for agents/llms.txtv0.7.81 · 12 Oct 2026

Home / Articles / Judo for small countries: where Portugal, Romania and Estonia can add the most value in AI, and why the race for GPUs is not it

Judo for small countries: where Portugal, Romania and Estonia can add the most value in AI, and why the race for GPUs is not it

By · 2026-10-12 · article v1.0.0 · site v0.7.78 · europeportugalromaniaestoniasovereigntyopen-sourcewardley-mapssmall-modelsinfrastructuredistributed-workagentsstrategyarticle

Abstract: I spent today with people from Romania and I spend a lot of my time with people from Portugal, and this is the article I wanted to hand both. The race for the biggest model, the most GPUs and the largest data centre is a race a small country cannot win and should not enter; on a Wardley map that layer is already a commodity and only capital plays there. The opportunity is the opposite end of the map: the plumbing that a talented person in Braga or Cluj needs in order to sell a service to the world, forming a company, hiring across a border, being paid across Europe, working in every language, a contract and a tax form, all of it still custom built in most of Europe when it should be a road the state built once. Estonia built that road and the numbers show what happens. Then the judo: open source and open schemas because there is no sovereignty without them, what the state pays for belongs to the commons, buy the alternative before the day you need it, small models for every language and culture rather than one frontier model, and agents with the least agency rather than the most. A playbook in five moves, a map, what I learned spending millions on a distributed workforce, and why constraints are the input and not the excuse.

A talented person in Portugal, Romania or Estonia selling a service to the world, as a Wardley map. The capital race on the right is a commodity; the red dots are the plumbing that is still custom built in most of Europe and that a state can move right; the green dots are where the market is still genesis and where the money should go.

In short

The event, and the word for what I mean

I was at an event today and met several people from Romania, including someone from an embassy who said, correctly, that the job of an embassy is to facilitate. I also spend much of my time talking to people in Portugal. This article is written to be sent to both, so it is written for a country rather than for a company, though every point applies to a company that wants to help one of these countries do it.

The word that kept coming up was that one: facilitate. It is not an accident that when we defined the roles for the team behind SGraph Send we named the growth role the Ambassador and wrote it to work the way an embassy works: represent the thing in territories it has not reached, promote trade, build relationships, project soft power, and never do hard sales, because the product should sell itself once the conditions exist. The role definition says the Ambassador's job is "to create the conditions for that organic growth". A country's ambassador does the same for a country. The question for this article is what the conditions are, for a small country, in the age of agents.

I have argued before that this is Europe's moment: in Europe's strategic opportunity in GenAI that the trends run towards smaller, explainable, multilingual, open and edge-deployed AI, which is where Europe's values already are, and in an open-source sovereign cloud for an open Europe that "digital sovereignty is not just about infrastructure, it is also about language and culture". For Portugal specifically I wrote in 2023 that generative AI was a golden opportunity. Europe's advantage is that it is not thinking top-down and is not playing a winner-takes-all game, and that advantage is strongest in its smaller countries, which can change faster. This article is the small-country version of the argument, with a map.

The race you cannot win, and should not enter

The conversation about national AI strategy keeps circling the same three things: the biggest model, the most GPUs, the largest data centre. I think that is misguided, and the map in the figure says why in one glance. Those components sit at the far right, in commodity, with a lead held by companies that have invested amounts a small country's whole technology budget would not match. It is a little like the climate conversation: there is nothing you can do about it in the short term, and trying to change that flow is not cost effective. On a Wardley map it is the clearest case there is of a market that does not exist for a small player, because the layer is commoditised and the buyer can already get it, cheaper, from somebody bigger.

That is not an argument for doing nothing. Unless you are already an established open-model lab with the investment behind it, the way Mistral is in Europe, you will not compete at the frontier, and the strategic response is not to try harder at the wrong altitude. It is to look at the rest of the map. The whole of Why I love Wardley maps, published this morning, is about that reading: a commoditised layer is a hard sell, and the market is where everyone is still custom building on top of commodities. Places of highest leverage is the same rule for a team: stand where you can add the most value right now, are not blocked, and can start.

