Applied Computing wants to give oil and gas operators an AI model for the entire plant

London's Applied Computing wants to sell oil and gas operators something the industry has chased for decades: a single AI model that understands the whole plant. Not a dashboard. Not a point solution for predictive maintenance on compressor number four. The entire facility — thousands of sensors, miles of piping, the physics, the chemistry, the operator logs, the engineering drawings — all talking to each other in real time.

The pitch is seductive. Callum Adamson, the co-founder and CEO, says operators use less than 8 percent of the data their facilities already generate. The rest sits trapped in histrians, PDFs, and the heads of engineers who rotate out every few years. Orbital, Applied Computing's foundation model, proposes to stitch three dissimilar intelligence types together: a time-series engine that reads sensor streams, a physics-based layer that respects mass and energy balances, and a language model that ingests documentation and operator notes. The combination, Adamson argues, lets technicians simulate a valve change in the hydrocracker and see the ripple effect on the sulfur recovery unit minutes later — not weeks.

Speed is the product. The company claims Orbital can flag an anomaly, diagnose the root cause, and stress-test a proposed fix before the shift supervisor finishes coffee. If true, that compresses investigation cycles that currently burn days of engineering time. Energy savings and throughput gains follow. The market has noticed: stealth to double-digit millions in annual recurring revenue in under 18 months. Some "large, publicly listed" upstream, downstream, and petrochemical names are reportedly running it. Adamson won't say how many.

The $20 million Series A led by KBR, with Databricks Ventures participating, is not the story. KBR is. The engineering giant has baked Orbital into INSITE 3.0, its digital platform for energy projects, and is using it for ammonia production. That is distribution wrapped in credibility. Wipro is also a partner. A major U.S. upstream operator is in the works. A European oil major partnership announcement is teased for the coming weeks. These are not pilot programs — they are embedding events.

But the industrial software landscape is littered with the wreckage of "unified model" ambitions. AspenTech has owned simulation and modeling for upstream, refining, and chemicals for decades. AVEVA sells physics-based process simulation. Both have deep domain libraries, installed bases, and contracts that survive regime changes. Focused AI startups swarm the periphery — vibration analytics here, flare optimization there. Applied Computing enters this thicket claiming a horizontal layer that sits above all of them.

The technical architecture is the gamble. Large language models predict tokens. Time-series models forecast values. Physics simulators solve conservation equations. Fusing them without hallucination or computational explosion is a research problem, not an engineering one. Adamson says Orbital keeps physics and chemistry "in mind" while recognizing equipment constraints and operator activity. That phrasing does heavy lifting. The proof lives in whether the model respects a hard constraint — a relief valve set pressure, a metallurgical temperature limit — when the language component suggests a workaround that the time-series data says is profitable.

Data fragmentation is real. The 8 percent figure, whether precise or directional, captures a truth: historians don't talk to P&IDs, which don't talk to shift logs. But integration is a grinding slog of parsers, ontologies, and vendor negotiations — not a model capability. Applied Computing will live or die on the connective tissue it builds around Orbital, not the model weights themselves.

The revenue traction suggests operators are buying the promise. Desperation for margin improvement in refining and petrochemicals is high. Carbon intensity pressure is higher. If Orbital can genuinely show that a 2 percent feed rate adjustment cuts furnace duty by 5 percent without violating a catalyst temperature envelope — and do it while the operator watches — the price of admission becomes irrelevant.

KBR's ammonia deployment is the first public stress test. Ammonia loops are unforgiving: synthesis loop pressure, converter temperature, recycle gas composition, all coupled. A model that navigates that space in minutes earns the right to be taken seriously. The European major partnership will be the second. The U.S. upstream operator the third.

Applied Computing's bet is that the industry will pay for a reasoning layer that speaks sensor, physics, and human — simultaneously. The incumbents bet that their entrenched workflows and verified libraries will absorb any AI advance through acquisition or replication. The operators bet on whichever reduces the 2 a.m. phone call when the distillation column floods.

The next 18 months will reveal whether a foundation model for the plant is a product category or a category error. The sensors are waiting.