Inequalities Arising from Frontier AI Access
A handful of companies’ decisions on pricing, access, and deployment now shape a country’s productivity frontier as directly as any trade agreement. And Singapore’s new Economic Strategy Review does not mention this.
The Access Problem Is Already Here
It is becoming clear that access to frontier technology is a source of geopolitical leverage, with geoblocking of frontier AI (e.g. Claude in China), limited previews of advanced AI capabilities (e.g. Claude Mythos), and restricted product surfaces for global consumption (e.g. rollout of Google’s NotebookLM). Even the engineers within these frontier AI labs are worried about falling behind because they cannot access one another’s tools.
More than a productivity frontier, it also determines questions of national security. Consider Anthropic’s Claude Mythos model, which is expert in finding zero-day (i.e. previously undiscovered) security vulnerabilities autonomously. It was only distributed in limited preview to US firms (under Project Glasswing). OpenAI did the same with its Daybreak initiative. These are not just productivity questions — they include national security.
While China remains a major source of open-weight models, the lab that is perceived to be leading is the closed-source lab from Bytedance — meaning even the open-weight alternative may not remain the path to the frontier.
Last week in Parliament, I spoke about this. This was rooted in my experience with Claude Code, which has made my coding and other tasks much more efficient — but which also made tangible to me how much a country’s competitive position now depends on which tools its people can access.
The Subsidy Moment
Beyond geopolitics, there is also the possibility that we are living in an unsustainably private-money and VC-subsidised moment. I recall the case of Uber’s strategy: an era of cheap rides subsidised with VC money, capturing the market, with prices rising once the market had consolidated.
For instance, I find that it is worth it to pay for Claude Code through a Max plan which currently costs about $200 USD a month, as it makes my coding and other tasks more efficient. It however seems heavily subsidised, and the equivalent in API costs has to be significantly more. If frontier AI pricing corrects to true cost within two years, and prices of frontier AI remain substantially high, Singapore’s entire diffusion model — which assumes cheap, abundant access trickling down from Champions to the broader economy — prices out everyone below the enterprise tier.
Not everyone will be able to access premium AI models. The gap between those with access and those without will be enormous. And if open-source models go closed-weight, this will reduce the alternative pathway.
What the ESR Misses
The Economic Strategy Review report was released yesterday. There is a lot of discussion on AI, but it is silent on the most consequential risk facing Singapore’s AI ambitions: whether we will continue to have access to frontier AI at all.
No mention of supply-side risk to frontier AI access — the entire strategy treats frontier model availability as a given input, never as a variable that could be constrained. The ESR treats Singapore as a consumer and deployer of frontier capabilities produced elsewhere, not as an orchestrator of AI capability flows. Its Thrust 8 (resilience) identifies energy, supply chains, food, and climate as domains requiring strategic buffers and diversification — but does not extend this logic to AI compute/model access, which, if we are to take the AI Hub framing seriously, is this country’s critical emerging dependency.
The ESR is written entirely within the assumption that the current era of abundant access continues — that frontier AI is a commodity input Singapore can procure, and the differentiator is what you do with it. If that assumption breaks, the entire three-tier deployment model (Champions → sector-level solutions → economy-wide diffusion) is compromised at the foundation. The “best problems to solve” framing only works if you also have access to the best tools to solve them with.
It is a major omission. And as the examples above show, the problem is not theoretical.
What Can Be Done
Anton Leicht of the Carnegie Endowment has a useful treatment of this problem. He argues that frontier AI access will be constrained by three compounding forces: security concerns that motivate withholding (as we saw with Mythos and Daybreak), compute scarcity that makes serving frontier models genuinely zero-sum, and the eventual instrumentalisation of access controls by the U.S. government for broader strategic purposes. Critically, he points out that efficiency gains do not resolve this — they cheapen last-generation capabilities, not frontier ones.
Leicht proposes four responses, which I find broadly right. First, reduce the justification for restriction — harden cyber and biosecurity infrastructure so the case for withholding frontier models from countries like Singapore is weaker. Second, build datacenters at scale to alleviate the compute crunch. Third, build compute in exchange for access — offer hyperscalers favourable terms for regional buildout in return for contractual frontier access guarantees. Fourth, retain some independent capability to build as a contingency, because if the first three fail, a country without any domestic frontier capacity has no fallback. On fallback capacity, AI Singapore and the NSCC’s efforts need to provide a strategic buffer, and should be resourced accordingly.
On the third point — access for compute — this is where I think Singapore has some cards, and where Leicht and I converge. In Parliament last week, I made the case that Singapore Inc builds good data centres, and that we are among the world’s leaders in water reuse and integrated water management — which is a binding constraint on data centre expansion across water-scarce Southeast Asia. If we position ourselves as the infrastructure partner of choice for this region, that is real leverage — something we bring to the table in exchange for access, for pricing, and for presence. When a company like Anthropic or OpenAI approaches us, we should be their preferred regional partner in rolling out and scaling their data centre buildout, as well as all the infrastructure needed to make these data centres work.
It will likely be an ongoing goal of middle powers to proliferate, within a broader network of states, the production of frontier technology and capabilities such as frontier AI models. Whether it can be done is unknown, but there will be serious attempts to try. In any case, there will need to be continued foreign policy engagement to maintain ecosystem access — with both the US and China.
Many AI researchers and AI users I have spoken to, both Singaporean and foreign, have inequality of AI access as their number one worry. If we don’t have a frontier AI access resilience plan — sitting alongside the existing frameworks for energy, food, and supply chains - we are building an AI hub on an assumption we do not control.