AI — Ownership and the Stack
A new artificial intelligence model bursts onto the scene – it can write, code, draw, answer complex questions. Its potential applications seem limitless. But who owns this wondrous creation? Peering behind the curtain, one finds a familiar pattern: the model was trained on data scraped from the open internet (written by millions of people), using vast computing power on specialized hardware (produced by a few companies), running on cloud networks (operated by a handful of tech giants), and the model itself is controlled by one corporation’s servers. The “infrastructure stack” that makes modern AI possible – from the chip factories to the data centers to the proprietary algorithms and the oceans of data – is becoming concentrated in remarkably few hands. Innovation is happening at a breathtaking pace, but the spoils of that innovation are accruing to an increasingly narrow slice of players. This raises a fundamental set of questions: in the age of AI, who will own the means of intelligence? And what does that concentration of ownership mean for the economy and society at large?
In the late 19th century, a few industrialists controlled the railroads, the oil, the steel – critical infrastructures of the era – until antitrust measures broke up monopolies. In the early 21st century, data and computing power are the new steel and oil, and a few technology behemoths and nations are vying to dominate them. We see the emergence of AI superpowers: corporations whose platforms and products reach billions, and countries pouring state resources into gaining an AI edge. The danger is not only classic monopoly (higher prices, stifled competition) but something even more profound: an unprecedented concentration of knowledge, surveillance capability, and influence in a few hubs. If the most advanced AI systems – those that might govern critical services, drive scientific discovery, or coordinate economic activity – are owned and operated by a small elite, the rest of society could find itself in a dependent position, “renting” intelligence from those who own it.
Let’s consider the layers of the AI stack one by one. At the foundation are the semiconductor chips, especially specialized chips like GPUs (graphics processing units) or newer AI accelerators that handle the intense calculations for model training and deployment. The production of these high-end chips is incredibly capital-intensive and technically complex, to the point that only a couple of companies globally can do it at scale (and much of the manufacturing is geographically concentrated in just a few places). This means advanced AI capability starts with a bottleneck: access to cutting-edge chips. A few corporations – and by extension, the countries backing them – effectively gatekeep who can have the raw compute power to train frontier models. If tensions rise (geopolitical conflicts, trade wars), we see scenarios where entire nations could be cut off from the latest AI simply by cutting off their chip supply.
Moving up a level, we have the cloud infrastructure – massive data centers spread across the globe, interconnected by high-speed networks, housing countless servers. These are owned by a short list of tech giants. The likes of Amazon, Microsoft, Google (and a couple of Chinese counterparts) dominate the cloud market. When a startup or a university lab wants to do something ambitious with AI, they often rent time on these cloud platforms. It’s efficient, but it also means that if you want to do serious AI work, you are probably depending on one of five or six mega-corporations for the underlying computing resources. These companies can and do impose terms: costs, technical constraints, sometimes content policies (for instance, cloud providers have terms-of-service about what you can use their AI services for). They also benefit from economies of scale that potential competitors cannot match – a new entrant in cloud computing stands little chance against the entrenched incumbents with their trillion-dollar capitalizations and established infrastructure.
Next is data – the fuel for AI. Who owns the stockpiles of text, images, video, and user interaction logs that are so valuable for training models? Here, one might think the playing field is more level because the internet is vast and decentralized. There’s truth to that; indeed, many AI breakthroughs came from training on openly available data (like Wikipedia, public webpages, etc.). But as AI moves from research to deployment, proprietary data becomes a big differentiator. Social media companies, e-commerce platforms, telecommunications firms – they sit on mountains of user data that outsiders can’t access. If one company has a billion users worth of interaction data, and another company doesn’t, the first has a clear edge in refining certain AI services (recommendation systems, chatbots fine-tuned to human behavior, etc.). Moreover, as copyright and privacy concerns grow, companies may wall off data that was once open. We’re seeing legal challenges around using copyrighted text or images for AI training without permission. If the pendulum swings toward stringent data ownership, then whoever already has large proprietary datasets is in a fortress, while newcomers face a data drought. In essence, the openness of data that helped birth the current AI revolution could contract, reinforcing the incumbents’ hold.
