Competition: algorithmic collusion and the foundation-model stack

Lesson 2 of 5 in AI, Elections, Labor, Competition, and the Planet.

Competition law meets AI at two very different altitudes. Down low: algorithms that fix prices. Up high: whether the foundation-model market can be competitive at all. An expert needs both, because clients confuse them constantly.

Start low. In United States v. Topkins (2015) — the first algorithmic price-fixing prosecution — poster sellers on Amazon Marketplace agreed to use coordinating pricing software. Easy case: humans agreed, the algorithm merely executed. The frontier case is RealPage: the DOJ and a coalition of states sued in 2024, alleging that landlords feeding non-public rent and occupancy data into RealPage’s revenue-management software — which then recommended aligned rents back to all of them — amounted to concerted action through a shared vendor. This is the hub-and-spoke theory: competitors never speak to each other; the algorithm is the hub that does the coordinating. Cities responded faster than courts — San Francisco banned algorithmic rent-setting software built on non-public competitor data in 2024, with Philadelphia and others following. Check the current litigation posture before advising: settlements have been reshaping what the software may lawfully ingest.

The genuinely unsolved problem sits one step further out: autonomous algorithmic collusion. Laboratory studies show reinforcement-learning pricing agents learning supracompetitive prices with no agreement, no communication, and no human intent — they simply discover that punishment-and-reward pricing patterns maximise profit. US Sherman Act §1 requires an agreement; EU Article 101 requires concertation. Two algorithms tacitly converging on high prices may violate neither. Regulators know it: expect the doctrinal fight of the next decade to be whether using a self-learning pricing tool with foreseeable collusive tendencies is itself the culpable act.

Now the high altitude. Foundation models looked, for a moment, like a competitive explosion — until authorities mapped the stack: training-scale compute concentrated in three hyperscalers, frontier models concentrated in a handful of labs, and the two layers fused by billion-dollar cloud–model partnerships (Microsoft–OpenAI, Amazon and Google with Anthropic). These deals are not mergers — no control changes hands — so they slid past merger-notification thresholds. The response was study rather than suit: the UK CMA’s foundation-model reviews (2023–24) articulated principles of access, diversity, choice, and fair dealing; the FTC used its 6(b) study power to compel the partnership documents, reporting in January 2025 on exclusivity, compute dependency, and rights over model IP; the European Commission examined Microsoft–OpenAI and concluded it could not be reviewed as a merger absent a change of control. And in the landmark Google search remedies decision of 2025, the court explicitly shaped relief around the fear that yesterday’s search monopoly could roll forward into tomorrow’s AI-assistant market.

Key terms: algorithmic collusion, Hub-and-spoke (operating model), market study, quasi merger, vertical foreclosure

Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.