# Chips, Cloud, Control: Who Owns the Hardware Owns the AI
There is a peculiar inversion at the heart of the current technological moment. The more intangible the output of artificial intelligence appears, the more physical its foundations become. Behind every generated sentence, every classified image, every predictive score stands a precise arrangement of silicon wafers, lithography machines, copper cabling, diesel generators and water cooling loops. The cloud, despite its name, is not a cloud. It is a geography of concrete buildings in specific jurisdictions, powered by specific grids, supplied by specific fabs, financed by specific balance sheets. In my book ALGORITHMUS. Wer die KI kontrolliert kontrolliert die Zukunft, I argue that the question of who owns this layer is not a technical question but the central power question of the twenty first century. This essay extends that argument for readers who allocate capital, and who increasingly sense that the hardware stratum of AI has become an asset class of its own.
## The Return of the Physical
For more than two decades the dominant narrative of the digital economy was the progressive dematerialisation of value. Software was said to eat the world, platforms were praised for their asset light balance sheets, and intangibles became the preferred vocabulary of equity analysts. Artificial intelligence has quietly reversed that movement. The frontier of value creation has migrated back into the physical: into fabrication plants, into lithography halls, into data centre campuses whose footprint is measured in hectares and whose electricity draw rivals that of small cities. What looked like a cloud has turned out to be a continent.
The numbers in the canon of my book are unambiguous. NVIDIA's market capitalisation exceeded three trillion dollars in 2024, surpassing the combined market value of all DAX companies. Its data centre revenue quadrupled within twelve months, from 4.3 billion dollars in the third quarter of 2022 to 18.4 billion dollars in the third quarter of 2023. These are not the figures of a speculative fashion. They are the revaluation of an industrial base whose strategic role the market had previously underestimated, and which now reasserts itself with the gravitational force of a new standard of account.
## Three Companies, One Bottleneck
The hardware layer of artificial intelligence rests, at its upper end, on an extraordinarily narrow foundation. Three firms define the frontier. TSMC in Taiwan fabricates roughly ninety percent of the world's advanced logic chips. ASML in the Netherlands is the sole producer of the extreme ultraviolet lithography machines without which these chips cannot be manufactured; each unit costs around 150 million euros, contains more than 100,000 parts from over 800 suppliers, and is produced in annual quantities of only fifty to sixty. NVIDIA in the United States designs the graphics processors on which frontier models are trained; its H100 chip sold in 2023 for between 25,000 and 40,000 dollars and remained effectively unavailable for months.
From an allocator's perspective, this is not a diversified supply chain. It is a geographically concentrated, politically exposed triangle whose disruption, for any reason, would cascade through every dependent layer above it. The American export control regime of 7 October 2022, which I analyse in detail in the book, codified what had previously been implicit: the hardware bottleneck is a sovereign instrument. When Washington restricts the sale of advanced chips to China, when it bars American nationals from working for specified Chinese semiconductor firms, it is not conducting trade policy in the traditional sense. It is exercising a form of pre military deterrence through the control of compute.
## The Hyperscaler Oligopoly
One layer above the silicon, a second concentration has formed. The training and serving of frontier models is economically viable only on infrastructure that a handful of hyperscalers can provide. Microsoft's cumulative investment in OpenAI of more than thirteen billion dollars is not, in substance, a venture bet. It is the monetisation of Azure capacity through a preferred tenant whose workloads fill the racks that Microsoft has built. Google, Amazon and Meta have pursued structurally identical strategies, with Amazon's investment in Anthropic rising to four billion dollars and Google committing three hundred million in the immediate aftermath of ChatGPT's release.
The consequence is an oligopoly whose members are simultaneously landlords, financiers and competitors of the model labs they host. For any industrial or financial actor outside this inner circle, access to frontier compute is mediated by commercial terms that are negotiated, not discovered. This is a classical structure of infrastructure power, familiar from railways, pipelines and undersea cables, and it deserves to be analysed with the same seriousness. The cloud is not a market in the textbook sense. It is a regulated commons without a regulator.
## Capital Intensity as a Moat
The training of GPT-4, according to estimates cited in my book, cost between 63 and 100 million dollars in compute alone. The next generation of frontier models is projected by Epoch AI to require investments exceeding one billion dollars per training run. These figures are not incidental. They define the boundary of the field. A firm that cannot commit a billion dollars to a single experiment, and absorb the possibility that the experiment yields a model inferior to the previous generation, cannot participate at the frontier. The barrier is not intellectual. It is financial, and therefore structural.