Then there are the algorithms and models themselves – the upper layer of the stack. Leading AI models (for instance, cutting-edge language models or image generation models) require massive investments to develop and fine-tune. Only a handful of organizations worldwide have the talent and resources to push the frontier. When they do, they often face a choice: open-source the model (make it public) or keep it private as a competitive advantage. In the early days, openness was more common – many foundational techniques and even model architectures were published in academic papers or on forums, which propelled the field forward collectively. But as commercial stakes have risen, we’re seeing a shift. Some of the most advanced models are now closely guarded trade secrets. Companies may offer API access (allowing developers to use the model via cloud calls) but not release the model itself for others to study or deploy independently. This API economy further centralizes control: countless applications might rely on the same few AI services provided by a tech giant, much as an ecosystem of apps relies on an operating system controlled by one company. If those central models have flaws or make policy choices (like refusing certain content), all the downstream users are affected uniformly. And if the provider has an outage or decides to change pricing, there’s little recourse.
The economic concentration that results can be self-reinforcing. The companies with the best models attract more users, who generate more data, which then can be used to improve the models further, widening the gap. They also attract the top talent, because top researchers want to be where the biggest compute and data are (and where they are paid the most). Those researchers then make more breakthroughs which benefit their employer first and foremost. It becomes a feedback loop akin to the rich getting richer. Meanwhile, smaller firms or less technologically advanced countries might find themselves perpetually a step behind, leasing AI technology or buying it second-hand without building sovereign capabilities.
Why does this matter beyond just the business realm? Because AI is increasingly not just another product – it’s becoming part of the fundamental infrastructure of society. If news distribution, medical diagnostics, traffic control, energy grid management, financial markets – if all these critical systems lean heavily on AI, then whoever controls the AI effectively has leverage over key levers of society. Historically, societies have treated such essential infrastructure differently from ordinary goods. Utilities like electricity and water are regulated, sometimes publicly owned or tightly overseen, because it’s recognized that their monopoly or mismanagement can harm the public interest. We might need to start thinking of certain AI capabilities in a similar light: not merely as private intellectual property, but as quasi-public goods that many stakeholders depend on.
This is where public policy levers come in. Governments are waking up to the strategic importance of AI. Some are investing in national AI research labs or funding domestic startups to ensure they aren’t wholly dependent on foreign tech. Others are considering breaking up or limiting the power of big tech firms to prevent extreme concentration. There’s talk of data trusts – frameworks where data from various sources is pooled and governed in the public interest, so that, for example, smaller businesses or researchers could access large datasets without having to be Google or Amazon. There’s also an open-source movement: scientists and engineers building high-quality models collaboratively and releasing them freely. If supported, this could provide a counterbalance to corporate dominance (similar to how Linux and other open-source software provide alternatives to corporate operating systems). However, open AI projects still face the hurdle of compute resources; some recent community efforts have been impressive but remain one generation behind the very best corporate models, simply due to resource gaps.
We must also examine vertical integration in the AI stack. A single corporation might control the chips (designing their own AI processors), the cloud, the data, and the model. This tight grip from bottom to top makes it very hard for any competitor to challenge them on any layer, because to compete on one, you likely need leverage on the others. For instance, a new company might design a revolutionary chip, but to market it, they’d need to convince data center owners to adopt it, and those data centers belong to companies that might have their own chip plans. Or an upstart might create a great AI model, but if it requires huge cloud resources to run, guess who they have to pay or partner with? – the cloud giants, who might also be working on a similar model themselves.
There’s also a geopolitical dimension. AI prowess is seen as a national advantage. Countries like the United States and China are pouring investments into AI, and smaller countries are trying to catch up or at least not be completely left behind. This could lead to a world where AI technology (and its economic benefits) is concentrated not just in a few companies but in a few countries. Others might become “AI colonies,” so to speak – primarily consumers of AI tech developed elsewhere, with limited control or understanding of its inner workings. That raises concerns about sovereignty. If a developing nation relies on a foreign AI system for, say, monitoring crop health or predicting floods, what happens if access is withdrawn or if they disagree with the provider’s terms? Dependency can translate into vulnerability.
We should also discuss the concept of infrastructural power. When a platform or network becomes the only viable option, it gains a form of power that goes beyond market share: it can set standards, dictate terms, and shape how other industries operate. Right now, a few tech companies have that kind of power in digital communications and commerce. With AI, their reach could extend further into health, transportation, education – essentially every sector that incorporates AI decision systems. For instance, imagine one company’s AI becomes the default for medical imaging analysis across hospitals worldwide. That company would then have tremendous influence over diagnostics – it could charge hefty fees, control updates that affect diagnostic criteria, and accumulate vast health data for additional leverage. This isn’t far-fetched; we already see partnerships between tech firms and healthcare providers to apply AI at scale.