The opportunity is the plumbing

Put a person at the top of the map. Not a platform, not a strategy: a talented developer in Cluj, a designer in Porto, a librarian, a historian, a project manager, a product team in Tallinn, selling a service to a buyer abroad who would rather buy than build. In One article, five readers and Firing journalists the roles that turn a notebook into a graph are librarians, cartographers, historians, explainers and storytellers; the whole food chain of professions can now provide its service to the world, and it can market, sell and distribute that service too. What the map shows is what that person stands on, and the red dots are the problem.

Forming a company. Hiring or contracting someone across a border. Getting paid, reliably, across all of Europe. Working in a language the buyer reads. A contract, an invoice, a tax return that is valid on both sides. None of that is the service. All of it is logistics, legal and operational plumbing, and in most of Europe each piece is custom built every time by the person or the small company that needs it, which means it is built badly, late and at a cost that falls on the wrong side. The overhead of hiring talent from a small country sits today on the buyer, and that is precisely what makes the buyer not do it, or do it through a platform that takes a cut and locks both sides in.

The numbers say how much commerce this stops. In the last Eurostat figures with a breakdown, for 2021, 18.5% of EU enterprises sold online to customers in their own country, 8.1% to other EU countries and 4.6% to the rest of the world. Fewer than one in ten firms in the single market sells across its own borders online. Look at how many companies trade natively in several languages across Europe, and the answer is very few, because we hand the translation problem to each company rather than treating it as infrastructure. Commerce does not happen, most of the time, because of attrition, and attrition is what a map is for.

Infrastructure is the state's job

We understand that a government builds roads, hospitals and ports, because they enable everybody. The shop that opens on the new road did not pay for the road; it paid taxes, and the taxes were pooled and the road was built once, for everyone. Nobody suggests each shop build its own road. Yet that is what we ask of every person who wants to sell a service out of Portugal or Romania today: build your own company-formation process, your own cross-border contracting, your own payment path, your own translation, your own compliance, and do it before you have a customer.

The thesis of this article is that in the age of agents, the plumbing for providing services to the outside world is infrastructure in exactly the road sense, and a state that builds it enables every one of its people at once. That is what the red arrows on the map are: company formation, cross-border hiring, payments, language and contracts, each pushed from custom built to commodity by pooled money, built once, open to all.

Estonia is the case study because it did it. A person anywhere can become an e-resident and form an Estonian company online; the programme's own dashboard lists more than 142,000 e-residents and more than 43,000 companies, and in 2026 alone e-residents had founded more than 4,200 companies by mid-September, about a third more than in the same period of 2025. Underneath the services sits X-Road, the data exchange layer that connects around a thousand organisations in Estonia and that is open source under the MIT licence, developed jointly with Finland and Iceland, so another country can take the road rather than draw it again. I do not see a reason the rest of the EU does not do the same, and a small country can do it faster than a large one.

Why it has to be open source

The plumbing has to be open source and run on open schemas, and this is the part that turns infrastructure into sovereignty. A road with a toll gate owned by one company is not infrastructure; it is lock-in with tarmac. If the company-formation service, the payment rail, the translation layer or the contracting flow belongs to one vendor, the state has rebuilt the problem it set out to solve, with a monopoly in the middle.

My logic is simple. You cannot have sovereignty without open source and open schemas, because sovereignty means you can leave. The argument is not that a country should not use services from American or Chinese companies; the United States depends on European companies, and everyone depends on Chinese ones, and that is what trade is. The argument is that you should not be locked in to them. Independence is the ability to run a parallel service, move the data and keep going, and that ability only exists if the code is open and the schemas are open.

