This is not merely a price war between service stations. The lawsuit filed on June 22, 2026 in the federal court in Sacramento shines a light on a more subtle and pervasive mechanism: the use of artificial intelligence to steer pump prices and, according to the plaintiffs, keep them artificially high at the expense of consumers. The names of the defendants—BP, Circle K, Marathon Petroleum, 7-Eleven, Walmart, and Albertsons—underscore the scale of the dispute, which could reshape the rules for deploying algorithms in competition-sensitive sectors.

What the accusation describes: an algorithm to orchestrate price hikes

At the heart of the case is the use of an AI pricing tool that was either shared or interoperable across different chains. The claim does not point to an explicit cartel with clandestine meetings, but to a more modern form of signalling: a software system that, processing real-time data on demand, inventories, and competitors’ prices, recommended or enforced aligned increases, eliminating the healthy competitive uncertainty that should benefit consumers with lower costs. In effect, the technology would have acted as a clearing house, making any verbal agreement unnecessary.

The underlying technical issue is what analysts call tacit collusion by algorithm. In heavily digitized markets, machine learning models trained on historical series and sector data can spontaneously converge toward above-cost pricing policies, simply by maximizing profit without the need for explicit instructions. When multiple players rely on the same provider or the same framework, the risk of parallel behaviour multiplies.

Why algorithmic collusion is so hard to prove

For antitrust authorities, the boundary between harmless parallelism and illegal coordination has always been blurred. With AI, that boundary becomes fluid. Regulators must demonstrate that the software did not merely react to market dynamics but created the conditions for a stable alignment that harms consumers. Yet the technical complexity makes the internal workings opaque: a pricing model can be a black box whose behaviour emerges from millions of parameters and interactions, not from a simple cartel instruction. That is why the California lawsuit—still in its early stages—could set an important evidentiary precedent, pushing authorities to demand full algorithmic auditability.

Transparency and control: the on-premise card

It is precisely here that deployment choices take on strategic weight. Those who manage the AI infrastructure internally—in a self-hosted, on-premise scenario—maintain full visibility into the code, the training data, and the decision logic. Conversely, reliance on third-party cloud services, often with proprietary components, can limit the ability to demonstrate that one’s practices are compliant. In an increasingly watchful regulatory environment (consider the GDPR and the EU’s AI Act guidelines), the ability to reconstruct exactly how a pricing system processes inputs is not a luxury but a defensive necessity.

A model running on dedicated in-house hardware, with forced logging of every recommendation, price movement and market signal, provides a forensic trail that can be used in court. The trade-off, of course, is a higher TCO and a demanding technical burden that many organizations are not equipped to shoulder. That is why AI-RADAR offers comparative analyses and frameworks to assess when the investment in a local deployment is justified not only in terms of performance but also for data sovereignty and legal risk management.

Beyond the specific case: what it means for those building AI internally

Beyond the legal dispute, the story sends a strong signal to every team building business intelligence, dynamic pricing, and recommendation tools: control over the algorithm is no longer just an engineering matter, but a compliance asset. The choice between an open, verifiable model—possibly fine-tuned on proprietary data in a controlled environment—and an off-the-shelf SaaS solution can become decisive when an authority investigates. Those designing AI architectures for sensitive functions today should include explainability and governance mechanisms from the outset, not just for ethics, but to reduce exposure to the kind of challenges now unsettling California’s fuel giants.