Essay No. 49 · AI
Make the Machines Compete
Lower production costs do not guarantee lower prices. Competition policy must make buyer exit and genuine market entry enforceable.
The factory fired its last employee. Your replacement pump still costs a thousand dollars.
That is not evidence that automation failed; it can be rational business behavior. A seller has every reason to protect its pricing power, and once a firm eliminates its payroll, nothing about that cost saving compels it to hand the savings to the customer. Investment and upkeep do not disappear. The owner still has an incentive to extract the maximum amount the buyer is willing to pay before walking away.
Look at Nvidia. Its own 2026 annual report describes a full-stack platform built around GPUs, CUDA, libraries, frameworks, and APIs. The competitive choice can involve an entire software ecosystem, not merely a piece of silicon. Where moving that ecosystem is costly, the incumbent gains pricing power. That is the policy concern; the filing does not isolate how much of Nvidia’s margin comes from switching costs rather than performance, supply, or other factors.
Asking builders and sellers to voluntarily engineer easy exit ramps for their customers is naive. Rational founders have an incentive to build moats, not bypasses. If we want an automated economy to produce cheap, abundant goods instead of high-margin monopolies, regulation has to make open exits enforceable rather than optional.
The Thousand-Dollar Pump
Imagine two automated pump manufacturers. Their software buys materials, schedules machining, rejects defective seals, and quotes delivery dates without human intervention. Both factories can build an identical pump for $600. The established incumbent charges $1,000. A newcomer offers $750.
On paper, the newcomer has an easy opening: earn $150 before fixed costs while saving the buyer $250. Then the buyer tries to leave. The pump’s maintenance history is trapped inside the incumbent’s proprietary portal. To switch, the buyer has to reconstruct operating specifications, verify compatibility, re-qualify the part, and renegotiate a service contract. If that friction and risk cost $300, the $750 pump actually costs $1,050. Staying with the overpriced incumbent becomes the rational choice.
Now change the rules of the market. Mandate that the buyer owns a portable, standardized equipment record. Require open interface specifications and prohibit incumbents from voiding unrelated warranties when a buyer uses a qualified third-party part. Suppose those changes drop the switching cost to $20. Suddenly, the challenger’s pump costs $770 all-in, saving the buyer $230.
| Market regime | Build cost | Quoted price | Switching friction | All-in buyer cost | Market outcome |
|---|---|---|---|---|---|
| Incumbent baseline | $600 | $1,000 | $0 | $1,000 | Incumbent retains $400 before fixed costs |
| Challenger (locked records) | $600 | $750 | $300 | $1,050 | Buyer trapped; order stays put |
| Challenger (mandated exit) | $600 | $750 | $20 | $770 | Buyer saves $230; rival wins |
The math is illustrative, not a forecast, and it ignores fixed factory costs. The full integrated human-free firm remains untested. Still, it exposes the fatal flaw in assuming automation automatically lowers prices: a production cost advantage means nothing if switching costs eat it alive on the way to the customer.
Ten Shell Companies Are Not a Market
The standard Silicon Valley rejoinder is that we don’t need regulation because AI will make starting a company so cheap that an army of newcomers will overwhelm every tollkeeper. That confuses cheap back-office software with actual market entry. Even if one person with an AI stack can handle purchasing, scheduling, and support, that does not give a challenger spare factory capacity, put it in front of buyers, or prove its parts won’t fail.
Thousands of storefronts do not necessarily equal competition. A single parent operation can spin up ten brands selling the same product. Ten names under one controller are not ten independent rivals. Ownership must not become untraceable when a brand disappears after a safety failure. Meanwhile, a platform can collect fulfillment fees and advertising payments from all those brands while their apparent rivals remain under common control. Counting storefronts can hide the tollbooth instead of measuring competition.
Counting incorporation certificates or API keys tells regulators nothing. A real competitor requires independent ownership, physical capacity, verifiable compliance, and a direct route to the customer that doesn’t require paying tribute to a gatekeeper.