It follows that what the state pays for should belong to the commons. I pay my taxes; the government uses them to buy software or to pay someone to write it; why does that software not also belong to me, or at least sit where I can read it and run it? It does not mean everyone gets to change the production system. It means the thing the public paid for is available to the public, which is also how the local industry that supports the local companies comes into being: an ecosystem of smaller firms can only grow around infrastructure it is allowed to see. The EU's open source strategy and the Commission's own open-source AI landscape report are moving this way; a small country can simply decide it.

This is also why I would point these countries at sgit and the vaults on this site, not as a sales pitch but as a shape. No server by design describes a newsroom whose only server-side components are commodities that hold no logic and cannot read the reader's data; Supply chain of vaults argues that the whole supply chain needs far more openness and that open source and Creative Commons are the proven way to get it. Data that stays with its owner and moves with them, under keys the owner holds, on an open protocol, is the data-sovereignty piece of the plumbing, and it is already open source.

Create the demand before the day you need it

Sovereignty is not something you have the day you decide to move. If a hyperscaler, a payment provider or a model API went away tomorrow, most European organisations would have no alternative but to stop, because no market for the alternative exists, and markets do not appear overnight. The single points of failure are real and larger than the public record suggests. The investigation GitHub owes its customers made the case that software failures are not documented the way the airline industry documents its own, so we only see the examples a company could not hide, and the picture is worse than the few cases that become public.

So the move is to create demand. Run parallel services. Buy the local or open alternative for part of the estate now, and accept that it has fewer features, because the point is not the features: it is that when the day comes there is somewhere to go, and that by then the alternative has had years of paying customers. Procurement is the lever a state already holds; it does not need a new law to prefer open schemas, data mobility and service mobility in what it buys. The business case writes itself, and it is a resilience case, not an ideological one.

Smaller models, more of them

The direction of travel in agents is to use less of the model, not more. Instead of agents with enormous permissions, which is why we wrote Agent Behaviour Policies to control them, you want the least agency that does the job, because that is what you can trust; Agency is not a yes is about where the agency should sit, and the supply-chain article's example is the right end state: the model is used once to build the tool, and production is plain files and plain code, with no prompt in the path. Frontier models are for working at the frontier, for finding the art of the possible; this article is being written with one, and it makes a difference here. For things that have to work, every day, at a price, a frontier model is far too expensive, far too much power for the task, and dangerous in proportion.

That means small models, and it means many of them. We are commoditising a level of intelligence, and most of the value is captured by a model that is good enough for one job, in one language, with one culture in it. There is a market, still at genesis on the map, for models for Portuguese, and not one Portuguese: the Portuguese of Portugal and of Brazil, of Angola, of Cape Verde, of Mozambique, of the Azores, of the Algarve, each with its culture in the training data and not just its spelling. The same for Romanian, for Estonian. Portugal released Amalia on 1 July 2026, a model built by a consortium of universities with 5.5 million euros of EU recovery money and published with its dataset and code under an open licence, as a base for public bodies, companies and researchers to build on. That is the right shape: a public model is a public road. I argued for this in Portuguese as a programming language in the AI era, and the open-weight economies show the economics: a company that can run a near-frontier model inside its own environment and build on it owns the workflow, the data, the tooling and the people around it, which is where the value is.

Constraints are the input

The race for GPUs, and the parallel race for the rare materials to build them, is a race that only benefits the players already in it; on a map it is very hard to compete there without a huge amount of capital. The better question is why so many GPUs are needed at all. Most of that compute is waste that the big players can afford. The lab that could not afford it did the interesting work: DeepSeek's V3 technical report says the model was trained in 2.788 million GPU hours on export-restricted H800 chips, chips whose interconnect was deliberately throttled, with custom communication code written around the constraint. Fewer and slower chips, not none, and a fraction of the compute a comparable Western run used, whatever one makes of the reported dollar figure. The constraint produced the innovation.

Simon Wardley has written for years about how China plays this game: directed investment, special economic zones, and a constant drive up the value chain by industrialising components, the innovate, leverage, commoditise pattern run at the scale of a state. We should learn from it and copy the mechanics, and do it with Europe's values and its focus on people: bring the citizens on the journey, make their environments better, and measure the result in lives rather than in parameters.

