When we read articles or watch videos on YouTube about artificial intelligence, it is easy to come across a narrative full of promises: ‘the model will write code’, ‘the agent will complete the task’, ‘the company will grow thanks to automation’. But let us step one logical level down – to the server rooms scattered across the world. Let us feel their heat, see the space they occupy and, as if following the cables – both internet and power – trace our way to their owners, operators and clients… Let us see how it all works and who pays for it.
This text accompanies the research on language models that I am currently conducting. In another article I looked at aspects of the technology itself as well as the science and innovation behind AI. Here I would also like to use a geographical axis of comparison between the United States, Europe and China. If you would like to know how the dollar circulates in the United States, why China triggered a price shock and why Europe is turning towards ‘sovereign AI’ for companies and institutions – read on.
AI is driven by powerful infrastructure. My investor’s instinct urges me to peer through the fogged-up windows and see how it all fits together.
The United States and ‘financial spaghetti’
Have you ever heard of ‘circular financing’? The mechanism is relatively simple although in the view of some experts it ultimately creates a financial bubble. Technology giants invest in AI companies and providers of computing power. These companies – often relatively young and much smaller – then spend a large portion of the capital they have raised on infrastructure (servers, GPUs and data centre services) frequently purchased from… the very same giants who had just become their shareholders.
The relationship between NVIDIA and CoreWeave is a fairly vivid example. NVIDIA took an equity stake in CoreWeave and CoreWeave buys thousands of accelerators precisely from NVIDIA. Revenue grows on NVIDIA’s books. On the stock market the dominant narrative is that ‘demand is insatiable’. CoreWeave can pledge the hardware as collateral for debt and expand its computing farms even faster as a result. This is not fraud – it is simply a very clever financial structure.
Circular financing in the United States resembles intricately intertwined strands of spaghetti where capital and infrastructure drive successive valuations.
A similar circulation of cash can be seen among the largest providers of computing infrastructure. Microsoft and Amazon invest money in companies building models after which a significant share of that money returns to them as fees for using their data centres. When OpenAI announced another funding round at the end of February 2026 alongside long-term infrastructure agreements with Amazon (AWS) and NVIDIA it became very clear how closely hardware and capital are linked here.
This ‘spaghetti’ has two side effects: at every stage of the flow of money revenue is recorded and the market receives fuel to push valuations higher for example on the Nasdaq. But how real is that revenue? And everything works perfectly as long as capital keeps flowing.
And that is not all. The market is currently investing in AI as if it already assumes that within a few years a new multi-trillion-dollar segment of the economy will emerge. Yet the maths is brutal: the costs of data centres, chips, energy and cooling are for now rising faster than the recurring profits from services that provide access to AI. (Spoiler alert – we will consider whether the model of charging a monthly subscription will remain with us for good.)
OpenAI is a useful reference point. Its financial results suggest that revenue in 2025 hovered around $13 billion. At the same time based on the February transaction mentioned above OpenAI’s valuation can be considered to be $730 billion – at least according to the official statement on openai.com (in February 2026 alone SoftBank, Amazon and NVIDIA invested a combined $110 billion in OpenAI which is 8.5× annual revenue).
If we compare such a valuation with revenue the obvious question arises: where will the money come from to cover the costs and allow investors to recover their funds? We already know that some of it will return to certain investors in the form of purchases. But that cannot be the whole story. For billions of dollars in infrastructure investment to make economic sense the industry would at some point need to generate not tens but hundreds of billions in revenue. That is clearly not visible yet. What is visible is capital expansion and… an ever-growing appetite for electricity (more on that later).
If companies fail to show a credible path to positive cash flow relatively quickly we may witness a hard slowdown: reduced investment, layoffs, frozen projects and in the worst case problems with a debt spiral. For many investors the year 2000 and the dot-com bubble remain vivid in memory. The comparison is certainly not perfect but the mechanism of market expectations is quite similar: a narrative about a ‘new economy’ followed by a large-scale reality check.
Incidentally it is partly by observing capital flows that I base my assumptions about AGI which – whenever I discuss them – often meet with scepticism from other experts. I remember for instance a conversation with the IT director of one of the banks. Our estimates diverged significantly and although almost two years have passed since that discussion I still stand by my view. In my opinion if the AGI stage is not reached within the next few years (and I mean genuinely soon – on the order of two or three years) these investments could collapse. A great deal has been placed on a single bet.
In addition some companies in the AI sector are considering going public. Before an IPO almost everyone has an interest in making the valuation look as strong as possible.
The big can do more
The American model is based on concentration: of capital, infrastructure and people (in the previous article I wrote that even knowledge tends to concentrate around companies rather than universities – this is a simplification, read the full piece). In generative AI it is relatively easy to build a ‘scale advantage’ – whoever has more high-quality processors and better energy contracts can train, test, deploy and ultimately sell their solutions faster. Naturally this comes at a cost.
