The Periphery Is a Choice
Most of the AI-anxiety industry is pointed inward. Will the model take my job, my team’s jobs, my industry? It’s a reasonable worry, but it quietly assumes you live in a country that builds frontier AI. Anton Leicht, in a sharp essay for Asterisk called “Beware the Permanent Periphery”, flips the frame: the more consequential divide isn’t between workers and machines inside rich countries. It’s between the handful of nations that build frontier AI and the roughly 193 that never will.
His argument, compressed: a displaced worker inside the US still lives inside the polity that captures AI’s gains. She votes, she protests, she gets redistributed to, however imperfectly. A displaced country has none of that leverage. It absorbs the disruption, automated-away export industries, AI-empowered adversaries, competitive pressure in every tradable service, without any claim on the upside. And unlike a domestic underclass, a peripheral nation has no mechanism to force the winners to share. That’s what makes the periphery permanent.
I spend my working days on the physical layer of exactly this divide, datacenter capacity in Europe, and I think Leicht’s essay is the most useful thing written on AI geopolitics this year. Not because the diagnosis is novel, but because of where he locates the danger. The periphery isn’t something done to countries. Mostly, it’s something countries do to themselves.
The compounding gap
The core mechanism is recursion. Capable models help build more capable models. Revenue buys compute; compute trains better systems; better systems generate more revenue. User data attracts users. Leicht points to the access asymmetries that opened up this year, restricted tiers of frontier models, government-exclusive arrangements, and notes that even a two-month head start compounds. In domains where advantage is relative rather than absolute, cyber, trading, contested markets, there is no “good enough.” The floor is set by the strength of the other side’s AI.
This is the part that fast-follower strategists keep underestimating. The gap between frontier and follower isn’t a fixed distance you can close with a national champion and a five-year plan. It’s a growth rate. You don’t catch up to a growth rate by running hard for a while; you catch up by changing the structure of your participation.
The protectionist trap
Here’s where the essay gets uncomfortable for a European reader. The instinctive policy response to dependence, sovereignty requirements, domestic procurement preferences, regulatory moats around local players, feels like self-defense. Leicht argues it’s the opposite: a self-reinforcing spiral into the periphery. Every wall that keeps frontier AI out also keeps your firms training on weaker tools, your talent working on yesterday’s stack, and your economy paying the disruption costs of global AI without collecting any of its dividends.
His historical analogy is blunt: societies that opted out of industrialization in the 18th century didn’t get to keep their pre-industrial equilibrium. They got to meet industrialization later, on much worse terms. Protectionism, he notes, either fails quickly or ends in a forced re-entry into global markets, after the compounding has run against you for years.
I’d add a field observation: I have sat in enough European “sovereign cloud” conversations to recognize the pattern. The sovereignty debate almost always centers on control of the stack and almost never on access to the frontier. We negotiate hard over where the data sits and barely at all over whether our companies get first-tier model capability. It’s optimizing the fence while the harvest moves elsewhere.
Megawatts are the entry ticket
The constructive part of Leicht’s essay is the one closest to my day job, and I think he’s more right than he knows. What can a middle power actually trade for frontier access? Not model weights, not talent at scale, not capital the labs don’t already have. What’s scarce is physical: sites, grid connections, permits, power. Compute is the binding constraint of the frontier, and compute is made of land and electricity, things mid-sized countries control absolutely.
That makes datacenter infrastructure something unusual: a foreign-policy instrument that a country of ten million can wield. Offer the sites, the megawatts, and the permitting speed, and you get frontier capacity on your soil, procurement gravity, an operational talent base, and a seat near the table. The window matters, though. Infrastructure is leverage precisely because compute is scarce. If the constraint eases, more efficient models, more supply, the ticket price of admission goes up and the seller’s market closes. The countries moving now (and a few are moving impressively fast) are trading a temporarily scarce asset for a permanent position. The ones running two-year feasibility studies are watching their one bargaining chip depreciate.
Bend, don’t brace
Leicht’s second prescription is about labor markets, and it rhymes with something I wrote about capacity two weeks ago. The economies that will handle AI disruption best are not the ones that armor specific jobs against change, that’s the employment version of hardcoding a single VM shape and calling it architecture. It’s Denmark’s flexicurity model he points to: let the jobs move, insure the people generously, retrain fast. Rigid protection locks an economy into its current shape at exactly the moment the shape is becoming obsolete. Design for obtainability, but for careers: define what you’re actually trying to preserve (income, dignity, adaptability) and allow everything else to flex.
The third play is bottleneck positioning, finding the node in the AI-era supply chain where you’re hard to route around. Taiwan and South Korea did it in semiconductors, India is doing it in biopharma manufacturing, and there’s a long tail of Baumol-resistant services where humans stay expensive because AI makes everything else cheap. Not every country gets a frontier lab. Every country gets to choose what it’s indispensable for.
The reality check
You can push back on parts of this. Maybe open-weight models keep the follower gap tolerable. Maybe the recursion stalls. Maybe US restrictions get so aggressive that alignment with the frontier stops being available at any price, and hedging looks wiser than Leicht allows. These are real uncertainties, and he’s honest about them.
But notice the asymmetry in the bet. If you pursue engagement, infrastructure deals, flexible labor markets, bottleneck positions, and the frontier gap turns out to be manageable, you’ve lost little: you own useful infrastructure and an adaptable economy. If you pursue walls and the gap compounds, you’ve converted a temporary disadvantage into a structural one, with your leverage spent and your talent gone. One error is recoverable. The other is the permanent periphery.
The countries that thrive in the next decade won’t be the ones that built the tallest fences around their digital economies. They will be the ones that traded what was scarce while it was scarce, kept their economies loose enough to reshape, and made themselves impossible to route around.
Opinions are my own.