Beyond the Hype: The Real AI Race is an Infrastructure Play

The public discourse surrounding Artificial Intelligence is dominated by two compelling, yet ultimately misleading, narratives. On one hand, we hear of a geopolitical “AI arms race,” with superpowers like the United States and China locked in a fierce battle for technological supremacy. On the other, the focus shifts to a handful of corporate behemoths—OpenAI, Anthropic, Google, and a few Chinese giants—who seemingly hold the keys to humanity’s AI-driven future. Both narratives share a common, foundational premise: AI is a finite resource, something to be captured and controlled. And the current race is simply about who captures it first.

However, these prevailing stories, while not entirely devoid of truth, fundamentally misunderstand the terrain. They mistake the map for the territory. The real “AI race” isn’t a zero-sum game of national stockpiles or corporate monopolies. Instead, true leverage and strategic advantage lie in understanding and exploiting the intricate, globally distributed, and inherently porous infrastructure upon which all AI development depends. This is what we call “the gap”—a space where conventional power dynamics loosen, and agile players can achieve disproportionate impact.

Consider France’s Mistral AI. In just eighteen months, with a fraction of the resources of its Silicon Valley counterparts, Mistral has shipped models that rival labs fifty times its size. They didn’t engage in the “AI race” as defined by the dominant narratives. Instead, they exploited the gap: focusing on open-weight models deployable anywhere, leveraging European rhetoric for sovereignty while securing transatlantic capital from Microsoft. By the time the “who’s winning the AI race” commentators took notice, Mistral was already a formidable player, demonstrating precisely what operating in this gap looks like.

Deconstructing the Flawed Narratives

Let’s examine why the popular narratives fall short:

1. The National Arms Race: This view paints AI as a strategic asset, a digital arsenal that nations must accumulate and defend to avoid being “left behind.” Every country, under this logic, needs its own AI strategy to compete for dominance.
2. The Corporate Concentration: This narrative suggests an inevitable consolidation of AI power, where a few elite companies dictate the future, and everyone else simply buys access to their innovations. Human progress, in this view, is indexed to the roadmaps of a select few.

Both perspectives, while grounded in observable facts like national investments and corporate valuations, fail to account for the deeper, more complex reality of AI’s underlying infrastructure. They assume AI can be neatly contained and controlled, a notion that a closer look at the ecosystem quickly disproves.

The Ecosystem Lens: Where the Real Game is Played

To understand where the true leverage lies, we must look beyond the headlines and into the infrastructure itself—its layers, dependencies, and inherent escape points. This ecosystem lens reveals the incompleteness of both state and firm control:

• Chips: The foundation of modern AI. TSMC in Taiwan fabricates most advanced AI semiconductors, while ASML in the Netherlands produces the indispensable machines for etching them. NVIDIA designs the architectures but is entirely dependent on this non-American supply chain. Export controls can disrupt, but the ultimate lever runs through Taipei and Eindhoven, not solely Washington or Beijing.
• Capital: OpenAI, often touted as an “American AI” champion, has Microsoft as its largest historical investor, and its latest mega-rounds involve conglomerates like SoftBank (Japan) and sovereign wealth funds from Saudi Arabia and the UAE. “National champion” becomes a convenient label, not an accurate description of AI’s truly globalized financial backing.
• Models: The frontier of AI development is far more open than it appears. Meta’s release of LLaMA, a partially open large language model, and the emergence of DeepSeek from a Hangzhou hedge fund outside China’s state apparatus, demonstrate that the “moat” of proprietary models is constantly being challenged and circumvented.
• Talent: The architects of foundational AI technologies, such as the Transformer architecture at Google, came from diverse global backgrounds. Leading AI labs like Anthropic, OpenAI, and DeepMind are international melting pots. China’s top AI talent often trained in the United States, with some returning and others staying. Immigration policies, therefore, often shape AI capacity more profoundly than national R&D budgets.
• Data: AI models are trained on the vast, borderless expanse of the global internet. Platforms like Common Crawl do not check passports. The crucial labor of data annotation is distributed across Kenya, the Philippines, and Latin America. The human effort that makes AI legible is geographically dispersed, far from the headquarters that ultimately capture the financial returns.

These realities are not secret, yet they are often set aside because they disrupt the simpler, more convenient narratives of national and corporate control.

Coercion vs. Control: The Limits of Power

It’s important to acknowledge that state leverage is real. Export controls can impact corporate bottom lines, tariffs can disrupt supply chains, and antitrust laws can force divestitures. Security concerns regarding dual-use AI technologies are legitimate. However, coercive power, while disruptive, does not equate to absolute control.

The state can disrupt the ecosystem, but it cannot cheaply re-territorialize it. A decade of American pressure on Apple, for instance, resulted in only slow and partial supply chain diversification to India and Vietnam, with no significant return to domestic manufacturing. The deeply entrenched Shenzhen ecosystem, built over decades, simply cannot be willed back inside national borders by decree.

AI infrastructure will face similar challenges. Every “sovereign AI” initiative quickly encounters this wall. Funding national compute capacity, for example, often means purchasing chips from NVIDIA (US) manufactured in Taiwan. National researchers inevitably collaborate with international peers. While a state can restrict chip exports, it cannot erase knowledge that has already diffused globally.

