# AI as Asset Class: How Institutional Investors Allocate Intelligence
When Microsoft committed ten billion dollars to OpenAI in January 2023, the transaction was recorded in the financial press as an investment. It was something else. It was the public ratification of a hypothesis that had been forming in private capital councils since at least 2020: that artificial intelligence constitutes not a sector, not a theme, not a technological cycle comparable to cloud computing or mobile, but an asset class in its own right. This essay, grounded in the book ALGORITHMUS. Wer die KI kontrolliert kontrolliert die Zukunft, examines what it means for institutional allocators to treat intelligence itself as a category of capital.
## The Repricing of Intelligence: What 2022 to 2024 Revealed
The numbers from the canon are, on their own, the clearest expression of what happened. OpenAI moved from a valuation of one billion dollars in 2019 to twenty billion in 2021, seventy billion in early 2023, and above ninety billion later that year. NVIDIA's data center quarterly revenue rose from 4.3 billion dollars in the third quarter of 2022 to 18.4 billion dollars twelve months later. The share price of NVIDIA increased by 240 percent in 2023, and its market capitalisation passed three trillion dollars at the end of 2024, exceeding the combined capitalisation of all DAX companies. In 2023 alone, 91.9 billion dollars flowed into AI startups, even as the broader venture capital market contracted by more than thirty percent.
These figures are not anomalies in a familiar landscape. They are the signature of a structural repricing. When capital behaves this way, it is no longer responding to a product cycle. It is responding to a recognition that the underlying substrate of economic activity has shifted, and that the instruments through which value will be captured in the coming decade are being assembled in real time. What appears to the uninformed observer as speculative excess is, at closer reading, the market translating a power shift into price.
Dr. Raphael Nagel (LL.M.) argues in the canon that this movement of capital is of a different order than the dot-com cycle or the mobile revolution. The ratio of allocation to AI within US venture capital, which exceeded forty percent in 2023, is without precedent. Such concentration reveals that professional allocators have already made the conceptual leap: they are no longer treating AI as a sector within technology, but as a category alongside technology.
## Three Layers, Three Thesis Families
An asset class is defined by its internal differentiation as much as by its external boundaries. Equities are not a monolith; they divide into geographies, capitalisation bands, styles. Real estate distinguishes between core, value-add and opportunistic. AI as an asset class, in the framework developed across the canon, resolves into three distinct layers, each with its own risk profile, its own capital intensity, and its own competitive dynamics.
The first layer is infrastructure: semiconductors, foundries, lithography, energy supply to data centres. This layer is characterised by extreme concentration and extreme capital intensity. TSMC fabricates approximately ninety percent of advanced logic chips. ASML holds a monopoly on EUV lithography, producing fifty to sixty machines per year at roughly 150 million euros each. The investment thesis here is infrastructural scarcity combined with geopolitical exposure. Returns follow the logic of essential bottlenecks.
The second layer is the foundation model: the proprietary large models produced by OpenAI, Anthropic, Google DeepMind, Meta and a narrowing set of Chinese laboratories. Epoch AI projects that training a frontier model will soon exceed one billion dollars per run. This cost structure confines competition to a handful of globally operating actors and produces the platform monopoly dynamics analysed in Part II of the canon. The investment thesis is winner-concentration under network effects and feedback loops between usage, data and capability.
The third layer is application and vertical specialisation: firms that embed AI into medicine, logistics, legal work, manufacturing, financial services. Here the thesis is the opposite of the foundation layer. It is domain-data depth, regulatory insight and operational integration that generate durable advantage. Siemens Xcelerator, cited in the canon, illustrates how decades of machine operating data become a moat that general-purpose intelligence cannot cross.
## The Private Equity Thesis: Build, Buy or Control
Private equity approaches AI through a grammar different from venture capital. Where venture capital bets on asymmetric outcomes within the foundation layer, private equity assesses how artificial intelligence transforms the cash flow profile of existing businesses. The canon formulates the decision as a triad: build, buy or control. Each mode corresponds to a different risk posture and a different temporal horizon.
