I think compute will become a more important measure of national power than GDP.
The country with the larger economy will not necessarily be the country that can develop new technology faster, adapt its industry sooner, or sustain its essential systems when access to foreign AI becomes uncertain. An increasing share of that advantage will depend on the computational capacity it can actually use.
By compute, I mean the processing capacity needed to train and run AI: the chips, servers and connected systems that turn models into working services. For a country, the relevant quantity is how much useful work that infrastructure can deliver, and how reliably the country can call on it.
Over the next decade, I expect this capacity to move toward the center of economic and geopolitical competition. The reason is the growing range of valuable work that AI could perform, and the speed at which successful systems could be deployed across an economy.
GDP will remain essential for measuring economic activity. It tells us the value of production within a country over a period of time. That production supports incomes, tax revenues, investment and defense. But GDP was never a complete inventory of a country’s capabilities, dependencies or freedom to act. Its definition does not include a test of who can switch off an essential service.
As intelligence becomes something we can deploy through infrastructure, that omission becomes much more consequential.
Consider two countries with similar GDP.
Both have successful businesses, universities and public services. Both use advanced AI. In one, domestic institutions can operate capable models on infrastructure they control, buy additional capacity from several suppliers, and expand a functioning electricity system. In the other, most critical AI workflows depend on a few external services, with little ability to move them elsewhere.
Their economic statistics might look similar today. Their options during a supply disruption, a diplomatic dispute or a sudden surge in demand would be very different.
They would also face different opportunities in ordinary times. One could decide to give its researchers more computational capacity, subsidize access for new firms, or prioritize a national engineering project. The other would need those decisions to fit someone else’s commercial and technical constraints.
That difference is a form of power.
For most of modern history, expanding a country’s supply of skilled cognitive work required educating more people, attracting talent and giving those people better tools. These remain important. But when AI can perform a task reliably, additional instances of that task can be run without waiting for another generation of specialists to graduate.
A useful system can be deployed across many teams. More capacity can support more experiments, more software development, more design alternatives and more analysis. The work still needs direction and verification, and physical experiments still take time. The advantage comes from increasing the volume of work that meets a useful standard.
Yet even partial success changes the strategic calculation. A country able to run many more productive attempts can shorten the time between identifying a problem and finding a workable solution.
That matters in pharmaceuticals, energy systems, manufacturing and defense. The advantage can accumulate: better tools help engineers improve infrastructure, which supports further development and deployment. I expect access to effective compute to become one of the strongest predictors of how quickly a country can improve its other capabilities.
This is the reason I expect it to overtake GDP as an indicator of power. A larger existing economy gives a country more resources. The ability to accelerate research, engineering and coordination can change how quickly those resources become new capabilities. If AI makes that process substantially faster, the rate at which a country can adapt becomes decisive. Compute would influence the machinery of economic growth itself.
GDP will eventually reflect some of the resulting production. By then, the capacity that made it possible may already be concentrated elsewhere.
There is a reasonable objection here. Rich countries can buy computing services, just as they buy other inputs. Why should every country build its own?
They should certainly trade. Buying access can be cheaper and better than operating everything domestically. International suppliers and alliances can substantially increase a country’s capabilities.
But the ability to pay and the ability to obtain a critical resource are different things. A budget cannot instantly create a grid connection, manufacture unavailable equipment or change the terms under which a supplier may serve you. When supply is constrained, the countries that already have operational capacity possess an option that money alone cannot immediately reproduce.
A serious compute strategy gives a country several credible ways to keep working.
Domestic capacity is one of them. Diversified suppliers, usable models, technical expertise and agreements with allies are others. For Poland, building within a European system makes far more sense than trying to reproduce the entire semiconductor supply chain behind a national border. We should still know which capabilities are ours to operate and which depend on continuing permission from elsewhere.
The address of a data center does not settle that question.
A foreign-owned facility can bring valuable investment, infrastructure and services. Its location alone does not tell us whether Polish researchers can afford to use it, whether critical public services have guaranteed capacity, or who decides what happens during a shortage. Domestic ownership alone is also insufficient if the operator cannot maintain the systems or replace a crucial supplier.
The practical questions concern access, operating rights, jurisdiction, maintenance and the ability to move workloads. Those belong in the definition of national compute capacity.
Ownership also determines who benefits.
In my report Poland Facing AGI, I argue that infrastructure, public finance and ownership have to be considered together. A country could experience rising AI-assisted production while much of the income from the underlying capital flows to owners abroad. Workers and the state would not automatically receive a proportionate share of the gains.
Anthropic’s economic scenarios illustrate why this deserves attention: under its most transformative assumptions, rapid GDP growth can coexist with severe disruption to knowledge workers’ employment and wages. These are conditional scenarios, not an established forecast. They do show why a larger GDP number cannot answer every question about economic security.
A national compute strategy therefore needs a public return. That might include broad access for businesses and researchers, better public services, public equity where taxpayers bear investment risk, or wider citizen participation in ownership. The arrangement can vary. The benefit should be specified before public support is committed.
All of this ultimately requires electricity.
