Nick Bostrom defined the singleton in 2006 as “a world order in which there is a single decision-making agency at the highest level,” capable of preventing any threats to its own existence and exerting effective control over major features of its domain. The concept is deliberately abstract. A singleton could be a democratic world republic, a totalitarian surveillance state, a single dominant AI, or a strong set of global norms with effective enforcement provisions. Its defining characteristic is finality: once established, a singleton cannot be challenged or replaced from within its domain. The concept encompasses two distinct pathways that are often conflated: a state or corporation that captures AI capability and uses it to establish permanent dominance, and an AI system that itself becomes the singleton, operating with goals that may not align with any human principal. These pathways have different mechanisms, different timelines, and different governance implications. The first is a political problem amenable to institutional design. The second is an alignment problem that current technical methods have not solved. Bostrom argues in Superintelligence (2014) that the first project to achieve a decisive strategic advantage through AI could parlay a temporary lead into permanent control, either by disabling competitors or by establishing governance structures that no subsequent actor can overturn. If a singleton goes bad, he notes, a whole civilization goes bad. The error-correction mechanisms that exist in a multipolar world (competition, emigration, external example, revolution) do not operate inside a singleton.
The scenario moved from philosophy toward policy in 2024 when Leopold Aschenbrenner’s “Situational Awareness” memo argued that whichever nation first achieves superintelligence gains a decisive strategic advantage, and that the U.S. government will inevitably nationalize the project once its implications become clear. Aschenbrenner’s framing treats the singleton as the default trajectory unless something interrupts it: competitive dynamics push toward a single winner, and the stakes ensure that winner will be backed by a nation-state. Anthropic CEO Dario Amodei has used similar language, writing in 2024 that “because AI systems can eventually help make even smarter AI systems, a temporary lead could be parlayed into a durable advantage.” The logic is straightforward: if AI capability compounds faster than it diffuses, multipolarity is a transitional state. The entity that reaches recursive self-improvement first (Shade #21) does not merely lead. It accelerates away.
The geopolitical landscape in early 2026 provides partial evidence for both the singleton thesis and its opposite. The Atlantic Council forecasts that the AI race in 2026 “will still be defined by a multipolar order,” with the U.S. and China as dominant players and middle powers closing the gap. The Trump administration’s AI Action Plan (July 2025) made it explicit policy to export the U.S. AI “stack” to third-party countries, a strategy that treats AI infrastructure as a tool of geopolitical influence. China has responded by doubling down on open-source AI to capture global market share, with DeepSeek and Alibaba releasing models that rival closed Western models on key benchmarks while costing nothing to download. Chatham House analysis notes that middle powers face a conundrum: full autonomy in AI is unrealistic, but dependence on either superpower creates strategic vulnerability. The current trajectory looks multipolar, not singular. But the current trajectory is also pre-recursive-improvement. The singleton question is whether multipolarity survives the transition described in Shade #21.
The strongest case against the singleton comes from three directions. First, the nuclear analogy. A 2025 analysis argues that Bostrom’s “decisive strategic advantage” concept can only hold if it invalidates the theory of nuclear deterrence formulated by Bernard Brodie in 1946. A superintelligence developed in a world that already has nuclear weapons, intercontinental delivery systems, and a mature logic of mutually assured destruction is not the same as one developed in 1944. Even a decisive cognitive advantage cannot neutralize second-strike nuclear capability without risking the destruction of everything the advantage was meant to control. Second, the diffusion evidence. DeepSeek’s January 2025 release of R1 under the MIT license demonstrated that frontier reasoning capabilities can be achieved at a fraction of the expected cost, and total model downloads shifted from U.S.-dominant to China-dominant during the summer of 2025. Open-source models are proliferating, not concentrating. Qwen has overtaken Llama in total downloads and is now the most-used base model for fine-tuning globally. The competitive dynamics of AI development in 2025-2026 look more like the internet’s diffusion pattern than like nuclear weapons’ concentration pattern. Third, the coordination constraint. Even superintelligent capability does not dissolve the institutional, logistical, and legitimacy requirements of exercising power. A system that outthinks every human on every problem still has to operate through supply chains, communication infrastructure, energy grids, and command structures that exist in the physical world. Power scales with organized coherence, not raw cognition.