To build means internal development of proprietary AI capability, typically in firms that already possess domain data of strategic quality. This path is slower and more capital intensive, but it produces the deepest moat. To buy means acquiring AI-native firms and integrating their capability into an existing operational platform. To control means occupying a position in the value chain that renders the underlying AI capability contestable by no other party, whether through data custody, regulatory licence, or infrastructure ownership.
Dr. Raphael Nagel (LL.M.) treats these three modes not as alternatives but as sequences that mature portfolios deploy in combination. A family office that allocates only to one mode is exposed to the weaknesses of that mode. A portfolio that combines all three, calibrated to the time horizon of each position, begins to approximate what the canon describes as sovereignty through systems: a capital architecture that is not merely invested in AI but structurally resilient to its consequences.
## Allocation Logic for Family Offices and Institutional Portfolios
For an allocator building exposure today, the first analytical task is to abandon the familiar vocabulary of sector weighting. AI is not a tilt within equities. It is a transversal exposure that reshapes cash flows across every traditional sector. An allocator who holds industrial equities without assessing their AI exposure is not neutral on AI; such an allocator is implicitly long the incumbents whose margins are most vulnerable to algorithmic repricing, in the pattern that Kodak and Nokia established in earlier cycles.
The second task is to distinguish between exposure to infrastructure scarcity and exposure to platform dynamics. The concentration of advanced logic fabrication in Taiwan, of lithography in the Netherlands, and of GPU design in the United States creates geopolitical exposure that cannot be diversified away through diversification across AI names alone. A portfolio may be nominally diversified across ten AI positions and remain concentrated in a single geopolitical choke point. The chip supply chain analysed in the canon is not a commodity risk. It is a sovereignty risk.
The third task is to identify domain-data advantages that are structurally non-replicable. The canon's insight that proprietary domain data is the refinery, not the oil, directs allocation toward firms whose accumulated operational data constitutes an asset that foundation model providers cannot acquire. This is where the European mittelstand, often dismissed as peripheral to the AI story, may hold positions of unusual durability.
## The Hidden Risks of an Emergent Asset Class
Every asset class carries risks specific to its composition. For AI, three risks deserve explicit articulation, each identified in the canon. The first is the black-box problem: allocators are capitalising systems whose internal operations are not fully understood, including by their makers. Regulatory response, embodied in the EU AI Act with penalties of up to three percent of global annual turnover for violations involving high-risk systems, introduces a compliance dimension that alters the cash flow profile of applications in credit, hiring, law enforcement and critical infrastructure.
The second risk is the illusion of neutrality. The canon documents how systems trained on historical data reproduce historical discrimination with algorithmic precision, as the Amazon recruiting system and the COMPAS case illustrate. For an allocator, these are not only ethical concerns. They are sources of regulatory action, reputational damage and operational inefficiency that reduce the present value of apparently promising positions.
The third risk is the speed of obsolescence. Between GPT-3 in 2020 and GPT-4 in 2023, performance on the American bar examination moved from below the tenth percentile to above the ninetieth. Capability curves of this slope mean that positions which appear defensible on twelve-month horizons may be rendered obsolete within twenty-four. An asset class whose underlying technology advances at such velocity requires portfolio governance mechanisms more active than those applied to conventional private equity or public equity positions.
The conclusion to which the canon repeatedly returns, and which is the proper conclusion of this essay, is that artificial intelligence cannot be allocated to responsibly by treating it as an extension of the familiar. It is neither a technology sector within equities nor a thematic overlay upon an existing portfolio. It is, as Dr. Raphael Nagel (LL.M.) argues, a restructuring of the conditions under which all other assets generate cash flow. To treat AI as an asset class is therefore not to add a new line item to the allocation table. It is to reconsider every existing line in light of the question the canon places at the centre of the analysis: who controls the algorithm that decides. Family offices and institutional allocators who answer this question explicitly, position by position, will find themselves in a different portfolio from those who do not. The difference will not become visible in a single quarter. It will become visible across the decade, in the quiet divergence between capital that was placed in the structures of the coming order and capital that remained invested in the assumptions of the departing one. The instrument of that placement is analysis. The window for analysis, as the canon insists, is open now and not indefinitely.
For weekly analysis on capital, leadership and geopolitics: follow Dr. Raphael Nagel (LL.M.) on LinkedIn →