Large players behave like providers of general-purpose infrastructure for the rest of the market: without their data centres, networks and platforms it is difficult to build products that use AI. The barrier to entry is also rising because the cost of training a top-tier model can reach hundreds of millions and sometimes approach a billion dollars depending on the approach, the data and the number of experiments.
The concentration of resources in the hands of technology giants means rising barriers to entry for smaller players in the AI market. Microsoft for example builds large data centres out of containers – if something fails the entire container is replaced.
Smaller teams can compete through quality, interesting ideas for model architectures, data or specialisation but they will not win the ‘scale race’ without capital. For now however this seems to apply more to Europe because when it comes to China they use both ingenuity and vast amounts of money.
The ‘DeepSeek effect’ and the price war
China is challenging the logic of the American SaaS model by offering low-cost solutions. Instead of playing the game of ‘who has the bigger server room’ they have focused on higher efficiency. DeepSeek has become a symbol of this approach: the V3/R1 models were presented as offering capabilities close to Western systems at much lower training and inference costs. The capabilities are not quite that straightforward though. What is more, analyses have been circulating online suggesting that the full costs may have been higher than officially stated. But the impression has stuck.
The biggest blow to the West comes not from benchmark results but from price lists. Chinese providers have started offering APIs at prices that force margin pressure.
China’s AI strategy is built around cost efficiency and price pressure which is reshaping the Western SaaS market. As a result AI can be widely used to automate a range of tasks. We are beginning to see the so-called commoditisation of AI.
For example: the rates for the very high-quality Qwen model are on the order of fractions of a dollar per million input tokens and a few dollars for output tokens – I mean the prices when you use the model directly from the creator; it is open so it may be even cheaper from other providers. DeepSeek is cheaper still. For OpenAI’s models the rates can be significantly higher especially for output tokens in the strongest (and oldest ;)) variants of GPT and oX. Europe is also trying to catch up on price.
Status as at 3 March 2026 (I tried to pick the closest possible models):
| Model | Country | Rate per 1m input tokens | Rate per 1m output tokens |
|---|---|---|---|
| GPT-5.2 | USA | 1,75$ | 14,00$ |
| Claude Sonnet 4.6* | USA | 1,50$ | 15,00$ |
| Mistral 3 Large | Francja | 0,50$ | 1,50$ |
| Qwen 3.5 Plus | Chiny | 0,40$ | 2,40$ |
| DeepSeek V3.2 | Chiny | 0,28$ | 0,42$ |
* Price for prompts up to 200k tokens. Longer prompts are more expensive.
If AI accessed via API is getting cheaper I will venture a thesis (overheard and, since I agree with it, borrowed): language models are starting to resemble a resource like electricity or tap water; we will treat them as something obvious that we barely think about. Such a commoditisation process is dangerous for the West because it undermines the classic SaaS economics.
In China state-backed funds focused on strategic sectors play a significant role. In such a system profitability does not have to be calculated quarter by quarter; what matters is that the sums add up over the long term. More broadly China is betting on the state’s ability to build capabilities vertically, aiming for the strongest possible position of entire sectors.
On top of that there is vertical integration within corporations: Alibaba, Tencent, Huawei and other players can combine hardware, super-apps (such as WeChat, which is used for a great many things including payments) and services. This structure makes it possible to shift resources between business lines and for a time subsidise costs in order – again – to gain a large market share in the long run.
Europe: a capital gap and a flight into B2B
On the Old Continent there is a lack of large-scale private funding and there is not much government or EU funding either. Some ambitious companies therefore move their headquarters and operations to the United States because it is easier there to raise multiple rounds of financing which in Europe is a rarity. One example is ElevenLabs, founded by Poles in the United States.
The response of European companies within the existing financial environment (and also the regulatory one, more on that in another text) is pragmatic: if we are not going to win the mass consumer services segment let us focus on the public sector and industry where regulation, data confidentiality and the ability to run models in an isolated environment matter more. The term ‘sovereign AI’ has even been coined – an approach in which an institution has control over the data and the model rather than just over the invoice for the service.
In their business narrative European LLM builders focus on AI sovereignty, offering services among others to the public sector, which often requires AI to be hosted on its own infrastructure.
Companies such as Mistral and SAP are building offerings aimed at organisations that do not want to send sensitive data to external providers. European regulations (including the AI Act) reinforce the need to follow this path.
The road to the agentic economy
In this section I will touch on a thread which – I remember well – I jotted down on my reMarkable a few years ago while travelling by train. If I had written about it then people would have thought I had lost the plot. Now is a good moment and the view is no longer quite so isolated (although at some point my thinking takes a turn in a different direction which I may dare to write more about soon). For now though let us stick with AI.