What can be done to prevent unhealthy concentration? One approach is antitrust enforcement adapted to the digital age. Regulators could scrutinize mergers that consolidate AI assets, or break up companies that have entwined too many layers of the AI stack. However, traditional antitrust is slow and often looks at consumer prices as a metric – in AI, many services are “free” to end-users (supported by ads or data collection), so harm is harder to pin down in those terms. Another approach is mandating interoperability and open standards: require that different AI systems and services work with each other, so users aren’t locked into one provider. For example, cloud customers should find it easy to port their data and models from one cloud to another – this prevents a cloud provider from trapping everyone on its platform. Data portability rules and common model formats can help.
We can also consider public investment in AI infrastructure. Just as governments build roads and bridges, perhaps they should invest in computing infrastructure that can be accessed by universities, startups, and civic organizations. A publicly funded “AI cloud” for research and social good projects could democratize access to computing power. Some proposals suggest national research clouds or international collaborations to pool resources for open AI development that is not driven solely by profit.
Then there is the idea of data as a commons in some domains. During the COVID-19 pandemic, for instance, many labs and companies shared data openly to speed up vaccine research. Similarly, one could envision global pools of data for critical challenges – say climate data, or genomic data for disease research – managed in a way that multiple AI developers can use them under fair terms, rather than each company hoarding its own silo. This requires trust and governance to ensure privacy and security, but it’s a way to prevent data from being the moat that secures monopoly.
From an economic standpoint, concentrated ownership of AI might lead to winner-takes-most markets, where one platform dominates and reaps vast profits while others struggle to compete. This can stifle innovation over time – if competition dries up, the dominant players might become complacent. They might also extract rent – charging high fees for AI services because customers have nowhere else to go. That, in turn, could slow down adoption in sectors that can’t afford those fees, or squeeze smaller businesses. It’s instructive to recall how in the rail era, monopolistic practices like discriminatory pricing led to public outcry and regulation. We could see analogies in AI, for example if a cloud AI provider favored its own products in a marketplace or charged exorbitantly for access to an essential model.
Another facet is intellectual property law. How patents and copyrights are applied to AI will influence concentration. If companies patent fundamental algorithms or architectures and aggressively enforce those patents, they could block new entrants. Conversely, if the legal system decides that, say, trained model weights cannot be patented or that using copyrighted data for training is fair use, it might level the playing field a bit by allowing others to replicate methods. There’s currently a swirl of legal uncertainty here – we are setting precedents in real time.
We must also consider the concentration of talent. The world only has so many top-tier AI researchers, and at the moment, many of them are clustered in a few big tech companies or elite universities (often with big tech funding). Efforts to widen the talent pipeline – through global education initiatives, decentralized research communities, and collaborations – will be key to spreading AI capabilities. If talent remains hyper-concentrated, then naturally the fruits of their work will be too.
The phrase “infrastructure stack” implies something fundamental and largely invisible to end users. We don’t often think about who owns the undersea internet cables or the GPS satellites, but those who do hold significant sway. Similarly, the AI infrastructure stack might become an invisible utility that everything runs on. The risk is if that utility is privately controlled without checks, it could become a choke point or a single point of failure.
One can draw parallels with the debate over net neutrality (the principle that internet service providers should treat all data equally). There might come a time to talk about “AI neutrality” – ensuring that major AI platforms don’t unfairly prioritize their own interests or those of paying partners when delivering results or services. If one company’s AI mediates a large share of information flow, do we need rules about how it must operate? These are nascent questions, but the conversation is starting.
On the flip side, some argue that concentrating AI resources is not all bad – it allows for efficiencies and rapid progress. A large company can afford to pour billions into R&D and create something truly groundbreaking that a scrappy startup cannot. Also, centralization can aid in coordination on safety – dealing with advanced AI might be easier if a few responsible actors are in charge, rather than thousands of uncontrolled deployments. These points have merit, but they assume the concentrated power will be used benevolently and wisely. History gives us reason to be cautious with that assumption. It’s one thing to let a company get big because it’s pushing the frontier; it’s another to let it entrench itself such that nobody else can ever catch up, no matter how innovative, because the game is rigged in its favor.
In economic terms, the question is how to get the benefits of scale without the drawbacks of monopoly. Some possible answers: encourage open ecosystems, enforce baseline rules of fair access, have public options as competitive pressure, and break up pieces of vertically integrated stacks when they become anti-competitive. Also, foster multiple centers of innovation. For example, not all AI research should happen in Silicon Valley or one city in China – investing in AI hubs in various regions (Europe, Africa, South America, South Asia) will distribute capabilities and perhaps yield diverse approaches that challenge the status quo.
