The following analysis summarizes my current thinking about the risks
& opportunities around increasing LLM/AI usage over time, specifically from the perspective of government policymakers.
The initiative to do this work came after I read the Cut Off article by Anton
Leicht. I am still unsure about some of that article's conclusions
because they depend on an aggressive reading of the recent Mythos and
Daybreak releases by Anthropic and OpenAI, but several of his
arguments intrigued me.
Assuming his premise holds, how do the points translate for
policymakers outside the US? Are there also adjacent risks and
opportunities that Anton Leicht may have missed?
I draw these findings from a variety of sources besides Anton Leicht
himself, both from published articles and discourse between industry
professionals and between policy-adjacent experts in my network.
Note that we are assuming here that increased LLM/AI usage is generally
desirable and leads to increased economic output at the level of entire
countries/regions. Validating this assumption is left out of scope for
this analysis.
Your main take-aways:
- The era of broad, cheap, unrestricted frontier-AI access is
closing.
- Compute, not models, is the binding constraint: AI sovereignty
is becoming like energy sovereignty.
- Defenses against IP theft and defenses against equitable access
to AI are the same defenses.
- Non-US countries have a closing window to trade datacenter buildout
for access guarantees.
- Open-weight models are a credible fallback for most uses.
- The harness/tooling layer is a more achievable moat than
training a frontier model.
Terms
- Frontier model
- The most capable LLM (generative AI system) available at any given moment. Currently produced by a handful of US labs (Anthropic, OpenAI, Google) and one or two Chinese labs (DeepSeek, Alibaba/Qwen).
- Open-weight model
- An LLM whose internal parameters are published, so anyone can run it themselves rather than going through a company's API.
- Distillation
- Training a smaller, cheaper model by having it study the outputs of a larger one. This is a key mechanism by which trailing models catch up to leading ones.
- Compute
- The GPUs and datacenters needed to train and run these systems; the binding physical resource.
- Harness
- The software middleware that adapts a general-purpose LLM
to a specific application.
Risks
Frontier access becomes a foreign policy lever, not a market good.
Observing that recent model releases were restricted (Anthropic's
Mythos cybersecurity model went only to a handful of US companies;
OpenAI's Daybreak cybersecurity model followed suit), we might expect further
government-mandated access control.
Three forces drive this: genuine misuse concerns (you don't want a
model that finds zero-day vulnerabilities available to everyone),
commercial protection against copying by competitors, and eventual
US government use of access as a bargaining chip in unrelated
negotiations. The Trump administration's pattern of bundling trade,
intelligence, and tech access into single deals makes the third risk
concrete rather than hypothetical.
For policymakers in non-US countries, the operative implication is:
assuming continuous access to the best AI is the same kind of
mistake as assuming continuous access to any other strategically
controlled good.
Conversely, if Mythos and Daybreak turn out to be one-offs rather
than a trend, the risk weakens considerably.
References:
Some experts suggest that we can fall back to weaker models if
frontier models become controlled. This hedge may be weaker than it
looks.
The natural fallback, that is using open-weight models from Meta,
DeepSeek, or Qwen which lag the frontier by three to six months, is a
common suggestion. But consider: cost curves bring yesterday's
frontier capabilities down to reasonable prices, but tomorrow's
frontier keeps getting more expensive.
If economic and security competition rewards having the best AI
rather than adequate AI, being six months behind is structurally
losing, not merely “almost as good.”
Whether this matters depends entirely on what you're using AI
for. For drafting emails or generating reports, six months behind is
fine. For high-stakes uses like cyber defense, biosecurity or
intelligence analysis, it may not be.
References:
Compute, not models, is the binding constraint.
This is a point professionals often underweight. Even Anthropic,
sitting on top of the market, is reportedly buying compute time from
third parties because it struggles to serve its own
customers. Running a frontier model for a new country or customer is
genuinely expensive at the margin (unlike e.g. Microsoft Office,
where the marginal cost is near zero).
For a policymaker, this reframes the question: “AI sovereignty”
cannot mean only training your own models. It has to include
datacenters, GPU supply, and the energy to power both.
In other words, AI sovereignty is becoming more like energy
sovereignty than software sovereignty.
References:
Distillation defenses will tighten the screws on third parties.
Distillation is the mechanism that keeps open-weight models,
currently mostly Chinese, close to the frontier, currently mostly
US-sourced. If US firms or the US government crack down on
distillation (through identity verification, geographic
restrictions, query-pattern monitoring, etc.), this will widen the
gap between open and closed models everywhere.
In other words, defenses against intellectual property theft and
defenses against equitable global access are the same defenses. You
cannot have one without the other.
References:
The wealth-concentration dynamic.
