Closed or open? AI capability is concentrating in a small number of companies, chip manufacturers, and state actors. Open-weight models are distributing it outward. Neither trend is winning outright — and the debate about power in AI is actually three separate questions that rarely get separated.
Craig Stanley Studio · The Debate · Issue 04 of 09 · craigstanley.work
The power debate in AI is usually conducted as if it were a single question. It is at least three: who controls the models, who controls the compute needed to build and run them, and who sets the rules about what AI can be used for. Open-source advocates have a compelling partial answer to the first. They have a weaker answer to the second. Nobody has a complete answer to the third.
The data on model transparency is striking. The proportion of model downloads accompanied by disclosure of training data fell from 79.3% in 2022 to just 39% in 2025, according to arXiv research on the economies of open intelligence. At the same time, open-weight models — DeepSeek-R1, Qwen3, Llama 4 — are matching or outperforming closed models on many benchmarks at a fraction of the cost. The performance advantage that justified closed-model pricing is shrinking faster than closed-model companies expected.
But the compute question is largely unaddressed by the open-source response. NVIDIA controls advanced GPU manufacture. TSMC controls the fabrication of those chips. US export controls targeting China have turned AI compute access into a geopolitical instrument. None of that is changed by releasing model weights under an open licence. You can download the weights; you cannot download the chip fab. The distribution of AI capability at the model layer and the concentration of AI capability at the compute layer are both real — and they are in tension in a way the debate rarely acknowledges.
The sourced claims on both sides. The concentration column covers market structure, compute, data transparency, and geopolitics. The distribution column covers open-weight performance, hybrid architectures, sovereign AI, and the first signs of global governance.
Data transparency has collapsed: the proportion of model downloads disclosing training data fell from 79.3% in 2022 to just 39% in 2025. Users and researchers deploying these models cannot verify what they were trained on — and therefore cannot fully audit the values, biases, or intellectual property they encode.
Compute concentration: advanced GPU manufacture is controlled by one company (NVIDIA) and the fabrication of those GPUs by one foundry (TSMC). This is a structural choke point that no open-source licence can circumvent. The ability to train frontier models at scale depends on access to hardware that a very small number of actors control.
US chip export controls targeting China have transformed AI compute into a geopolitical instrument — fragmenting the global AI ecosystem and placing smaller nations at a structural disadvantage. Countries caught between the US and Chinese blocs face restricted access to advanced hardware regardless of their domestic policy choices.
Closed-model companies control training data, model weights, and deployment infrastructure simultaneously. They determine what AI does, for whom, at what price, and with what content restrictions — without meaningful external oversight of those decisions. The concentration of all three levers in the same corporate entity is the defining feature of the closed-model structure.
DeepSeek-R1, Qwen3, and Llama 4 are open-weight models that match or outperform closed models on many standard benchmarks at a fraction of the cost. The performance advantage that sustained closed-model pricing premiums is eroding faster than the frontier labs anticipated — and the distribution of that capability is genuine, if uneven.
Most serious enterprise AI adopters already use hybrid architectures — combining open-weight and closed models by task, cost, and sensitivity. Pure closed-model dependency is neither inevitable nor universal. The practical AI stack in 2026 is more heterogeneous than the "open vs. closed" framing suggests.
2026 marks the first genuinely global phase of AI governance: the UN-backed Global Dialogue on AI Governance and the Independent International Scientific Panel on AI are both active with state participation across blocs. Governance infrastructure is being built outside the US-China bilateral framework.
Sovereign states and public institutions — particularly in Europe and the Global South — are actively funding open-weight model development to avoid strategic dependency on for-profit US or Chinese companies. The motivation is explicitly geopolitical: domestic AI capability as infrastructure, not as product.
Open weights addresses part of the concentration problem — it gives developers, researchers, and sovereign states something to build with without paying closed-model prices or accepting closed-model terms. That is real and significant. It does not address compute: you can download a Llama model but you still need NVIDIA GPUs to train it at scale, and TSMC to manufacture them.
The power debate in AI is three questions running simultaneously. Who controls the models? Open source has made genuine progress here — the performance gap has closed, the number of capable open-weight models has expanded, and hybrid architectures are mainstream. Who controls the compute? This question is substantially unanswered by the open-source response, and it is in many ways the more fundamental one: without access to advanced silicon, no one can train or fine-tune at frontier scale, regardless of what licence their starting weights carry. Who sets the rules? This is being contested in three capitals simultaneously — Washington, Beijing, and Brussels — with the outcome genuinely unknown.
The most important structural observation is this: the three questions have different answers, different actors, and different timelines. Conflating them — treating "open vs. closed" as the single axis of the power debate — obscures the compute layer, which is where concentration is deepest and where no distributional mechanism currently operates.
The Europe-US-China strategic divide is real and sharpening. US export controls on advanced chips, Chinese state investment in domestic semiconductor alternatives, and European sovereign AI programmes are all moving simultaneously. Whether this resolves into negotiated international standards or a permanent AI iron curtain — with different models, different chips, and different governance frameworks running on parallel tracks — is genuinely unknown in 2026.
Open-weight models have reached performance parity with closed models on most mainstream benchmarks as of mid-2026. The frontier remains contested — GPT-5, Claude 4, and Gemini Ultra are still materially ahead on the most demanding reasoning and multimodal tasks — but the gap is narrowing. For the vast majority of practical applications, open-weight models are already sufficient, and the governance implications of that are significant: the barrier to deploying capable AI without closed-model oversight or pricing is lower than it has ever been. That distributes capability. It also distributes the risks associated with capable AI into contexts where closed-model guardrails do not apply.
This is where concentration is deepest and moving least. NVIDIA's H100 and B200 series dominate frontier training. TSMC's 3nm and 2nm processes fabricate them. Neither is subject to meaningful competitive pressure at the frontier within the next two to three years. US export controls mean that the compute available to Chinese AI developers — and to smaller nations — is structurally constrained regardless of their domestic ambitions. The Atlantic Council's 2026 geopolitics analysis identifies this as the single most significant structural factor in the global AI balance — more durable than any model-layer advantage.
The UN Global Dialogue on AI Governance and the Independent International Scientific Panel on AI represent the first genuinely multilateral governance mechanisms with state-level participation. They do not yet have enforcement powers, and they are operating in parallel to rather than above national regulatory frameworks. The EU AI Act is the most advanced binding rule set. The US has no equivalent federal legislation. China operates a distinct domestic regulatory framework. Whether these three frameworks converge on shared international standards, or whether regulatory fragmentation becomes a permanent feature of the global AI landscape, is the central governance question of the next three years — and one that no actor currently controls.
Primary sources cited in this issue. Dates and URLs verified at time of publication.
Open weights is not the same as open power. The real concentration is in chips, data, and governance — none of which an open-source licence can fix. This issue separates the three questions and names what each one actually requires.
Every issue of The Debate presents the strongest sourced case on both sides of a live question in AI. The ledger is not a verdict. The bridge names what both sides leave unresolved. The line says where law and policy currently sit.
Craig Stanley Studio · craigstanley.work · The Debate series · Issue 04 of 09 · 2026