Cut Off: access to frontier AI will be scarce and selective

Cut Off

From the Threading the Needle newsletter. Challenges the popular “AI tokens will be abundant, the future belongs to those who use them best” mantra — loudest among the world’s “middle powers” betting on navigating the AI revolution with “good-enough” models. Leicht argues access to frontier AI will soon be limited by economic and security constraints.

The “Mythos moment”

• In early April, Anthropic announced Mythos, a leading cybersecurity model, and made its vulnerability-patching ability available only to a [select few] mostly US-based companies (its “Glasswing” partners) — leaving startups, systems integrators, and allied capitals off the list.
• OpenAI followed with a similarly limited release (its “Daybreak” initiative for a gpt-5.5-cyber model), dispelling hopes this was a fluke.
• The US government is reportedly planning to “do something at some point.” Leicht argues this reveals structural trends, not just current events.

Three compounding constraints

1. Security & distillation

• Misuse risks: a highly capable model (useful for cyberattacks or bioweapon design) gets rolled out first to defenders to shore up vulnerabilities, then to trusted customers, and only later — once no longer state-of-the-art — to everyone.
• US government interest: restricted access serves national security. The NSA might want to know which zero-days a model like Mythos can find so it can exploit them first (cf. EternalBlue), rather than ensure equitable global diffusion.
• Theft, espionage, distillation: the biggest concern. “Fast followers” (e.g. DeepSeek, 6–9 months behind) reportedly rely on distillation requiring broad API access. Distillation is economically untenable for frontier labs (can’t recoup R&D if distilled in six months), so expect crackdowns — stricter KYC, more restrictive default access, geopolitically conditioned access.

2. Compute crunch

• Unlike software (low marginal cost, so market-size logic favours mass rollout), serving frontier AI is near zero-sum — high marginal compute cost per token.
• Labs repeatedly hit compute crunches, cut offerings, and struggle to subsidise consumer subscriptions; Anthropic is reportedly shopping ad-hoc datacenter deals (even with rival xAI). This worsens if AI begins to rival human workers’ output.
• Efficiency curves won’t save you: they make last year’s capabilities cheap, but frontier (“Mythos 2”) capabilities keep getting more expensive month-to-month. If you need “the best” AI (not just good-enough) to compete with rivals and attackers, efficiency gains don’t bail you out.
• So the marginal cost of a new user (firm or country) is high, sharply diminishing the market-power leverage that buyers usually wield. Competition over who gets tokens will emerge — Leicht foresees a revival of GAIN Act–style logic (Americans as “buyers of first resort” of American-produced intelligence), or razor-thin margins favouring whoever monetises tokens best (not cash-strapped governments or constrained European businesses).

3. US government leverage

• What begins as genuine security restriction may not stay that way. With a formal role over frontier-token flow, the US could wield access control for political/strategic ends.
• The Trump administration’s style is to bundle leverage (breaking trade deadlocks with intelligence threats, stalling tech deals over food-safety standards). So frontier access stays “fundamentally contingent” wherever governments’ strategic interests diverge — even if security and economics are handled “right.”

The next equilibrium

• Unlimited API access becomes the exception. A new model flows first to the US national security apparatus, then to trusted defenders (US firms, maybe a few internationals), then to firms clearing high KYC/security bars. Everyone else — consumers, scrappy startups, nervous governments — may only ever get limited product layers (chatbots, coding agents), never clean API access, until the next generation enters the pipeline.
• This is a bad future. Economically, the accelerationist critique is right: restricting frontier access throttles innovation and the “Hayekian” process of figuring out how to live and work with AI. But the fault isn’t Anthropic’s alleged push for nationalisation — it’s the underlying market and security dynamics.
• Geopolitically: countries split into frontier “haves” and “have-nots,” with the haves much wealthier and safer. Historically, unevenly distributed industrial revolutions triggered mass migration, reopened conflicts, and destabilised democracies.

Proposed solutions (framed as accel/safety convergence)

1. Make the world safer so security-motivated restraint is less necessary — harden against bio pathways, screen protein synthesis, deploy technical fixes for distillation before heavy-handed rules arrive, and improve importers’ own datacenter/cyber security.
2. Build lots of datacenters to ease the compute crunch — “not complicated, just fairly hard”; every GPU online now makes equitable diffusion later more likely.
3. Non-US countries build compute in exchange for access — allies offer hyperscalers favourable datacenter terms (subsidised energy) for contractual frontier-access guarantees, with mutual hostage-like incentives to prevent reneging.
4. Middle powers retain contingency options (leverage, and some domestic build capacity) for edge cases where all else fails — the subject of his next post.

Bottom line: the fix isn’t clever new ideas but executing the long-standing agenda — build infrastructure to host advanced AI at scale, and build a world that can handle it — before the “Andy Warhol era of AI access” (where rich and poor use the same product) ends.