Who will win the AI race? Why the Switzerland strategy could hold the key
In the AI era, it is widely assumed that the race will be won by those who build the largest models, deploy the fastest, or spend the most on compute. Yet this logic does not always hold – and it is a scenario that could realistically unfold in the evolution of AI, writes Mindsets managing director Mahmoud Deghaim.
The AI race is being shaped by two fundamental constraints that pull in opposite directions:
Physics pushes toward concentration. Training frontier AI models requires staggering amounts of energy, specialized chips, and capital that only a handful of organizations can marshal. The economics favor scale: bigger datacenters, more GPUs, exclusive energy deals. Physics wants centralization.
Policy pushes toward fragmentation. The US prioritizes free speech and open research. China demands censorship and state control. Europe insists on privacy and precautionary regulation. These aren’t technical preferences you can patch with an update. They’re deeply held values that must be encoded during training. Policy wants balkanization.
This creates an impossible tension. You can’t simultaneously:
- Build the most powerful AI (requires massive centralized infrastructure)
- Comply with every jurisdiction’s requirements (requires customization and constraint)
- Move fast enough to win the race (requires cutting corners somewhere)
The current players have chosen their positions: OpenAI races on capability and shapes regulation in its favor. Anthropic emphasizes safety but still competes commercially. Meta open-sources to commoditize competitors. China builds in relative opacity, unrestricted by public debate about alignment.
Every strategy accepts trade-offs. Every player has chosen which game to win and which to lose.
The Game Theory Problem
Here’s where it gets interesting. The AI race isn’t a single game: it’s multiple games happening simultaneously, with contradictory incentives.
The physics game is cooperative at the frontier. Everyone benefits when energy becomes cheaper or training becomes more efficient. Better algorithms, once published, help all players. There’s a natural pressure toward sharing research and coordinating on infrastructure.
The policy game is zero-sum. My values encoded in AI aren’t your values. My country’s regulations aren’t your country’s. If I deploy first, I set the norms. If you regulate me, you help my competitors. This creates a race dynamic where moving fast and loose beats moving slow and safe, even when everyone knows the slow approach is better for humanity.
This is a textbook prisoner’s dilemma. Both players cooperating (safe, coordinated development) is optimal. But each player individually benefits from defecting (racing ahead). So both defect, and we all end up in a worse state.
The pattern repeats geographically. The US, China, and Europe are locked in a three-way game where cooperation on global AI standards would benefit everyone, but each region’s incentive is to defect: to build AI that reflects its values and gives it strategic advantage. The result isn’t a single global AI ecosystem. It’s three incompatible systems; each confined to its regulatory bubble.
We’re not heading toward one AI to rule them all. We’re heading toward Splinternet, but for intelligence.
What Switzerland understood about neutrality
In 1815, Switzerland made a choice. Surrounded by great powers, it could have allied with one against the others. Instead, it declared permanent neutrality. Not weakness: neutrality backed by preparedness and defensive capability. The strategy wasn’t to win the great power competition. It was to be valuable to all sides precisely by not taking sides.
This created a unique position: the place where adversaries could meet, where money could flow across borders, where international organizations could operate without appearing to favor any bloc. Switzerland didn’t win by being the strongest. It won by being the most trusted.
The AI race needs its Switzerland.
The Switzerland of AI
Imagine an AI system optimized not for maximum capability or maximum speed, but for maximum compatibility. Not the biggest model, but the most adaptable one. Not the fastest deployment, but the most trustworthy one.
This would look different from current approaches:
Values-neutral by design. Rather than encoding one set of cultural values, the system would be built to be customizable by jurisdiction while maintaining core safety properties. Like international banking standards, the same infrastructure works in every country, but each nation applies its own regulations on top.
Efficient over massive. Instead of racing to train trillion-parameter models, focus on efficiency. Smaller models that deliver 80% of the capability at 20% of the cost. This sidesteps the physics arms race while remaining accessible to more players.
Regulatory compliance as a feature, not a bug. Most labs treat regulation as an obstacle. A Switzerland strategy would treat it as the product. Build the AI that can prove it complies with EU privacy rules, Chinese content restrictions, and US safety standards not because it’s weak, but because it’s architected for adaptability.
Open architecture, proprietary trust. The technology itself might be open source, but the governance, auditing, and certification systems would be the moat. Like the International Organization for Standardization where anyone can use the standards, but ISO’s value comes from being the trusted certifier.
The business model isn’t deploy the smartest AI. It’s provide the AI that regulators accept, enterprises trust, and citizens don’t fear.
