About a year ago, I started worrying that intelligence would end up being very similar to capital in that it would be concentrated, difficult to redistribute, and quietly decisive. At the time, this seemed rather far-fetched. The largest models were impressive but inconsistent. Open-weight releases kept coming. The field was so active that the idea of a few firms gradually amassing the strongest systems seemed unlikely.
If you have been paying attention to recent changes, you will have noticed that this feels less like a prediction and more like a description of current developments. The best models are not closing in the ‘cartoon’ sense, and open-weight work remains important. Nevertheless, the top tier is more restricted than it was a year ago. Anthropic is only releasing some frontier capabilities to a select group of security partners. Alibaba, often cited as evidence that strong open weights still matter, only makes some of its strongest systems available via hosted APIs. The concentration of computing power, talent and cloud relationships within a few organisations is causing bottlenecks in access, deployment and governance. The infrastructure surrounding these models — the aspect that determines who can experiment with them, audit them, and deploy them — is beginning to resemble private infrastructure with rented access rather than a commons.
We are witnessing the transformation of intelligence into a gated asset class. It’s not fully closed, but it’s becoming increasingly gated.
From ‘AI for everyone’ to a tiered stack.
If you had simply believed the marketing of the last few years, you might have thought that universal access was on the horizon. Llama was open source. DeepSeek-R1 briefly challenged assumptions about who could develop a capable reasoning model. It seemed likely that frontier capability would spread almost as quickly as it advanced.
The landscape is changing, nevertheless Open-weight frontier work is still ongoing. AI2’s Olmo 3, for example, is an excellent illustration of an open-source contribution: it comes with the entire pipeline, including weights, checkpoints, and data. Openness across the field is decreasing, and the strongest systems are being pulled into exclusive tiers.
Anthropic’s Mythos Preview is the prime example of this trend. Anthropic describes Mythos Preview as a high-capability frontier system and is limiting access to a select group of defensive cybersecurity partners via Project Glasswing. The stated reasoning is risk-based: the model is too dual-use to be released widely. You may or may not agree with this decision. What is harder to ignore is the institutional fact that a private firm has decided which institutions get to use one of the world’s most capable systems, and which do not.
Imagine a layered intelligence stack. There are capable open-weight models available at the bottom, but not everyone can run them. Above that are premium hosted commercial systems. Above that is an even more exclusive tier, mostly reserved for firms, security partners and, eventually, states. Studies by the OECD and Carnegie say the same thing: ‘Open vs. closed’ is now too simplistic.
The political economy of intelligence.
If you look at the bigger picture, it’s less about the products themselves and more about the underlying concentration of resources — compute and human capital.
Sevilla et al. measured training compute over decades and found that, since the start of the deep learning era, it has roughly doubled every six months, with another significant increase for the largest systems around 2015 (arXiv). You can’t ride that curve without significant capital, GPUs, and power. Luccioni, Jernite and Strubell (arXiv) also demonstrated that the inference cost of general-purpose generative systems is significantly higher than that of task-specific models. Training was the obvious industrial stage. Deployment is the next stage, and it involves inference, serving models, correlated security and integration, which deepen the concentration beyond the training alone. While none of this is inherently bad, the cost profile favours hyperscalers, frontier labs and partnered governments over universities, civil society and smaller firms.
The Public AI white paper makes this point clear: the most advanced systems are developed and controlled by a small number of private companies, which hold power over both the models and, most of the time, the underlying computing infrastructure. The International AI Safety Report, written by a large coalition of researchers and governments, identifies the concentration of capability and computing power as structural risks in their own right, in addition to the existing debates about misuse and accidents in the field.
While this comparison to capital is helpful, I also want to stop relying on it. The useful aspect is structural at it’s core:: owning productive infrastructure determines who sets the terms and decides what counts as valuable work. Intelligence is starting to look similar. If a frontier model increases a researcher’s productivity tenfold or renders an entire department redundant, then deciding who has access to it will be crucial for future development and progress. The right comparison is not with nuclear weapons or a priesthood of knowledge. Rather, it is the mundane tale of how control over a foundational asset redistributes power.
Access is agency.
