Applied Computing, a London-based startup founded in 2023, has just raised $20 million in a Series A round to build a foundation model tailored for refineries. The round was led by engineering giant KBR, with participation from Databricks Ventures. Behind the funding announcement lies a sharper question: why does an industrial plant generating millions of data points every day – temperature, pressure, velocity, viscosity – use less than 8% of that data in decision-making?

The answer is not merely technical but structural. Refinery data is born, lives, and dies within the physical perimeter of the plant. It is noisy, intermittent, and scattered across thousands of sensors stretched over kilometers of piping and distillation columns. Moving it to the cloud for training or inference with a generalist LLM is almost never feasible: latency, bandwidth costs, regulatory compliance, and security risks make it a non-starter. That’s the gap Applied Computing aims to fill.

The 8% trap

The numbers are striking: a single refinery can host more than 10,000 sensors, yet operators rely on a tiny fraction of those signals. The reason is that traditional analytics tools – static dashboards, DCS systems, fixed rules – cannot untangle the complex correlations among flows, pressures, and temperatures that shift in real time. A foundation model trained specifically on the petrochemical domain can capture those interdependencies, predict failures, optimize distillate yields, and cut downtime. All without a single bit leaving the premises.

To achieve that, though, local deployment is non-negotiable. Latency cannot afford cloud round-trips: an overpressure alarm must trigger in milliseconds, and connectivity at remote sites is often spotty. It’s the perfect use case for self-hosted AI, where the model runs on edge nodes or on-premise servers, tapping directly into sensor streams.

Industrial sovereignty injection

The fact that a corporate fund like Databricks Ventures is participating in such a deal says a lot. Databricks is synonymous with cloud data lakehouses, but here it’s investing in a project that by its very nature will push compute capacity to the edge, far from centralized data centers. It’s a recognition that high-value industrial data remains anchored to its place of origin, and that monetizing it without relocation requires hybrid architectures and vertical-specific models.

KBR, for its part, brings decades of process engineering expertise: it knows the peripherals, the protocols, the tolerances. Together, they are betting that the next leap in efficiency won’t come from a generic chatbot, but from an LLM that truly “understands” refining, and does so without violating data sovereignty.

Who wins and who loses

For makers of edge inferencing hardware – from NVIDIA Jetson modules to FPGA solutions – this is music to the ears. An entire vertical sector beginning to demand on-premise models at scale could fuel a new generation of rugged chips, certified for ATEX environments and optimized for low power. Winners also include those building orchestration for hybrid deployments: tools to update models remotely, monitor data drift, and manage versioning without halting production.

Losers are those who bet that heavy industry would pour its operational secrets into the public cloud. Resistance has always been there, but now it’s finding support in technology that makes on-site inference not just possible but economically rational. The structural message is clear: the AI race isn’t only about ever-larger models, but about models ever closer to the data. Applied Computing is just the latest name, but the shift is already underway.