The IEA’s 2026 outlook projects global data-center electricity consumption rising from about 485 TWh in 2025 to 950 TWh in 2030. Those figures cover all data centers; consumption at AI-focused facilities is projected to triple over the same period. They are projections, with substantial uncertainty about deployment and efficiency.
The physical requirements exist regardless of which forecast proves closest. Servers need dependable power, cooling and connections. Building the electricity infrastructure takes time, and the IEA identifies bottlenecks in grids and equipment as constraints on the expansion of AI capacity.
This is why energy policy belongs at the center of an AI strategy. The price, reliability and availability of electricity help determine how much computational capacity a country can operate competitively, and how quickly it can add more.
In Poland Facing AGI, I proposed a national ambition of 1 GW of IT capacity by 2030. That is a proposal, not a claim about funded or installed infrastructure. Nor does one gigawatt translate into a fixed number of intelligent tasks: the hardware, models and applications matter enormously.
It does make the energy requirement concrete. If the IT equipment drew a full 1 GW continuously for a year, it would consume 8.76 TWh, before cooling and other facility overheads. Actual consumption would depend on utilization. A government adopting such an ambition would need a credible plan for generation, grid access and operating costs alongside its investment plan for servers.
It would also need to explain how the capacity will be used.
Counting chips would be a poor substitute for counting GDP. Chips differ in performance, memory and efficiency. Systems differ in networking and software. A cluster designed for a large training run serves a different purpose from distributed infrastructure running everyday AI services.
Even the available inventories are incomplete. Epoch AI’s methodology explicitly describes gaps in global coverage and uncertainty about who uses particular facilities. A confident league table can conceal a great deal of missing information.
I would start with a national dashboard, with published definitions and independent verification:
Useful output. Test a common set of tasks, at a specified level of quality, speed and human oversight. How much work can the available systems actually do?
Operational capacity. Report what is already running separately from construction projects and announcements. What exists now, and what might become available later?
Access. How much capacity is available to domestic users, at what cost, and under what conditions? Can businesses, researchers and public institutions actually use it?
Resilience. Test what happens during a supplier or service outage. How much essential work continues under stress?
Electricity. Measure its cost and reliability, and the time needed to connect new capacity. Can the system operate competitively and expand?
Report the total, the amount per person and the rate at which usable capacity is growing. Keep hardware performance and application results visible separately, so better software does not disappear inside a count of machines. Some sensitive details would need protected reporting, but that should not prevent public scrutiny of the overall program.
The OECD has already set out a framework for national compute capacity covering availability and use, the conditions that make capacity effective, and resilience. We have a basis for measurement. We do not need to invent a single impressive-looking number before beginning.
My forecast depends on continued progress in useful AI capabilities. If that progress stalls, the case for compute overtaking GDP weakens. Another challenge comes from success: AI may become so efficient, inexpensive and widely available that compute stops being an important source of advantage. If adequate capacity is abundant, reliable access is universal and switching suppliers is easy, ownership of additional infrastructure will matter less.
That would be an excellent outcome. It is also a condition we should measure, rather than assume.
Cheaper execution of existing tasks can coexist with growing demand for more ambitious ones. A country may need less compute for routine translation while choosing to spend far more on engineering, scientific models and complex automated work. Whether supply outpaces those demands remains an open question.
My expectation is that the frontier of useful work will keep moving, and that countries able to supply and deploy more effective compute will retain a substantial advantage. Their success will also depend on institutions, talent, capital and industrial capacity. Compute becomes especially powerful through its ability to improve how those other resources are used.
For Poland, this argues for building now, with a program capable of learning as it expands.
We should secure sites and electricity connections, develop operating teams, and bring capacity online in stages. Universities, industrial firms and public institutions should be able to put it to work. Procurement should leave room for changing hardware and models. Every stage should report useful output, costs and access conditions, with funding for maintenance and replacement included from the beginning.
Strategic value can justify investment that takes longer to pay back than a commercial investor would accept. It cannot justify concealing poor performance. We should be able to explain both the economic value of everyday use and the value of having dependable capacity when outside supply becomes uncertain.
Buying hardware is only the beginning of that obligation. An idle cluster does little for national power. An operational system that helps thousands of firms improve their products, enables research and supports essential services can strengthen the country well beyond its own revenues.
Waiting has costs too. The engineers who learn to operate these systems, the companies that learn to build on them, and the institutions that learn to purchase useful outcomes accumulate experience. That capability will take time to develop even if hardware becomes cheaper next year.
I would begin with an exercise that makes our present position visible. Ask the institutions running critical AI-dependent workflows to demonstrate what happens if their principal external provider becomes unavailable for thirty days. Which services continue? Which can move, at what cost and with what loss of capability? How long would it take to restore the rest?
Then publish an investment program tied to the gaps the exercise reveals, with named operators, delivery dates and tests of whether the promised capacity actually works.
A country that can answer those questions has a much firmer basis for claiming technological sovereignty. As AI takes on more consequential work, I expect that demonstrated ability to count for more in the balance of power than a higher place in the GDP table.