Similarly, firms cannot fully capture these intricate stacks. While science is distributed, the capital required to industrialize it is brutally concentrated. Knowing how Transformers work is a global public good; having the $10 billion cluster to train a frontier model is a localized advantage. The ecosystem lens doesn’t dissolve concentration but highlights its incompleteness. Each time a frontier lab claims an unassailable “moat,” a competitive model often emerges for a fraction of the cost, challenging the narrative of inevitable lock-in.

A crucial caveat: this argument primarily applies to the application layer, where most real-world decisions are made. At the bleeding edge—training runs costing tens or hundreds of billions—the “moat” is a wall of silicon and electricity that only a few entities can build. The gap exists primarily where you deploy, not necessarily where you train.

This irreducible slippage—where states can disrupt but not capture, and firms can concentrate but not lock down—is structural. It is not going away. The question, then, is how to act accordingly.

The “Gap Strategy”: Acting Accordingly

The most successful players in the AI landscape are those who have instinctively understood and leveraged these gaps:

• The UAE: Rather than choosing sides in the perceived “US versus China” binary, Abu Dhabi strategically played both. They welcomed American hyperscalers to build data centers on Emirati soil, engaged in Chinese partnerships where beneficial, and poured sovereign wealth into Silicon Valley startups. The UAE didn’t choose a bloc; it exploited the interstitial space to become an indispensable node that neither side could ignore.
• NVIDIA: Caught between American export controls and insatiable Chinese demand, Jensen Huang didn’t passively comply or overtly resist. He innovated by developing different chips for different regimes, lobbied Washington while simultaneously restructuring supply chains, and bought time as the regulatory landscape shifted. Infrastructure, it turns out, has its own logic, and even the most powerful states must negotiate with it.

These aren’t exceptions; they represent competent strategy when one discards the restrictive “race narrative.” However, the gap is not without its perils. Companies like ZTE and Huawei learned that playing both sides can lead to exclusion from critical global systems. Exploiting the gap carries existential risk for corporations without sufficient countervailing leverage. States like the UAE possess this leverage; most corporations do not.

Practical Implications of the Gap Strategy

So, what does this understanding mean in practice for various stakeholders?

• For Public Compute Capacity: The question for nations like France or Germany isn’t whether to build a national supercomputer simply to “compete.” It’s about purpose. A machine designed for open access—serving researchers, startups, public institutions, and small countries pooling resources—serves a fundamentally different logic than one framed as a national trophy. Finland’s LUMI supercomputer, structured as a shared European resource, exemplifies working the gap effectively.
• For Open Models: Meta’s release of LLaMA, while strategically motivated to commoditize complements to its advertising business and set architectural standards, also created real options. Open weights, though not altruistic gifts, allow entities like Mistral and Hugging Face to build on these foundations in ways that defy the concentration narrative. The key is to build effectively on the substrate while retaining the capacity to adapt and move.
• For Capital Investment: Sovereign wealth funds and public investors inject substantial capital into AI. Most of this flows to incumbents without conditions. However, public investment, such as Bpifrance’s investment in Mistral, can come with explicit expectations for European presence and open-weight releases. Public money can and should demand commitments that serve collective strategies, such as open-source requirements or access provisions. The current lack of such conditions is a policy choice, not an inevitability.
• For Cross-Border Coalitions: Nations like India and the African Union are not passively waiting to be assigned a position in the US-China binary. They are actively negotiating compute deals with both sides, building domestic capacity, and positioning themselves as alternative talent hubs. These aren’t peripheral actions; they are astute exploitations of the structural slippage that dominant narratives often ignore.

Conclusion

Ultimately, the question is not “who wins the AI race,” but rather, “what can be built while everyone is running a race with no clear finish line?”

For corporations, AI vendor strategy is a dependency decision, not merely a technology choice. The critical question isn’t “Which hyperscaler has the best model this quarter?” but rather, “What level of lock-in is acceptable, what alternatives can we cultivate, and what leverage do we retain when the next supply chain shock hits?” CTOs who diversified their compute suppliers before the 2023 GPU shortage understood this; those who scrambled did not. For many large non-tech enterprises, the initial script of signing with a major vendor is giving way to a silent pivot towards “gap strategies” as data sovereignty concerns, escalating API costs, and memories of early cloud lock-in resurface.

In policy, the real choice isn’t a binary of “regulate or lose to China.” It’s about what infrastructure to fund, under what conditions, and for what access. Every subsidy given to a “national champion” without clear access requirements is a missed opportunity to build public capacity and collective strategic advantage. The race framing obscures this; the ecosystem lens makes it a visible, conscious choice.

Those advising governments and corporations alike would do well to stop asking who is ahead and start mapping where the true leverage points are: compute and processing capacity, yes; access to sovereign and inexpensive energy, yes; but also standards bodies, training data governance, immigration policy, and public procurement rules.

The gap between the simplistic “race narrative” and the complex reality of AI infrastructure is vast and enduring. It is within this gap that political and strategic room to maneuver truly lies. The challenge, and the opportunity, is whether we see it, each in our own context, and what we choose to do there.