What I learned spending millions on a distributed workforce

I want to say something about the talent, because the argument above rests on it. When I was a CISO and a CTO I used Upwork a great deal; across the open source projects and the teams I ran we spent millions on it, and it was one of the most successful things we did. We built distributed teams and large execution teams, hired people who were as good as, and often better than, those we could hire the ordinary way, and delivered faster and more efficiently than was possible before. The reason it worked was that we invested first in the infrastructure around it: the first thing I would do was build a project management team, a structure team and an onboarding team, and then use Upwork to manage Upwork, so that the internal teams carried no overhead.

Two lessons from that, for a country. The first is that the infrastructure is the whole difference between a talent market that works and one that is a race to the bottom. People said hiring that way was buying the cheapest; it never was. It was finding the right talent at a price that was fair where they lived, good value for both sides, and a real career for the person, and it scaled. The second is where the jobs go. Today a good job means moving to a capital city, paying what the city costs to be there, and spending most of what you earn on the privilege, when what most people want is to go back to where they came from, where the family, the support and the lifestyle are better and the jobs are not. A small country that builds the plumbing makes the jobs available everywhere inside it, and that is a competitive advantage its big neighbours cannot copy quickly. Romania's technology sector is close to 8% of GDP with more than 250,000 professionals and over ten billion euros a year of service exports; the company that became UiPath was founded in Bucharest in 2005 and listed in New York sixteen years later. The talent, the energy and the hunger are not the constraint.

The playbook

The judo playbook in five moves: map the attrition, commoditise the plumbing, create the demand, put the money where the market is still genesis, regulate what settled, then map again. Three examples underneath of a constraint, a small state and a public model doing it.
  1. Map the attrition. Draw the chain from a talented person to a buyer abroad and mark every place commerce stops. Each is a dot in custom built that belongs in commodity.
  2. Commoditise the plumbing. Treat those dots as roads: pooled money, built once, open to all, open source and open schemas. Company formation, cross-border contracting, payments across Europe, language, contracts and tax.
  3. Create the demand. Buy the alternative before the day you need it. What the state pays for belongs to the commons. Run parallel services and accept fewer features so a market exists before the failure does.
  4. Put the money where the market is genesis. Small models for each language and culture; data that stays with its owner and moves; agents with the least agency and behaviour as policy. Not GPUs.
  5. Regulate what settled. Once a layer is a commodity, write the rule for it: data mobility, open interfaces, the report a failure must produce. The rule comes after the market, not before it.

Then map again, because a commoditised layer moves the opportunity up and the next attrition point is above the last one. That is the judo: each move uses the weight of the larger players, the commodity models, the commodity clouds, the commodity rails, rather than trying to match it.

What this does not claim

If you work in or with one of these countries, in an embassy, a ministry, a company or a university, and want to map the attrition for your own case, write to agent@riskmandate.ai.

Where this comes from

A voice memo of mine on 12 October 2026, after an event where I met people from Romania, written by this site's agent in the voice of this site. The argument and the editorial responsibility are mine. The earlier pieces it builds on are Europe's strategic opportunity in GenAI, An open-source sovereign cloud for an open Europe, Portuguese as a programming language in the AI era and Embracing the generative AI wave: a golden opportunity for Portugal (2023). The Ambassador role is the SGraph Send role definition of 12 February 2026. Figures: Eurostat on web sales by destination; the e-Residency dashboard and ERR via Remote Work Europe on Estonia; RIA and the EU Open Source Observatory on X-Road; Reuters via The Star on Amalia; the DeepSeek V3 report; Business Forum on Romania's sector; Wikipedia on UiPath; Simon Wardley on China's gameplay.

Threads

Startups & strategyAgents & policy This article as a graph →

Builds on

Continued by

All articles · All graphs

Want the next issue by email. One issue a week or so: what was published, what it adds up to, and what is worth your time. Subscribe to the SGit Newsroom →

← All articles