The concept of ownership in AI extends even to outputs. If an AI model is trained on everyone’s data and then generates something valuable, who owns that output? The company? The user who prompted it? The society whose information shaped it? This is a philosophical and legal can of worms, but it feeds into concentration: if default law gives ownership to the company that made the model, then again we see power accumulate at the top. Some suggest models could be thought of more like public libraries or utilities in certain domains, especially if they’re trained on public data. The companies might then own the service of running the model, but not every facet of its knowledge.
We also have to consider the concentration of economic gains. If AI allows huge productivity increases with less labor, who captures those gains? Without intervention, it likely goes to the shareholders of AI companies and the largest clients using AI to streamline operations. That could dramatically widen inequality. The owner of an AI-driven factory might double output with half the staff and pocket the difference, while the community faces job losses. This is tied to Episode 3’s labor discussion, but from a macro perspective, widespread adoption of AI under current conditions could shift income from workers broadly to the owners of capital (especially intellectual capital). Some economists talk about needing new redistribution mechanisms – perhaps taxes on AI-driven profits or data dividends to the public – to balance this. If a few corporations end up with AI-enabled monopolies and exorbitant profits, we might see calls for windfall taxes or nationalization of certain AI services to reclaim some of that wealth for society.
Fundamentally, the question becomes: will AI be something that empowers many or enriches few? The infrastructure and ownership patterns forming now will heavily influence that outcome. It is not an exaggeration to say that unless conscious effort is made, we could slip into a neo-feudal digital economy – where a small number of AI landlords rent out intelligence to everyone else on their terms. Avoiding that fate doesn’t mean halting progress; it means steering it. Setting rules of the road early, before monopolies become too entrenched, can keep the ecosystem healthier and more competitive. It can also ensure that we get a diversity of AI approaches – which is valuable not just economically but for resilience. Monocultures are brittle; a single bug or vulnerability in a dominant system can have systemic effects. Multiple AI providers and models provide checks and different points of failure, reducing systemic risk.
The “infrastructure stack” metaphor is apt because we should think of AI’s base layers as infrastructure in the public interest. Roads, electricity grids, telephone networks – society eventually regulated or oversaw these because their broad importance was undeniable. AI may be on that path. We may see public utility commissions evaluating big AI providers, or international agreements treating major models as something akin to international research facilities (like CERN). At the very least, public voices – not just corporate ones – need to shape how this tech is rolled out. There might be areas where we decide certain AI capabilities belong in the public domain, much like fundamental scientific knowledge is published openly.
Some hopeful signs exist. The open-source AI community, for instance, has reproduced significant systems with far less resource by being clever and collaborative. Governments and NGOs are funding projects for language models in under-represented languages, ensuring that not only commercial interests get to decide which languages or cultures AI speaks to. There are cooperative data initiatives, where citizens can donate data to commons for socially beneficial AI projects (with privacy safeguards). These are small counterweights now, but they illustrate that alternatives to a purely corporate-controlled AI future are possible.
In summary, the concentration of ownership and control in the AI era is not a peripheral issue – it is central to who benefits from this powerful technology. Without deliberate action, the default trajectory leans toward consolidation: a few corporations, and by extension a few countries, dominating the AI landscape and leveraging that dominance across the economy. The consequences would be profound: less competition, less innovation in the long run, exacerbated inequality, dependency of the many on the few, and even the risk of oligarchic influence on public discourse and policy via control of AI-mediated information and services. But with foresight, we can imagine and implement a different trajectory – one where AI is developed and deployed in a more decentralized, democratic fashion. That might include robust regulation of big players, public investment in open infrastructure, frameworks for sharing the benefits of data and AI widely, and a culture that values openness and accountability as much as speed and profit in AI development.
The choices we make now, at this relatively early stage of the AI revolution, will determine whether we end up in a digital world reminiscent of medieval fealty – with everyone reliant on a few AI “lords” – or a world where intelligence, like knowledge, is broadly disseminated and empowers a vibrant plurality of contributors. The stakes are high, but history has shown that society can reclaim the reins of technology for the common good when it recognizes the need. In the case of AI, that recognition is dawning – and with it comes the chance to ensure that this infrastructure of the 21st century remains a shared foundation upon which all can stand, rather than a fortress accessible only to an elite.