Calling AI The New Columbus Is Too Accurate To Be A Compliment – OpEd

Key Takeaways:

• The Columbus metaphor is useful only as a checklist, not a compliment: AI really opens search spaces (AlphaFold, GNoME), but value sits on an uneven stack of chips, cloud, energy, data work, and models—industry made 90%+ of notable 2025 models; many countries consume capability they cannot reproduce.
• Dependence starts as convenience (APIs, hospitals, ministries) then lock-in: switching costs, egress, retraining. “Exit sovereignty” means you can refuse or replace a system without losing core functions—not autarky, but governable interdependence (OECD, EU Data Act, Europe’s AI Factories).
• Before a critical dependency, ask who supplies inputs, who owns infrastructure, who captures rents, how expensive exit is, and who still has authority when vendor policy and national law collide—so discovery does not become a one-way port.

Frontier language flatters artificial intelligence as discovery. The harder geopolitical question begins after adoption: who supplies the resources, controls the infrastructure, captures the value, and still has the power

Artificial intelligence is constantly sold through geography. It is a new frontier, a new world, an uncharted territory waiting to be explored. Follow that language one step further and Christopher Columbus becomes an almost irresistible metaphor: the navigator who crossed an ocean, opened a new domain to the powers behind him, and changed the map.

As a compliment, the comparison is badly incomplete. As a warning, it is surprisingly useful.

To be clear, today’s AI economy is not equivalent to European conquest. The violence, dispossession, forced labor, and destruction associated with colonization should not be trivialized by turning them into a metaphor for cloud computing. The comparison earns its value only as a structural checklist. The Age of Exploration did not end when ships reached unfamiliar shores. The consequential questions came afterward: Who controlled the routes? Who supplied the resources? Who captured the value? Who became dependent on the new system? And who retained the power to say no?

Those are becoming some of the most important questions in global AI policy.

Discovery Is Real

The flattering part of the exploration metaphor is not imaginary. AI genuinely is opening search spaces that were previously too large, slow, or expensive for humans to navigate efficiently. AlphaFold has made more than 200 million predicted protein structures available to researchers and is used across more than 190 countries. Google DeepMind’s GNoME system predicted 2.2 million new crystal structures, including hundreds of thousands of potentially stable materials. Similar systems are accelerating weather forecasting, drug research, coding, logistics, and scientific hypothesis generation.

There is real discovery here. The mistake is stopping the story at discovery.

The systems producing these breakthroughs do not resemble lone navigators. They sit on top of a global industrial stack: advanced chips, hyperscale cloud infrastructure, data centers, model developers, proprietary interfaces, energy networks, human data work, and enormous pools of digital information.

That stack is not evenly distributed.

Stanford’s 2026 AI Index reports that industry produced more than 90% of notable AI models in 2025, while the United States and China continued to dominate top-tier model development. UN Trade and Development has warned that AI infrastructure, skills, investment, and governance capacity remain concentrated, with many developing economies positioned much more as users and data suppliers than as owners of the systems generating the highest returns.

This does not mean concentration is permanent. Open-source models, national compute programs, smaller models, and falling inference costs can redistribute capability. But it does mean the political economy of AI begins with an asymmetry: many more institutions and countries can consume frontier capability than can independently reproduce it.

The New Raw Materials

The colonial analogy becomes tempting because AI converts globally distributed inputs into highly concentrated capabilities.

Data is one input, but not the only one. Human labor remains embedded throughout the AI supply chain. UNCTAD’s Technology and Innovation Report notes that data annotation and related work in developing economies can be precarious and poorly paid even though those workers contribute to systems whose value is captured elsewhere in the stack.

Energy is another input. The International Energy Agency projects global data-center electricity consumption to roughly double by 2030, driven significantly by AI. Its 2026 update found that capital expenditure by five large technology companies exceeded $400 billion in 2025 and was expected to rise sharply again in 2026.

Compute is another. Advanced AI depends on scarce accelerators, data-center capacity, networking, and cloud platforms. The OECD describes AI infrastructure as highly concentrated at multiple layers, with only a few suppliers in important parts of the stack.

The point is not that every flow is exploitative. Many are commercial exchanges entered voluntarily and beneficially. AlphaFold, for example, has generated enormous public value and its protein-structure database is freely accessible. The more useful question is whether the relationship remains reciprocal as dependence grows.