At some point, this stops being a purely technical argument and becomes a question of civic responsibility. The question is not whether the public should own every top model personally. Most people do not need a frontier system in the same way that they need running water. However, the question remains: what would happen to a society in which the systems influencing research, journalism, and state decision-making were largely controlled by a select group of private entities? We can use the tools, but we cannot govern them. This creates an even greater asymmetry and dependency. Leslie and her colleagues frame the issue as one of public interest (Harvard Data Science Review). They argue that the wave of frontier AI policy initiatives has produced grand rhetoric and weak regulation in the form of self-regulatory commitments, non-binding principles, and small national safety institutes.
Institutions that produce valid knowledge also shape the conditions under which others can know and act. This corresponds with Foucault’s framework regarding power and knowledge. These institutions shape the conditions under which others can know and act. If frontier AI becomes dominant in research, decision-making and public administration, the institutions controlling it will define what constitutes valid questions, evidence and answers.
Langdon Winner’s essay ‘Do Artifacts Have Politics?’ makes this argument more concretely. While some technical systems are politically neutral, many are not. Some require hierarchy and centralisation just to function. Winner’s example was nuclear power. As I previously argued, this comparison is probably not the right one, but Frontier AI is starting to fit a similar pattern involving hyperscale computing, security screening, partnered deployment and proprietary infrastructure. Whether any lab intends it or not, the architecture is political.
If intelligence becomes a form of infrastructure, then control over that intelligence becomes a form of governance. That is the line I keep coming back to:
Open weights, open power, and the steelman.
The obvious response is to open-source everything — problem solved. It sounds simple and good, but I don’t think it works.
Open weights matter because they reduce dependence on any single firm and enable external safety research. In Science, Bommasani et al. carefully make this case: open foundation models can spread power and raise competition, and policy that ignores them will badly distort the ecosystem. I think they are quite right. However, running a serious model still requires industrial computing power and supporting infrastructure. If only a few institutions can afford this, then ‘the weights are public’ could still describe a concentrated system.
I also want to consider the case for caution seriously. The strongest argument for restricting frontier AI is not corporate land grabbing, but an important safety argument. Anderljung and his colleagues set out the reasons clearly: dangerous capabilities can emerge unexpectedly, deployed models are difficult to control, and once those capabilities have spread, it is not feasible to implement a clean rollback. They therefore advocate implementing standards and registration procedures prior to the widespread release of frontier systems. Anthropic’s Mythos rationale sits within that tradition. There are real misuse vectors here. ‘Release everything immediately’ is not a serious answer, and it is not acceptable to give such a response.
However, this does not address the wider question of governance. Even if you opt for a slow release and limited initial access, you still need to decide who will make those decisions and to what oversight they will be subject. You also have to decide whether the private allocation of general-purpose intelligence should be the norm. What is missing is an institutional design that is publicly accountable, so that society as a whole benefits from it.
What would be a better alternative?
I don’t have a full set of policies to offer, and I’m wary of essays that claim to have one. However, the alternatives are not imaginary.
The Public AI white paper provides the clearest proposal. It suggests treating part of frontier AI as public infrastructure, providing researchers and smaller institutions with accessible computing resources, releasing models in the public Interest, and setting up talent pipelines that don’t all lead to the same five firms. Add governance that isn’t dependent on corporate agendas. None of this is a magic solution and public institutions can be slow or captured. Nevertheless, a credible public option changes everyone else’s position in negotiations. It also establishes a real framework for transparency obligations around training data and post-deployment impact.
Arvind Narayanan and Sayash Kapoor argue that AI should be understood as a normal technology; powerful and transformative over time, but not exempt from the social and institutional logic that has defined every previous technological era. I find this perspective insightful. It stops frontier AI from being treated as so extraordinary that the usual questions of governance, accountability and ownership no longer apply.
Personal thoughts.
I am writing this from outside the rooms where these decisions are made. I do not own a cluster. What motivates me is the concern that intelligence could become just as concentrated as capital: incrementally and plausibly, one reasonable decision at a time. Yet this is not the whole story; Transparency and public-interest framing continue to be promoted, sometimes from within the labs themselves. Societies still have time to decide what kind of governance they want over the infrastructure on which they are about to become increasingly dependent.
The problem is not the existence of gates. The fundamental issue here is how gates are being constructed: before the public has had an impact on ownership, management and the interests they serve.