deciding this quarter, the practical question is narrow. It is which vendor's data platform the plant historian feeds, and on what terms that can be unwound. The longer question is whose judgment sets the target, because if the company supplying the control system also supplies the model that judges whether the target is right, a review that used to happen between two suppliers now happens inside one.