The Bifurcation of AI in the Geopolitics of Power
In a single week of April 2026, three moves redrew the geopolitical board of artificial intelligence. The NSA activated Anthropic's Mythos Preview for intelligence operations while the Pentagon classified the same company as a "supply chain risk". Friedrich Merz opened the Hannover Messe demanding that the European Union…

In a single week of April 2026, three moves redrew the geopolitical board of artificial intelligence. The NSA activated Anthropic's Mythos Preview for intelligence operations while the Pentagon labeled the same company a «supply chain risk». Friedrich Merz opened the Hannover Messe by demanding that the European Union exempt industrial AI from the AI Act and quadruple its computing capacity by 2030. And OpenAI scaled GPT-5.4-Cyber to thousands of verified defenders, creating its own classification system for offensive cyber capabilities. Three scenes of one phenomenon: the governance of artificial intelligence is now the central battleground of 21st-century sovereignty.
The first scene reveals an internal fracture in the most powerful state on the planet. That the signals intelligence agency operates a model its own defense department considers risky exposes something deeper than fragmented bureaucracy. Two AI doctrines compete inside the same government: one that prioritizes immediate operational capacity and another that manages the risk of dependence on private suppliers. The resulting Nash equilibrium is inherently unstable — no agency has an incentive to yield, and the strategic paralysis is resolved case by case, contract by contract, with no coherent architecture. For those watching from the Caribbean and Latin America, the lesson is direct: if the United States lacks a unified AI doctrine for its own national security, expecting it to export a coherent framework to its allies is productive fantasy.
The second scene runs in a different register. Merz articulates what Brussels has avoided saying for two years: the AI Act, as conceived, penalizes European industry against competitors operating without comparable restrictions. Roland Busch's threat — Siemens would move investments to the United States and China — works because it is credible. At that same fair, NVIDIA showed how the convergence of CUDA-X, Omniverse and digital twins with OpenUSD cuts humanoid development from 18 to 7 months. Physical AI leaves the lab and enters the production floor. Europe faces a relentless regulatory thermodynamics: the energy it invests in regulating becomes competitive entropy if the pressure is not calibrated. The bifurcation Merz proposes — exempting industrial AI while keeping regulation of consumer AI — creates a two-speed system that finally acknowledges that the category «artificial intelligence» is too broad for a single regulatory framework.
The third scene is perhaps the most unsettling for what it prefigures. OpenAI is building a parallel security classification system: verified access, tiered capabilities, its own US$10 million fund for open source, free Codex for more than a thousand projects. A private company replicates functions that historically belonged to the State — classifying threats, granting access levels, deciding who can use which tool and for what purpose. Langdon Winner warned that artifacts have politics; here the artifact's politics include its own governance structure, complete with barriers to entry and mechanisms of exclusion.
What emerges from these three scenes is a map with three competing models. The American one fragments sovereignty between agencies and private suppliers who negotiate bilateral power. The European one tries to bifurcate regulation by sector to preserve competitiveness without abandoning the rights framework. The corporate one directly privatizes governance functions, creating parallel States of access and classification. None of the three is stable. All are in motion.
For countries like the Dominican Republic, the implication goes beyond academic analysis. The inputs trap operates here with almost mechanical precision: we consume imported AI — some US$300 million projected for 2025 — without taking part in defining the rules under which that AI operates. Every model we adopt arrives with its governance built in, its biases embedded, its classification decisions already made. The cognitive sovereignty we have discussed for years finds its most concrete expression this week: AI governance frameworks are being written now, in Hannover, in Fort Meade, in Sam Altman's offices. Whoever does not take part in that conversation will accept the rules others draft.
The Caribbean has a narrow window. AI governance is today the highest-value strategic input — it determines what can be built, who builds it and under what conditions. And here the inputs trap closes with a double turn: we import the technology and, with it, the rules that govern it. A national AI observatory that monitors these frameworks in real time, a position articulated in multilateral forums, regional alliances such as LatamGPT that generate negotiating critical mass: these are the minimum chips to sit at the table.
The alternative is familiar and predictable: wait for the rules to arrive bundled with the next software license, accept them without reading them, and discover years later that cognitive sovereignty was handed over in the fine print of a service contract.