Margins rose 38% where two German gas stations both bought pricing software

The identification sits in the gap between one adopter and two. Where a single German station automated, market-level margins did not move at all ^[5]. Where two competing stations both did, margins rose about 38% ^[4], so the whole measured effect belongs to the second adopter ^[1]. The economists found no meeting, message or agreement between the two ^[6]. The result was published in the Journal of Political Economy in 2024 and is among the first real-world measurements of a problem previously shown mostly in simulation ^[7]. The article does not give the base margin, so the 38% cannot be converted into cents at the pump.

Fortune calls the pattern consistent with each algorithm learning on its own that it earned more by backing off ^[8]. The learning is observed directly in the experiments. In the American Economic Review in 2020, four economists put reinforcement-learning algorithms into a standard model of repeated pricing, with no channel between them and one instruction, maximize profit ^[9]. The algorithms learned to charge above the competitive level and to enforce it, cutting when one undercut for share and returning to the higher level once it fell back into line ^[10]. Different costs, different demand and a changing number of competitors did not break the pattern ^[11].

Antitrust looks for conduct. Fortune sorts the ways competition fades into three: independently deployed algorithms that stop undercutting each other, one firm predicting how rivals will react and moving first, and competitors feeding data into a common provider ^[14]. Only the third has been reached in practice, with RealPage as the case ^[15]. Nessie is the second, and the FTC puts its excess profit above $1 billion while alleging Amazon paused the tool under scrutiny and restarted it ^[1][2]. Amazon says it was discontinued years ago ^[3].

I would put near-term enforcement risk low for an operator running an off-the-shelf pricing tool, because the first pattern leaves no agreement to subpoena and no vendor in the middle ^[14][16]. The exposure arrives if doctrine moves. Harrington has argued that competition law must be rethought for coordination that arises without agreement ^[13], and a rule written that way would land on pricing logs that already exist. The way this thesis breaks is selection: stations did not adopt at random, and if the early buyers sat in markets where margins were tightening anyway, the 38% describes where the software sold rather than what it did. The single-adopter comparison cuts against that, since those markets show no move at all ^[5].

The operational consequence is narrower than the legal one. Executives judge competition by the pressure they feel, and prices that hold alongside comfortable margins look like a market they have won ^[18]. Two algorithms that have stopped fighting produce the same picture, and Fortune's point is that the behavior may not appear on the dashboards used to judge whether the software works ^[17]. The harder failure is the system that does exactly what it was built to do, optimize margin, and reaches an outcome the company would struggle to justify in public ^[19].