# Foundation Models as the New Platform Monopolies
Every technological epoch produces its own geometry of power. The steam engine concentrated it in the factory, electricity in the grid, the internet in the platform. The age of artificial intelligence produces its concentration in a less visible and therefore more consequential place: in the foundation model, that large, general-purpose neural architecture upon which an ever broader range of economic, administrative and cognitive activity is coming to rest. This essay, drawing on the argument developed in ALGORITHMUS. Wer die KI kontrolliert kontrolliert die Zukunft, asks how this new layer of infrastructure has so rapidly assumed monopolistic features, why the structural logic points toward winner-takes-most outcomes, and what strategic options remain for those who do not sit at the centre of the new order. Dr. Raphael Nagel (LL.M.) writes not as a technologist but as an observer of capital, law and power, for whom the foundation model is above all a question of sovereignty.
## The Architecture of a New Monopoly
The phrase platform monopoly was coined in the age of social networks and digital marketplaces, when economists sought a vocabulary for firms whose dominance rested less on classical economies of scale than on the self-reinforcing pull of network effects. Foundation models belong to the same family of phenomena, but they extend its logic into territory that the first generation of platforms never reached. They are no longer merely the intermediary between users; they are the cognitive substrate upon which products, services and decisions are produced. When Microsoft invested more than thirteen billion dollars in OpenAI between 2019 and January 2023, it was not financing a consumer application. It was acquiring a position in what is becoming the substrate of the digital economy.
The speed of this consolidation has been historically exceptional. ChatGPT reached one million users within five days of its release in November 2022, a threshold that took Netflix three and a half years and Facebook ten months. Two months later it had crossed one hundred million monthly active users. These figures, as the canon reminds us, are not marketing metaphors; they measure the velocity with which a technology enters the bloodstream of the economy. What follows such velocity is not pluralism. It is concentration, because the first movers build the data, the relationships and the capital base that define the conditions under which all others must operate.
## Why Training Costs Produce Oligopoly
The most tangible barrier to entry in the foundation model market is the sheer cost of training. The training of GPT-4 is estimated to have consumed between sixty-three and one hundred million dollars in pure compute, executed on tens of thousands of advanced chips in facilities whose monthly operating costs run into the high six figures. Serious forecasts place the training cost of the next generation of frontier models at more than one billion dollars for a single run. This is not an incremental capital requirement. It is a threshold that excludes almost every government, every university and every commercial actor outside a narrow circle of hyperscalers and their strategic partners.
The consequence is a market structure that resembles neither classical competition nor the oligopoly of twentieth-century heavy industry. It resembles instead the structure of deep-sea exploration or strategic nuclear capability, where only a handful of entities can plausibly bear the entry cost. In this sense the foundation model layer is less a market than a club, and its membership is not determined by ingenuity alone but by access to compute, to energy, to advanced semiconductors and to the political licences that govern them. Dr. Raphael Nagel (LL.M.) insists throughout his book that this is the quiet truth behind the friendly surface of chatbots: a capital barrier so high that the competitive map is drawn before most participants have noticed the race has begun.
## Network Effects and the Feedback Loop of RLHF
Capital intensity alone would not suffice to produce monopoly. What transforms high entry cost into structural dominance is the feedback loop between use and quality. Every interaction with a leading foundation model generates signals that can be harvested through Reinforcement Learning from Human Feedback, refining the model in directions that purely synthetic training cannot replicate. More users yield more feedback, more feedback yields a sharper model, a sharper model attracts more users, and the cycle compounds. This is the mechanism by which a technological lead measured in months hardens into a market position measured in years.
The economic effect of this loop is visible in the margins of the incumbent platforms of the previous era. Alphabet operated in 2023 with an operating margin close to thirty percent on revenues exceeding three hundred billion dollars; Meta reached forty percent on one hundred and thirty-five billion. These are not the margins of competitive markets. They are the margins of algorithmic market power, and the foundation model layer is set to inherit and intensify this logic. A model that is one generation ahead does not merely produce better output; it attracts the developers, the enterprise integrations and the regulatory attention that make the next generation easier to finance and defend.