Faster Software Won’t Save the Buyer
Regulators also cannot assume that replacing human sales teams with pricing algorithms will spark fierce price wars. Fast software can punish an overpriced incumbent, or it can make a price cut effortless to detect and neutralize.
Real-world evidence already breaks the naive efficiency story. In their January 2021 working paper on German gas stations, Assad and coauthors infer algorithm adoption from pricing behavior and use headquarters adoption to address selection. They report higher margins in nonmonopoly markets; in their two-station result, margins rise when both stations adopt. They do not observe the algorithms’ learning process or establish that the software learned to collude.
Experimental research also shows why neither outcome is automatic. Eschenbaum and coauthors’ 2022 version 2 finds that collusive behavior can break down against different counterparts or contexts, but restricting the strategies available to algorithms can make it more robust.
Enforcers are confronting related risks. In March 2024, the FTC and DOJ argued in a hotel-pricing case that an alleged agreement can operate through a common algorithm provider without direct communications among competitors. That is the agencies’ legal position, not a finding that the hotels colluded. Separately, the FTC’s August 2026 proposed personalized-pricing policy addresses prices based on individual consumer data. The proposal says the extent of current use and consumer effects are not well understood. Personalized pricing and coordination among rivals are different mechanisms; one is not evidence of the other.
In an automated industrial market, structural lock-in can decide whether a cheaper offer reaches the buyer. The OECD Secretariat’s November 2026 background note identifies a related risk: accumulated histories and proprietary configurations can make purchasing agents themselves difficult to leave. Easy exit gives a qualified rival a better chance to win an order; it does not guarantee a price war. If records and platform tollbooths keep buyers trapped, removing employees may leave that pricing power intact.
What Competition Policy Actually Needs to Enforce
You cannot regulate an AI firm into feeling generous, and you shouldn’t try to cap its margins. Competition policy should start by enforcing three structural rules that prevent incumbents from turning switching costs into a private tax:
- Mandatory data and spec portability: Buyers—not vendors or platforms—must own their equipment histories, operating logs, and technical specifications in open, machine-readable formats. Exporting a dead PDF is not portability; any qualified competitor’s system must be able to ingest the record and bid on the replacement.
- Anti-tying and warranty protection: Incumbents must be barred from using proprietary software locks, artificial API restrictions, or punitive warranty voids to block compatible third-party hardware and parts.
- Beneficial ownership and real compliance: Marketplaces cannot hide behind an armada of disposable shell entities. Sellers must be tied to verified parent ownership with real liability for safety and quality compliance, stopping dominant operators from colonizing shelf space with fake rivals.
These rules need boundaries for trade secrets, privacy, and legitimate safety checks. Those protections must not become a pretext for blocking qualified rivals. The aim is to strip away the artificial friction that allows an incumbent to charge $1,000 for a $600 pump long after the last human worker has left the floor.
If policymakers measure economic progress by falling production costs or the sheer number of automated storefronts, they will miss the rise of the automated tollbooth. The whole point of building machines that can produce more for less is to make life cheaper and more abundant for everyone else—not to fund permanent monopoly rents. If we don’t regulate for open entry and buyer exit, the zero-employee economy will simply become the most efficient extraction machine ever built.
Sources
- Assad, Stephanie, Robert Clark, Daniel Ershov, and Lei Xu (January 2021). “Algorithmic Pricing and Competition: Empirical Evidence from the German Retail Gasoline Market.” Working paper.
- Eschenbaum, Nicolas, Filip Mellgren, and Philipp Zahn (2022). “Robust Algorithmic Collusion,” version 2.
- U.S. Department of Justice and Federal Trade Commission (March 28, 2024). Statement of interest in the hotel-room algorithmic price-fixing case.
- Federal Trade Commission (August 19, 2026). Proposed Enforcement Policy Statement Regarding Personalized Pricing, P034101.
- OECD Secretariat (November 2026). Competition and Agentic AI, background note DAF/COMP(2026)11.
- NVIDIA Corporation (2026). Form 10-K for the fiscal year ended January 25, 2026.