Engineering depth is the next step for Portuguese defence
Portugal has raised defence spending quickly and spends more of it at home. The next step is to keep engineering depth in the country while AI changes how software is built.
Portugal has raised defence spending quickly and has started to spend more of it at home. The harder part is to keep engineering depth in the country while AI changes how software is built. National defence depends on engineers, and on their work, staying in Portugal.
1. The software industry assumed a global world
For thirty years the software industry worked as one market. A team in Lisbon could write code on an American cloud, with American tools, for a German customer.
Portugal did well in this world. Its exports of IT and telecom services doubled between 2019 and 2023 (Banco de Portugal). ICT specialists are now 5.2% of employment, slightly above the EU average (Digital Decade 2025).
Few of the decisions in this model are made in Portugal. The customer and the infrastructure are usually abroad. That costs little until a supplier or a government restricts access.
2. Access is no longer guaranteed
In January 2025 the US Framework for AI Diffusion split the world into tiers for access to AI chips. Spain and France were in the first tier. Portugal, a member of the EU and NATO, was in the second tier, with caps. The rule was withdrawn in May 2025, but it showed that even an ally can limit what Portugal buys.
The EU now writes this risk into its own rules. Its defence loan programme SAFE limits components from outside the EU to 35% of the cost. The buyer must also be able to change the system without permission from a third country.
Portugal has acted quickly. The government brought the 2% of GDP defence target forward from 2029 to 2025 (ECO), and NATO estimates that Portugal reached it (NATO, 2026). Portugal was among the first countries to have its SAFE plan approved, for 5.8 billion euros. The plan puts drones, satellites and aircraft made in Portugal at its centre.
According to the Kiel Institute, defence spending raises growth when it buys new products developed at home. Portugal also hosts a NATO DIANA accelerator in Almada and a test centre for unmanned systems in Tróia (ECO).
3. Many engineers work on foreign products
About 75,000 Portuguese emigrated in 2023 and about 65,000 in 2024 (Observador). Many engineers who stay work for foreign companies. Firms with foreign capital made 85% of Portugal's IT and telecom services exports in 2025, according to Banco de Portugal (ECO).
In these companies the code and the patents usually belong to the head office. Portugal files 34 European patent applications per million people, against 153 in the EU (EPO, 2025). Most Portuguese unicorns have moved their headquarters abroad (ISCTE).
The IMF finds that Portuguese firms start small and rarely scale up. About 0.1% of new firms become fast-growing "gazelles", against about 0.5% in a typical European country (IMF, 2026). Portugal needs more companies that grow and keep their work at home.
Larger economies pay more and offer bigger projects. People and ideas move to them, by emigration or by contract.
4. Depth is hard to rebuild
Engineering depth means people who know why a system works the way it does, for example a compiler or a protocol stack. Such a team takes years to build and can fall apart after one reorganisation.
Most of this knowledge is never written down. A design document says what the system does. It rarely says which approaches failed, or why a parameter has its value. Senior engineers pass this on by working next to juniors for years.
A foreign development centre usually gets one part of the system, such as a module or a test campaign. Architecture decisions stay at head office. Amador and Stehrer (Banco de Portugal, 2014) found the same shape in the whole economy. Portugal adds value in the middle of chains that are designed and sold elsewhere.
When senior people leave, juniors lose their teachers, and the knowledge can disappear within a few years. The IMF finds that the hourly productivity gap between Portugal and the US has been stable since 1990 (IMF, 2026).
We think lost depth rarely comes back, because training people who understand a system takes far longer than buying the system.
5. AI makes the race harder
AI tools change how fast software gets built, and Portuguese companies have to use them to compete.
The frontier models are developed in the US and China, and so are most of the chips and data centres they run on. A Portuguese team that builds its work on these models depends on a foreign provider. The provider sets the price and the terms. The AI Diffusion rule showed how fast access can change.
Using these tools also means sending code and logs to the provider. Depending on the contract, the provider may store this material or use it to train the next model. Each team decides this alone. Together, these uploads move Portuguese engineering knowledge to foreign providers.
6. Full reliance on AI turns engineers into prompt writers
When a model writes the code and another model reviews it, the team stops learning how its own system works.
Junior engineers become senior by doing the hard part themselves, for example by tracking down a race condition. If a model does that work, the junior gets the answer without the understanding. Over the years the team loses the people who could check the model.
The prompts, corrections and accepted changes also go to the model's owner. A team in this position cannot work when the model is wrong or unavailable.
7. What Portugal can own
A small country can own the methods that check whether software is correct. This matches the direction the government has set for defence spending and for the PRR.
- Build on mathematics and algorithms. A frontier model needs a cluster of chips that Portugal may not be allowed to buy. A static analyzer or an SMT solver runs on one laptop. A small team can understand such a tool completely and teach it to the next team.
- Produce results that others can check again. An analysis or a proof gives the same answer on the same code every time, so an auditor can repeat it. The NATO software quality standard AQAP-2210 expects verification and test records that the buyer's inspector can review.
- Publish part of the work as open source. A vendor cannot withdraw public code, and engineers in any Portuguese company can learn from it.
- Put AI on top of verified facts. A model given checked findings, test results and runtime traces has less room to guess. The facts stay on local machines, and the team decides which model sees them.
This approach takes longer than wrapping someone else's API, and it keeps working when terms or export rules change. We build our own tools this way at Interpretica, in Portugal, under a PRR project. We have an obvious interest in this argument, so the numbers above link to their sources.
Portugal's defence budget has nearly tripled since 2014, from 2.3 to an estimated 6.8 billion euros in 2026 (NATO). The country has chosen to spend more of it at home. That choice lasts only if the engineers who understand the software stay in Portugal.