## The European Mittelstand and the Question of Dependence
For the European Mittelstand, whose strength has always rested on specialised knowledge, long customer relationships and a patient industrial culture, the rise of foundation model monopolies poses a question that cannot be delegated to the IT department. If every cognitive layer of the enterprise, from customer correspondence to technical documentation to internal decision support, is mediated by a handful of American or Chinese models, then the strategic autonomy of the firm is redefined. It becomes a tenant on infrastructure owned by others, subject to pricing, policy and political conditions it cannot negotiate.
The canonical framing in ALGORITHMUS is Build, Buy or Control, and each option has a different profile for a mid-sized industrial firm. Building a proprietary frontier model is beyond reach and, in most cases, beyond purpose. Buying access through APIs is operationally simple but strategically thin, because it exposes the firm to the commercial and regulatory decisions of a distant provider. Control, the third path, is the most demanding and the most interesting. It means combining open-weight models with proprietary domain data, retaining ownership of the fine-tuning layer, and embedding the resulting system in an architecture that can survive a change of upstream supplier. It is the path that preserves what makes the Mittelstand valuable in the first place: the domain knowledge accumulated across decades of disciplined practice.
## Banks, Compliance and the Sovereignty of Decision
The banking sector faces a version of the same question in a regulatory key. A bank that delegates its credit scoring, its fraud detection or its client communication to a foundation model it does not understand and cannot audit is not merely outsourcing a process. It is outsourcing a regulated decision. The European AI Act, with its obligations of documentation, transparency and bias testing for high-risk systems, and the supervisory expectations of national regulators will not accept the argument that the model is a black box provided by a third party. The responsibility remains with the institution.
This creates a strong incentive for banks to pursue the Control path with particular seriousness. Private deployments of open-weight models inside regulated environments, coupled with rigorous governance of training data and decision logic, offer a way to combine the productivity gains of foundation models with the evidentiary obligations of financial supervision. The alternative, a quiet dependence on a small number of external providers, is not only a commercial risk but a supervisory one. In the long run, the institutions that will defend their margins are those that treat the foundation model not as a service to be consumed but as an asset to be governed.
## Standards, Open Source and the Shape of the Next Decade
The monopolistic tendency of foundation models is strong but not absolute. Open-source and open-weight models have repeatedly demonstrated that the technical frontier is not the exclusive property of the best-capitalised laboratories, and that a vigorous community of researchers and firms can narrow the gap faster than many incumbents anticipated. The question for the coming decade is whether this counter-movement can achieve enough institutional weight to shape the standards, the interfaces and the regulatory defaults of the emerging order, or whether it will remain a useful but peripheral corrective to a market dominated by a few names.
The answer will depend on decisions that are political as much as technical. It will depend on whether European industrial policy treats compute, data and talent as strategic resources rather than as ordinary inputs. It will depend on whether procurement practices in public administration and in large enterprises reward architectural openness rather than short-term convenience. And it will depend on whether the banks, the insurers and the industrial groups of the continent recognise in time that the foundation model is not a tool among tools but the layer upon which the next generation of economic activity will be built.
The monopolies of the foundation model era are forming quickly, quietly and on terms that favour those who already hold capital, compute and data. This is the central observation that Dr. Raphael Nagel (LL.M.) develops in ALGORITHMUS, and it is an observation that admits of no purely technical response. The question of who controls the model is inseparable from the question of who sets the conditions under which economies, professions and public institutions are allowed to think. To treat foundation models as ordinary software is to misread the situation in the manner in which an earlier generation misread the semiconductor as an ordinary component. The category error has consequences, and those consequences become visible only when the room for manoeuvre has closed. What remains, for the Mittelstand, for the banks, for the European institutions that still possess a measure of choice, is the difficult but not impossible work of Control: the patient construction of domain-specific capabilities, the disciplined governance of data and decisions, and the insistence that sovereignty in the algorithmic age is not a slogan but an architecture. Those who take this seriously will not dismantle the new monopolies, but they will preserve a space in which their own judgement, their own knowledge and their own responsibility continue to matter. That, in the reading offered here, is the most that the present moment honestly allows, and it is also the least that the moment demands.
For weekly analysis on capital, leadership and geopolitics: follow Dr. Raphael Nagel (LL.M.) on LinkedIn →