Who supplies the inputs? Who controls the infrastructure that transforms them? Who captures the economic and strategic value? And can the supplier, customer, or country realistically change the relationship later?

Dependency Begins With Convenience

Technological dependence rarely begins with coercion. It begins with convenience.

A ministry adopts a foreign cloud service because it works. A hospital integrates an AI assistant because it saves clinicians time. A company builds internal workflows around a model API because developing an alternative would be irrationally expensive. A university teaches students on the tools employers already use.

Each decision can be sensible on its own.

Dependency appears later, when leaving requires data migration, retraining employees, rewriting software, rebuilding security controls, revalidating regulated workflows, replacing application interfaces, or accepting a temporary loss of capability.

The OECD has warned that structural dependence on integrated AI providers for models, cloud infrastructure, and APIs can create lock-in and raise switching costs. The European Union’s Data Act tackles a related problem directly: it requires measures to reduce obstacles to switching between cloud and other data-processing services because high egress charges, long procedures, and weak interoperability can make exit difficult.

That is the governance issue hidden inside the exploration metaphor. A frontier is easy to enter when someone else has already built the port. The important question is whether the port can later become a gate.

Exit Sovereignty

We need a better way to talk about technological sovereignty than asking whether a country has its own national AI strategy or whether a data center sits within its borders.

I use the term exit sovereignty for a narrower capability: the practical ability of an institution or country to refuse, replace, reconfigure, or leave an AI system without losing control of the essential functions that have become dependent on it.

Exit sovereignty does not require technological autarky. No serious economy should try to recreate every accelerator, cloud service, foundation model, software library, and data center domestically. Interdependence is not the same thing as subordination.

The test is whether dependence remains governable.

Can a government export its data and models in usable form? Can a hospital change providers without losing the workflows built around one vendor? Can a military or intelligence service continue a critical function if an external API is unavailable? Can a company audit the behavior that matters to it? Can national law meaningfully constrain a system if the decisive technical and contractual controls sit elsewhere? Can a country build or obtain an alternative before dependence becomes irreversible?

These are not hypothetical policy concerns. The OECD now treats the location and ownership of AI compute as matters with potential geopolitical significance. Stanford’s 2026 AI Index describes AI sovereignty as a defining feature of national policy. Europe is investing billions through AI Factories and proposed AI Gigafactories explicitly to expand access to computing infrastructure and strengthen technological sovereignty.

The lesson is not that every country should retreat behind a digital wall. It is that governments are beginning to understand that access to AI and authority over AI are different assets.

A Columbus Checklist for AI

If we insist on using the language of exploration, then we should use it rigorously.

Before an organization or state builds a critical dependency on an AI system, it should be able to answer five questions.

Who supplies the raw material? That includes data, human labor, energy, compute, and specialized knowledge.

Who controls the infrastructure? Not merely the visible application, but the model, cloud, chips, identity systems, interfaces, and update channels on which it depends.

Who captures the value? Productivity gains can be shared while rents, intellectual property, and strategic leverage remain concentrated.

How expensive is exit? Portability on paper is not enough if replacing the system requires years of retraining, reengineering, accreditation, and procurement.

And who retains authority? When provider policy, national law, organizational requirements, and model behavior diverge, who can actually decide what the system may do?

Those questions do not yield an anti-AI answer. They yield better contracts, interoperability requirements, open standards, diversified suppliers, domestic skills, fallback capacity, and clearer authority boundaries.

They also distinguish healthy interdependence from dependency that becomes politically consequential.

The Frontier Is Only the First Chapter

The problem with calling AI the new Columbus is not simply that Columbus is a morally complicated historical figure. It is that the compliment stops the story at exactly the point where the politics begins.

Discovery is the opening chapter. What follows determines the relationship.

Artificial intelligence may genuinely open new scientific and economic territory. That is reason for excitement. But the measure of a successful AI order cannot be how quickly the most capable systems spread or how many institutions become dependent on them.

It should also be whether the people, companies, and countries adopting those systems retain meaningful choices after adoption.

Who builds the ports? Who controls the routes? Who supplies the resources? Who captures the returns? And who can still leave?

The Columbus metaphor is too accurate to be a compliment if it encourages us to celebrate discovery while ignoring dependency.

Used honestly, it offers a better warning: the real measure of the AI frontier will not be how much new territory it opens. It will be how much authority its users retain after they enter it.