The AI Race: Winning Is Inconsequential

The paradox of AI abundance—and why quantum computing may be a very different race

The most expensive race in history may have no lasting winner

America treats artificial intelligence as a strategic contest. Export controls constrain access to advanced chips in China; technology companies compete to build vast data centers; politicians speak of leadership as though there were a single finish line. Yet the surprising outcome of this competition is that advanced AI is becoming more accessible, not less. Every attempt to build an unassailable lead seems to inspire a cheaper, faster or more open alternative.

Consider an earlier technological revolution. In 2011, venture capitalist Marc Andreessen wrote a line that became the defining slogan of an era:

“Software is eating the world.” — Marc Andreessen, 2011 [1]

He was right about software transforming industries. But software also offers a cautionary tale about where the economic rewards settle. Linux became indispensable without a single company owning Linux itself. The internet became essential without one company owning the internet. AI may be headed toward a similar destination: a universal capability that creates enormous economic value while frustrating attempts to monopolize its core intelligence.

The competitors keep distributing their weapons

Meta helped make powerful open-weight language models mainstream. Chinese teams behind DeepSeek, Qwen and Kimi have supplied another striking example: models capable of competing in significant tasks can be released for developers to download, modify and operate on their own infrastructure, subject to their respective licenses. An enterprise need not surrender every document to a proprietary cloud provider merely to benefit from modern AI.

There is an irony in NVIDIA’s position. It sells the scarce infrastructure on which much of the AI economy runs, yet it also releases software libraries, research tools, models and application frameworks to accelerate adoption. From robotics to computational science, free or openly available tools can stimulate demand for the expensive hardware beneath them. What looks like generosity at the software layer can be rational economics at the silicon layer.

Imagine two competing factories. One spends billions inventing a manufacturing process; the other publishes a working blueprint that the entire industry can adapt. Customers celebrate, but the original inventor must now explain why its product deserves a lasting premium. That is increasingly the strategic predicament of frontier-model developers.

DeepSeek: the anecdote that changed the arithmetic

DeepSeek’s rise provided a particularly vivid challenge to the assumption that the richest laboratory must always dominate. Its model releases and engineering disclosures triggered questions about the cost of training and serving capable AI. In March 2025, Reuters reported DeepSeek’s estimate of $87,072 in daily GPU rental costs against $562,027 in hypothetical daily inference revenue—a theoretical cost-profit ratio of 545% [2].

That is not evidence of an 85% realized corporate profit margin. DeepSeek explicitly cautioned that actual revenue was substantially lower because some access was free, some usage was discounted and only part of its services was monetized. The anecdote matters not because it proves spectacular realized profits, but because it punctured a powerful narrative: that frontier AI must remain intrinsically expensive and therefore permanently exclusive.

A breakthrough in inference efficiency does not merely help one company. Once the techniques are published, the entire ecosystem can learn from them. Yesterday’s competitive secret becomes tomorrow’s engineering baseline.

A moat that evaporates when copied

This is the heart of the argument. If the key techniques are published, model weights are downloadable, and alternatives become good enough for most enterprise work, what exactly does it mean to “win” AI? Benchmark leadership may last weeks or months. A premium API can face an open-weight substitute. A clever feature can be reproduced by a team armed with increasingly capable coding agents.

The claim is not that every AI business will fail. Distribution, trust, data rights, product design, customer relationships, proprietary workflows and physical infrastructure can remain powerful moats. Nor are open weights equivalent to fully open training data or reproducible training pipelines. The narrower but more consequential proposition is that general-purpose model intelligence itself may be a poor foundation for permanent monopoly.

Regulation can shape this contest, too. Incumbents may favor rules that smaller competitors find expensive to satisfy; supporters may argue the same rules are necessary for safety and accountability. Either way, regulatory advantage is not the same thing as technological superiority. A race decided by access, compliance or distribution is no longer simply a race to build the smartest model.

Quantum computing is a different kind of contest

Quantum computing presents a compelling contrast. Publishing an algorithm does not automatically give a rival a working quantum processor. The difficult parts may include fabrication, cryogenics, control electronics, calibration, error correction, integration and the sustained operation of reliable systems. These are embodied capabilities, not merely code that can be copied from a repository.

Richard Feynman captured the deeper motivation decades ago:

“Nature isn’t classical, dammit.” — Richard Feynman, 1981 lecture [3]

His observation was about simulating physical systems, not a prediction of commercial supremacy. But it illustrates why quantum computing could create a qualitatively different competitive frontier: the ability to manipulate physical processes that conventional machines struggle to represent efficiently. If fault-tolerant systems eventually deliver useful advantages at scale, the organizations controlling reliable hardware and its operational know-how may possess barriers to entry that are difficult to erase by publishing a paper.

This outcome is not guaranteed. Quantum technologies remain technically demanding, commercial advantage is application-dependent, and several hardware approaches may coexist. Open research, cloud access and competing suppliers could also reduce concentration. Nevertheless, the possibility of durable differentiation appears stronger where value depends on hard-to-reproduce physical systems rather than rapidly diffusing model weights.

The race worth asking about

The most provocative conclusion is not that AI is unimportant. Quite the opposite: AI may become so important, ubiquitous and affordable that no single country or company can claim enduring ownership of it. Winning a particular AI benchmark may be commercially useful without amounting to winning the future.

Quantum computing may offer a rarer prize: a difficult-to-replicate capability rooted in physics, manufacturing and systems engineering. The slogan that its winner could dominate the next century is rhetoric, not an established forecast. Yet the strategic question it raises is legitimate. In a world where intelligence becomes abundant, the enduring advantage may belong to those who can do something physically extraordinary that everyone else still cannot.

Perhaps the real mistake is confusing the race to popularize a technology with the race to own its hardest-to-copy capabilities.

Sources and quotation notes

[1] Marc Andreessen, “Why Software Is Eating the World,” August 20, 2011. https://a16z.com/why-software-is-eating-the-world/

[2] Reuters, “China’s DeepSeek claims theoretical cost-profit ratio of 545% per day,” March 1, 2025. https://www.reuters.com/technology/chinas-deepseek-claims-theoretical-cost-profit-ratio-545-per-day-2025-03-01/

[3] Richard Feynman, “Simulating Physics with Computers” (1982), quoted in Nature Physics, “Quantum simulation” (2012). https://www.nature.com/articles/nphys2258