“The best of all monopoly profits is a quiet life,” observed the late British economist and Nobel laureate Sir John Hicks. But there’s no quiet life in artificial intelligence.
When Chinese startup DeepSeek recently demonstrated it could train world-class AI models using a fraction of the computing resources required by industry leaders, it revealed something crucial about competition in AI: dominance is more fragile than markets and regulators believed. Through clever engineering, DeepSeek claims to have achieved performance rivaling top firms OpenAI and Anthropic. And it did so while reportedly spending just $5 million on compute—a rounding error compared with the budgets of leading AI labs.
This type of breakthrough challenges conventional antitrust wisdom, which sees in AI markets a system of already-entrenched monopolies. Despite their massive scale, even the AI companies that once appeared unassailable now find themselves racing to keep up with breakthroughs from unexpected directions.
Consider Nvidia, where gross margins of greater than 90 percent on its AI chips have propelled the company to a $2 trillion market value. The conventional antitrust view sees such outsize profits as evidence of classic monopoly power, enabled by control of key inputs—in this case, the specialized chips and software needed to train and run AI models. Indeed, competition authorities recently have warned about firms attempting to “restrict key inputs for the development of AI technologies.”
But this static view misses how competition actually works in technology markets. The DeepSeek news illustrates a crucial reality: competition occurs across different industrial inputs into a technology and can often come from unexpected sources. While Nvidia dominates chips, innovative AI makers are already finding ways to reduce dependency dramatically on expensive specialized hardware. When an AI model can run efficiently on consumer-grade processors that cost less than $1,000, instead of the $40,000 industry standard, the pricing power of the hardware monopolist is reduced.
DeepSeek itself may not ultimately challenge today’s AI incumbents. But that’s precisely the point: the competitive threat rarely comes from where incumbents expect it to. Just as Palm and BlackBerry didn’t see the iPhone coming, so, too, may today’s AI leaders find their positions undermined by innovations they didn’t anticipate and players they never worried about.
Contrast this with how antitrust enforcers—desperate to prove themselves relevant on the hottest topics—view AI. A joint statement last summer from U.S., EU, and U.K. competition authorities reveals a strikingly different worldview. In their framework, AI is an unchanging set of “key inputs” that must be carefully controlled to prevent powerful incumbents from extending their dominance. This misses how technological innovation constantly reshapes competitive dynamics.
Similarly, former FTC chair Lina Khan declared in an exit interview last month that “technological inflection moments are ones where enforcers have to be especially vigilant.” This gets things exactly backward. At such points, regulators are least able to predict how competition will evolve or what interventions might help or harm it.
Khan reinforced this perspective in a recent New York Times op-ed, arguing that DeepSeek’s breakthrough signals the dangers of entrenched monopolies. But her own example undermines her argument: if today’s supposed tech monopolies were as dominant and unassailable as she suggests, DeepSeek’s success wouldn’t be possible. Rather than a “canary in the coal mine” warning of stagnation, it’s a clear signal that competition is alive and well—just not in the way that regulators expect.
Instead of fixating on controlling today’s “key inputs,” regulators should focus on maintaining the conditions that enable creative destruction. That means keeping markets open to new approaches and ensuring that incumbents can’t use their market power to squash emerging alternatives before they can mature.
The early AI era already shows promising signs of robust competition. Open-source models are proliferating. New hardware is emerging. Novel training approaches are reducing processing-power requirements. Each innovation creates opportunities for new entrants, while disciplining the pricing power of the current leaders.
The DeepSeek example shows why the aggressive regulator’s instinct for vigilance at key points is misplaced. No regulator could have predicted that innovation would slash the compute requirements to train top-tier AI models.
This doesn’t mean we should ignore legitimate antitrust concerns in AI markets. But we should recognize that technology itself often proves the most effective check on monopoly power. As another Nobel laureate in economics, George Stigler, once noted: “Competition is a tough weed, not a delicate flower.” Even in AI’s seemingly winner-take-all markets, competitive forces find unexpected ways to break through the cracks in dominant firms’ market power. That’s something for regulators to keep in mind as they consider how best to promote competition in these rapidly evolving markets.
Photo by STR/AFP via Getty Images