For Dr. Raphael Nagel (LL.M.), this observation has a specific investment implication. Capital intensity of this magnitude produces, in the medium term, two consequences that allocators should price explicitly. The first is consolidation: the number of organisations capable of training frontier models will remain in single digits globally, and their bargaining power with respect to downstream users will therefore be considerable. The second is substitution risk: the same capital intensity that protects incumbents from new entrants also exposes them to architectural breakthroughs that devalue existing compute stockpiles. The H100 that is indispensable in 2024 may be an expensive museum piece by 2028. Moats built of silicon depreciate faster than moats built of brand or contract.
## Infrastructure as an Asset Class
What emerges from this analysis is the contour of a new asset class, although the term itself requires some care. Hardware exposure in the age of AI is not simply a sector allocation. It spans the listed equity of chip designers and fabricators, the private debt of data centre developers, the long term power purchase agreements that underwrite their operation, the industrial real estate on which the campuses are built, the water rights that cool them, and the sovereign instruments through which states subsidise or restrict their growth. The European Chips Act commits 43 billion euros through 2030. The American CHIPS and Science Act provides 52.7 billion dollars in direct subsidies alone. These are not line items. They are the terms of a new industrial policy that is redrawing the map of strategic capital.
For pension funds, family offices and sovereign allocators, the practical question is no longer whether to hold exposure to this layer, but how to hold it without concentrating in the most visible names at the most exposed valuations. A portfolio that owns NVIDIA through a technology index and nothing else has taken a single equity bet and mistaken it for infrastructure exposure. A portfolio that combines fab capacity, lithography, power, cooling, land, and selected sovereign debt of chip producing jurisdictions expresses a thesis about the physical substrate of intelligence. The difference between these two postures is the difference between following the narrative and understanding the structure.
## Sovereignty, Dependency and the European Question
Europe occupies a particular position in this architecture, and it is a position of considerable discomfort. The continent hosts ASML, one of the three irreplaceable nodes, and yet it lacks leading edge fabs, frontier model laboratories and hyperscale cloud champions of comparable weight. Its regulatory output, codified in the AI Act and in adjacent instruments, is globally consequential, but regulation without infrastructure is a form of influence that depends on the continued willingness of others to accept it. The automotive semiconductor crisis of 2020 to 2023, which cost the global industry more than 210 billion dollars in 2021 alone, was a preview of what structural hardware dependency looks like when it meets an external shock.
I have argued, and I repeat here, that the European response cannot be limited to defensive regulation. It requires a deliberate build out of domain specific compute, data and talent capacity, anchored in the industrial strengths that the continent still commands: mechanical engineering, pharmaceuticals, chemistry, specialised manufacturing. The strategic opportunity is not to replicate the hyperscalers, a contest already lost on its own terms, but to convert decades of proprietary industrial data into specialised models that no general platform can match. That opportunity, however, closes quickly if the underlying compute is rented from jurisdictions whose export controls may change by executive order.
The layer of chips and cloud is the layer at which the future of artificial intelligence is being allocated, in the precise financial sense of that verb. Capital flows to it, sovereign attention concentrates on it, and the terms of access to it will determine which firms, which sectors and which states remain authors of their own strategies rather than tenants in someone else's platform. For Dr. Raphael Nagel (LL.M.), the task of the serious allocator is neither to celebrate this concentration nor to lament it, but to understand it with sufficient precision to position capital, governance and time horizon accordingly. The hardware stratum will not become less important as models become more capable. It will become more important, because every increment of capability will be paid for in silicon, electricity and cooling water. Those who treat this layer as a technical detail will find, some years hence, that they have delegated the most consequential decisions of the decade to counterparties whose interests they never fully mapped. Those who treat it as the foundation it is will still face risk, but they will face it with open eyes. The algorithm belongs to someone. So does the machine on which it runs. The question of who, in each case, is the same question, and it is the one that this decade will answer whether we engage with it or not.
For weekly analysis on capital, leadership and geopolitics: follow Dr. Raphael Nagel (LL.M.) on LinkedIn →