If frontier AI is a genuine productivity multiplier, and access to
it is rationed by price and credentialing, then it functions as a
wealth amplifier for those who already have capital.
The risk: those with access pull further ahead, generate more
capital, secure even better access.
For policymakers, this raises the question of whether AI access
becomes something governments need to underwrite for citizens, the
way they underwrite roads or basic internet access, and if so, how.
References:
Geopolitical second-order effects.
Historically, when the benefits of major industrial revolutions were
distributed unevenly across nations, the result was mass migration
and destabilized democracies.
A world where some countries have the equivalent of post-scarcity
intellectual labor and others do not is not a stable equilibrium,
especially if the gap is visible and growing.
References:
Opportunities
The compute-for-access bargain.
Second-tier compute consumers (e.g. the Netherlands, UK, Germany, Japan, Singapore,
Australia, the Gulf states) can offer US hyperscalers favorable
terms (cheap energy, fast permitting, regulatory clarity) for
building datacenters on their soil in exchange for contractual
guarantees of continued frontier access.
The strategic logic: once Amazon, Microsoft, or Google has billions
of dollars sunk into a foreign datacenter, they become a US domestic
lobby against future administrations trying to cut access. The host
country is essentially buying lobbying capacity in Washington with
land and electricity. Singapore is already moving in this
direction. Caveat: this also makes the host country more
contractually dependent on US firms, which is a different
vulnerability (analogous to Europe's previous dependency on Russian
gas).
This may be worth the trade for most middle powers, but the
trade-off should be discussed democratically.
References:
Hardening defenses to reduce the security justification for restriction.
If countries invest seriously in biosecurity (screening synthetic
DNA orders, hardening laboratory supply chains), cybersecurity
(patching critical infrastructure faster), and datacenter physical
security, then the policy case for restricting model access on
safety grounds weakens.
This is one of the few areas where AI accelerationists and
safety-focused groups can agree on the same investment.
References:
Open-weight models as a strategic floor.
Even if frontier API access is cut off, capable open-weight models
from Meta (Llama), Alibaba (Qwen), and DeepSeek lag by months, not
years.
For most economic uses (in particular software development, customer
service, document processing, content generation), six months behind
the frontier might be functionally indistinguishable from the
frontier. If this view is generally right, the policy implication is that
frontier access is a luxury and open weights are the actual
strategic resource. The corresponding risk is that the view is
incorrect and frontier-vs-trailing is winner-takes-all. Reasonable
people disagree, and the answer probably varies by use case.
References:
Buildout speed as a policy variable.
The single most consequential thing a country can do is accelerate
datacenter and energy buildout. Permitting reform, grid expansion,
and skilled construction labor are the unglamorous
bottleneck. Safety-motivated objections to fast buildout look weaker
once you account for the access risks created by slow buildout.
References:
Harness and tooling as a domestic moat.
Even if LLMs become commodity resources, the surrounding software
(e.g. harnesses, agent frameworks, evaluation infrastructure,
integration tools, domain-specific scaffolding) is itself valuable
and harder to commoditize than the models.
A country or company that invests in this layer captures value
regardless of which model wins. This is a more achievable form of
“AI sovereignty” than training a frontier model from scratch.
References:
Universities and non-commercial providers.
A wildcard that discussions outside of the EU often miss: academic
institutions or international consortia could host/run open
frontier-adjacent models for the benefits of citizens and
organizations within their region.
References:
The main unresolved tension
There is strong disagreement on the question of whether an access
restriction on frontier models would “lock out” countries and
organization who do not have access from all the possible wealth
generation enabled by AI.
“AI frontier maximalists,” on one side, believe that LLM capabilities will
continue to increase on an exponential curve, and an access gate would
create a capability gap (between frontier and non-frontier users) that
would be forever impossible to close. See the risks above, as well
as the following references:
.
Industry practitioners, which we could name “good-enough realists,” in
contrast, frequently report that non-frontier models are satisfactory
for most use cases and will likely be available at low cost over time
due to commodization. See the opportunities above, as well as these
references: .
We currently have signals in the industry that support both
views. It will take more time (possibly years) to observe where the chips
will fall.
Shared strategic directions
From a policymaker perspective, there are shared strategic directions
worth considering in the short term, regardless of which view ends up
being correct:
- Invest in compute and energy infrastructure regardless.
- Treat frontier API access as a contingent resource rather than a
guaranteed input.
- Negotiate access guarantees through datacenter buildout deals while
the US needs international compute capacity.
- Invest in the harness/tooling layer where domestic firms can build durable competence.
- Harden against misuse domestically so the security argument for restriction loses force.
- Maintain a credible fallback to open-weight models for use cases
where six-months-behind-the-frontier is acceptable.