Many AI tools today are sold on a ‘per person’ or ‘per month’ price. It is a model that is convenient for the provider and predictable for the customer. The thing is that when AI agents carry out tasks autonomously pricing ‘by outcome’ starts to make more sense.
I think that over time a significant part of our work will boil down to delegating and supervising tasks assigned to AI agents. We will increasingly hear about billing for their work on an outcome basis.
Instead of paying a fixed amount for the ability to use a tool companies will more and more often pay per resolved customer ticket, a correctly prepared report, a deployed piece of code or a closed case. On the one hand this shifts risk from the customer to the provider: if the agent does not deliver (as people in IT corporates like to put it) it does not earn. On the other hand a provider with an excellent solution to a specific problem can win customers whom it would not have been able to persuade under another model due for example to their small scale of operations.
A ley feature will be reliability – repeatability and predictability of results. The market will start to reward systems that deliver rather than those that ‘talk’ nicely but lose the thread and make facts up.
Naturally classic subscriptions will remain.
Electricity sets the ceiling
It is hardly a revelation any more that data centres require vast amounts of energy. Transmission grids in many regions of the United States are overloaded and connection processes take years. No wonder technology companies have started looking for firm power – without queues and without risk.
In 2025 there was a lot of noise about AWS’s agreement with Talen Energy for 1.92 GW of electricity from the Susquehanna nuclear power plant along with a dispute among regulators over the supply arrangement (the ‘behind-the-meter’ variant and solutions that sidestep the classic grid-connection model). Another example is Microsoft’s deal to purchase 100% of the electricity from Unit 1 at the Three Mile Island plant in Pennsylvania which had previously been shut down.
The role of electricity as a strategic resource highlights risks for AI companies especially in the context of energy inequality.
Europe also has a problem with an ageing transmission grid and power supply can be unstable because a relatively large share of generation depends on the weather. I would suggest googling what cross-border electricity trading between countries looks like. At the same time we do not have huge corporations that could start building nuclear units. Most of the driving force lies with the governments of individual countries. Because everything takes a very long time I have a serious concern that an insufficient supply of electricity will be another bottleneck for European companies even if we somehow solve the capital problem.
Let us return to the United States because there we can in a sense see the privatisation of energy being carried out in an interesting way – by pulling the best, most stable sources of supply out of the market for one’s own needs. If this trend accelerates (for now it is only in its infancy) the rest of the economy may be left with more expensive energy and greater uncertainty of supply. That too is a risk and should not absolve the US government of investment at the federal level (or states at state level).
China is in a completely different and – I will venture – in my view the best position in terms of the amount of power generated and the stability of its sources. This is not a text about ecology (I like trees, I care about animals etc.) but about economics which will chart the future of AI. If we add the energy back-up to everything I have already written in this and other texts who knows whether the winner in the AI race will not be the Chinese…
Just to note: someone also has to build these data centres and transmission networks so it is a good environment for companies in that sector.
What this means for us, living in Poland
If we look at the United States we see a system in which capital circulates in loops, valuations rise faster than profitability, the ‘winner takes all’ logic reinforces the big players and forces smaller ones to look for niche paths and, to some extent, interdependence.
As for China – we see price pressure and an ‘AI as infrastructure’ model fuelled by state money and vertically integrated conglomerates.
Europe and Poland – we need a good plan. In this whole puzzle it is worth finding a place for ourselves and choosing a coherent strategy.
When we analyse Europe we see a lack of large private investment fuel and an attempt to build an advantage in sectors where regulation, confidentiality and control of infrastructure matter. (This direction also fits well with the narrative that AI is increasingly starting to operate in the physical world which strengthens the importance of safety and responsibility – I recommend the text on embodied AI.)
And when we think about the future of business models what comes to mind is a shift from paying ‘for access’ towards paying ‘for completion’ as well as a fight for energy as a strategic resource. If at some point the cost of a kilowatt-hour, cooling and hardware depreciation starts to matter so that revenues have to balance expenditure the narrative of a revolution may collide sharply with the financial statements.
If that happens and newspaper headlines start screaming that this whole AI thing was pointless you should still keep building your skills. Because even the Gartner hype cycle predicts a moment of mass disappointment. The dust will settle however and the players who somehow managed to make the Excel spreadsheet add up will remain.
I am keeping my fingers crossed for Polish and European AI models and companies. Perhaps one dimension of ‘sovereign AI’ would be directing our capital towards ourselves (a topic to google: how much European capital is on the US stock market). Because in rivalry with China cleverness and innovation alone will not be enough although naturally I wish us those as well.
As always I invite you to subscribe to the newsletter.