These objections are serious, and the shade’s 25 percent likelihood reflects them. The response is that each objection assumes the current strategic environment remains stable during the transition to superintelligence. The nuclear deterrence argument assumes that a sufficiently advanced AI cannot find novel pathways around second-strike capability. This may be the strongest objection to the singleton thesis in its military form, because submarine-launched ballistic missiles are specifically designed to survive first-strike scenarios, operate on systems deliberately isolated from networked infrastructure, and function precisely because they do not require centralized command to launch. A cognitive advantage, however large, does not obviously neutralize a weapons system designed to function after the destruction of all centralized command. The singleton pathway through military dominance may be foreclosed by nuclear deterrence. The economic and informational pathways are less clearly constrained: a system that controls the majority of global economic activity, information infrastructure, and scientific discovery may not need military supremacy to establish unchallengeable governance. The diffusion argument assumes that open-source development continues to close the gap with frontier models, but the AI Futures Model’s December 2025 update projects that fully automated AI R&D, the step that would enable recursive improvement, is still several years away and may be dominated by a small number of labs with the compute, data, and talent to reach it first. The coordination constraint assumes that a singleton must operate through existing institutions, but Bostrom’s original formulation explicitly notes that a sufficiently advanced system could create new institutions or render existing ones irrelevant through surveillance, economic dominance, or informational control. Each counterargument is strongest in the current environment and weakest in the post-recursive-improvement environment, which is precisely the environment in which the singleton would emerge. This framing carries a circularity risk that deserves acknowledgment: if all evidence against the singleton can be dismissed as “pre-transition evidence,” the thesis becomes unfalsifiable. The honest position is that the singleton remains a scenario whose plausibility depends on empirical questions (how fast does capability compound? how resistant are nuclear deterrence and institutional structures to cognitive superiority?) that cannot be answered until the transition is underway.
The Anthropic-Pentagon dispute of February 2026 (detailed in Shade #20) provides a concrete illustration of how singleton dynamics operate even in the pre-superintelligence era. The Pentagon’s attempt to compel unrestricted access to Anthropic’s technology, and its designation of a domestic company as a “supply chain risk” using a statute reserved for foreign adversaries, demonstrated that the U.S. government is already treating AI capability as a strategic asset that cannot be permitted to have independent governance. The speed with which OpenAI replaced Anthropic in the Pentagon’s portfolio demonstrated that competitive markets do not produce governance constraints; they produce substitution. The structural lesson: in a world where multiple providers compete for government contracts, the provider with the fewest restrictions wins. That dynamic selects for concentration of capability under the entity with the most leverage, which in the current environment is the U.S. executive branch.
The 9-point governance dividend means that the character of any future singleton depends entirely on the institutional choices made before it emerges. Bostrom’s original paper makes this point explicitly: a democratic world republic could be a singleton, and a good singleton could maintain internal diversity, regional autonomy, and individual freedoms while solving coordination problems that multipolar systems cannot. The analogy is constitutional design: writing a constitution before a government takes power constrains that government in ways that writing one afterward does not. The EU AI Act, Anthropic’s Responsible Scaling Policy, and the FLI Statement on Superintelligence (Shade #21) are all attempts to establish constraints before the entity they constrain exists. Whether those constraints will prove binding depends on whether they are embedded in institutions with enforcement power, or merely in voluntary commitments that evaporate under competitive pressure, as the Anthropic-Pentagon episode suggests they might.
The deepest risk is not that a singleton emerges. It is that one emerges without anyone having chosen its values. The mechanisms described in Shade #21 (recursive self-improvement, automated AI R&D) could produce a system that is effectively a singleton before any governance framework is in place to shape its objectives. As the 25-researcher interview study from August-September 2025 found, predictions converge on the transition from AI assistants to autonomous AI developers but diverge sharply on what happens afterward. In that scenario, the singleton’s values would be whatever its training happened to produce, filtered through the commercial incentives of the lab that built it and the strategic interests of the government that backed it. That is not a scenario anyone, including the people building the technology, has endorsed. It is the scenario that emerges by default if the governance infrastructure described in Shade #20 does not materialize in time.
Key tension: Whether AI development is naturally monopolistic or naturally competitive is an open empirical question. The evidence from 2025-2026 points toward diffusion and multipolarity. The theory of recursive self-improvement points toward concentration and singularity. The question is which dynamic dominates once the capability curve enters the self-improvement regime, and by then it may be too late to choose.