Why this strategy could win
The counterintuitive thesis: in a fragmenting world, the player who works everywhere beats the player who dominates somewhere.
Consider the economics. If US-China decoupling continues, American AI companies lose access to Chinese markets. Chinese AI companies face barriers in Western markets. European AI, if it emerges, will struggle outside the EU’s regulatory framework. Each optimizes for their home territory and loses the rest.
But an entity positioned like Switzerland? It operates in all three blocs. It becomes the common infrastructure when adversaries need to interact. It’s the fallback when local champions stumble.
More importantly, it solves the coordination problem that makes AI governance so difficult. Current proposals for international AI treaties face a verification problem: how do you monitor compliance when capabilities are invisible and dual-use? You can’t count AI like you can count nuclear warheads.
A Switzerland approach changes the equation. Instead of trying to verify what labs are building, you build a system that’s auditable by design, trusted by all parties, and valuable precisely because it exists outside any single nation’s control.
This isn’t about being weak or small. Switzerland maintained armed neutrality and capability without aggression. An AI Switzerland would need technical excellence, just deployed toward different ends. The goal isn’t the frontier of capability. It’s the frontier of trust.
Who holds the key?
The big question then is: who could execute this strategy? Not the current leaders. OpenAI is too tied to Microsoft and US interests. DeepMind is British, owned by Google. The Chinese labs are state directed. They’ve all chosen sides.
There’s an instructive precedent already playing out: Apple’s approach to AI. While others race to deploy the most powerful models, Apple is quietly building something different. Apple Intelligence isn’t trying to be the smartest AI. It’s trying to be the most private, the most integrated, and crucially, the most adaptable to local regulations.
Apple runs models on-device to satisfy privacy hawks. It partners with multiple AI providers rather than betting everything on one. It designs systems that can comply with China’s data localization requirements while meeting Europe’s GDPR standards. The strategy isn’t raw capability. It’s global deployment through regulatory compatibility using iPhone as global distribution channel.
This isn’t perfect neutrality. Apple is still an American company subject to US law. But it demonstrates the core insight: in a fragmenting world, the player who can operate everywhere captures more value than the player who dominates somewhere. Apple understood this with hardware. The same logic applies to intelligence.
The question is whether anyone will take this approach further, building AI infrastructure that’s neutral not just in practice but in governance.
The candidates are unconventional:
A small, neutral nation with technical capability. Switzerland itself, Singapore, or the UAE: countries with technical talent, stable governance, and no imperial ambitions. They could position a national AI lab as the trusted global option.
An international consortium. Think CERN for AI funded by multiple nations, governed by treaty, existing outside any single country’s control. Politically difficult but not impossible.
A new open protocol. Not a company or model, but a standard just like TCP/IP for intelligence. Define the interfaces, safety properties, and auditing requirements. Let many implementations compete, all compatible with every regulatory regime.
The winner might not be an organization at all. It might be an approach or a way of building AI that prioritizes compatibility over capability, trust over speed, durability over dominance. A winner could be a leadership or a country.
The Deeper Lesson
We’re in the early stages of AI’s geopolitical fragmentation. The physics game favors concentration where only a few players can afford frontier development. The policy game favors fragmentation where every jurisdiction wants control.
The current players are choosing to play one game or the other. Race on physics and hope policy follows. Or slow down on capability and hope regulation protects you.
The Switzerland strategy is refusing to accept that framing. It’s recognizing that in a world of competing powers, the real moat isn’t dominance: it’s indispensability. Not building AI that’s too powerful to challenge but building AI that’s too useful to exclude.
History suggests this is possible. The internet routed around centralized control. GPS is American infrastructure that the world depends on. The dollar is the global reserve currency not because of American military might alone, but because it’s more trusted and more liquid than the alternatives.
AI’s Switzerland moment is still available. But the window is closing. Once the major powers fully commit to incompatible approaches, once the regulatory walls go up, once the physics investments lock in current architectures: it becomes much harder to build the alternative.
The question isn’t whether someone will try the Switzerland strategy. It’s whether they’ll try it soon enough to matter.
Because here’s the thing about races: sometimes the winner is whoever realizes they’re running toward the wrong finish line. The real prize might not be building the most powerful AI. It might be building the AI that’s still running when everyone else has collapsed from the sprint, exhausted, overextended, and trapped in jurisdictions they can’t escape.
In a fragmenting world, the center holds unusual power. And right now, the center is empty.
