1.3.1 Winner-Takes-All Dynamics

In 2017, training a state-of-the-art AI model cost about $1,000. You could do it on a decent workstation. Researchers at universities, startups in garages, independent labs—everyone had a shot.

By 2024, training a frontier AI model cost nearly $200 million. That is not a typo. The price increased twenty thousand fold in seven years. If you don't have access to hundreds of millions of dollars, cutting-edge compute infrastructure, and the expertise to orchestrate it all, you're not in the game.

This is how winner-takes-all markets are born. Not through conspiracy or monopolistic scheming, but through the brutal mathematics of scale. Understanding who benefits and who doesn't requires examining the specific mechanisms through which AI creates and sustains competitive advantage—and how those mechanisms have shifted as the technology matured.

The Death of Data Moats

For years, the conventional wisdom was simple: data is the moat. Companies with the most data would build the best AI systems, which would attract more users, generating more data, creating a self-reinforcing cycle. Google, Facebook, Amazon—they sat on treasure troves of user behavior, searches, purchases, clicks. Their data advantage was supposed to be insurmountable.

Then ChatGPT launched. OpenAI didn't have Google's search data. They didn't have Facebook's social graph. They didn't have Amazon's purchase history. What they had was a clever architecture, significant compute, and access to the public internet. GPT-4 could reason about complex problems using training data from across the web. It didn't need proprietary user data—it needed text, billions of words of it, and most of that was publicly available or licensable.

Suddenly, the data moats that companies had spent billions building began to erode. Facebook's internal AI research couldn't keep pace with a startup. Google, with its decades of search data, found itself scrambling to respond. The incumbents' advantage wasn't as durable as everyone had assumed.

Data still matters, particularly for specialized applications. But the value of incremental data has declined while the cost of acquiring truly unique data has risen. Once a foundation model is trained on broad web-scale text, it can be fine-tuned for specific tasks with relatively modest additional data. The character of the data advantage has changed: it now tends to favor those who control genuinely unique, hard-to-replicate datasets—medical records, scientific literature, proprietary transaction data—rather than those with the most raw data volume.

The moat shifted. It's no longer primarily about having data. It's about having compute, distribution, and network effects.

The Compute Oligopoly

The new barrier to entry is infrastructure. Training large AI models requires thousands, sometimes tens of thousands, of high-end graphics processing units running in parallel for weeks or months. Each GPU costs around $40,000. A cluster capable of training a frontier model can represent hundreds of millions of dollars in hardware alone, before accounting for electricity, cooling, or engineering talent.

Who can afford that? Effectively, only the cloud computing giants. Amazon (via AWS), Microsoft (via Azure), and Google (via Google Cloud) control approximately two-thirds of the global cloud computing market. If you want to train a large AI model and don't already own a massive data center, you're renting capacity from one of them.

This creates a structural asymmetry. Large tech firms can acquire GPUs by the tens of thousands, build custom AI chips, and operate at scales where marginal capacity costs are far lower than for any smaller competitor. A startup trying to compete must either raise hundreds of millions—or billions—in venture capital, or accept dependence on cloud infrastructure controlled by potential rivals. The effective threshold of competitive scale has been described as the capacity to deploy around a million graphics processing units: those who can train bigger, cheaper, and faster than competitors set the pace, while those who cannot are followers.

This compute oligopoly shapes the broader competitive landscape in ways that extend beyond the cost of any single training run. Control over infrastructure translates into pricing power over every company that relies on that infrastructure, visibility into competitor usage patterns, and the ability to prioritize capacity for internal AI projects when demand is high.

The DeepSeek Surprise

Just when the narrative around compute costs seemed settled, something unexpected challenged it. In early 2025, the Chinese lab DeepSeek announced it had trained an advanced AI model using significantly fewer high-end chips than competing systems, relying instead on architectural innovations and optimized training techniques. The result performed comparably to much more expensive models on several benchmarks.

It wasn't magic—it was engineering efficiency. Rather than scaling compute to meet the problem, DeepSeek redesigned the problem to require less compute: better algorithmic choices, novel training strategies, and careful optimization produced a model that closed much of the gap at far lower resource cost.

The announcement briefly rattled investors who had bet billions on the assumption that AI progress required ever-larger capital expenditures. If a well-designed model could compete at a fraction of the cost, the capital barrier might be less permanent than assumed.

But context matters. DeepSeek's breakthrough required world-class research talent, novel and untested architectural ideas, and considerable institutional willingness to bet on an unproven approach. For every success like this, many well-funded teams attempt similar efficiency-first strategies and fail. Furthermore, "relatively few chips" still represents an enormous investment by the standards of most organizations—the threshold dropped, but it did not disappear. DeepSeek demonstrated that the frontier can occasionally be disrupted by efficiency rather than raw scale, but it has not fundamentally altered the structural advantages held by those commanding the most compute.

Economies of Scale Everywhere

What makes AI markets particularly prone to concentration is that economies of scale operate at every layer simultaneously, not just at the level of model training. Training is cheaper per parameter when you train a larger model. Inference—running the model to generate outputs—is cheaper per query when you're serving millions of users rather than thousands, because fixed infrastructure costs are spread across more requests. Data collection and curation is more efficient when you already have a platform with billions of active users generating behavioral signals at scale.

These individual advantages compound into a flywheel dynamic. A large platform deploys AI to improve its services, attracting more users. More users generate more data, enabling better models. Better models justify larger infrastructure investments, which lower marginal costs, which fund additional model improvements. Each turn of the flywheel widens the gap between the largest players and everyone else.

Research has documented this effect empirically. A study examining firm-level AI adoption found that the largest third of companies increased sales by 17% after deploying AI, while the smallest third saw no measurable increase. The technology is nominally available to all, but returns concentrate heavily at the top of the firm-size distribution—consistent with what economists call superstar firm dynamics, where small efficiency advantages, amplified by scale, produce disproportionate market outcomes.

The mechanism is not mysterious. When AI improves core business operations—recommendations, logistics, advertising targeting—the firm that can invest the most in AI improvement captures a disproportionate share of productivity gains, because those gains compound across its scale of operations. A one-percent improvement in ad targeting efficiency at a company running a hundred billion ad impressions annually is worth vastly more than the same improvement at a company running a billion, even if both firms start with an identical model.

The Distribution Game

If data moats have weakened and compute moats are expensive to overcome, distribution has emerged as a third, equally powerful form of advantage. Getting an AI solution widely adopted—through platform integrations, established user relationships, or simply being the first to solve a pressing need—creates self-reinforcing advantages through habit formation, switching costs, and ecosystem lock-in.

Microsoft's partnership with OpenAI illustrates the dynamic clearly. OpenAI built the technology; Microsoft provided the distribution: Windows, Office, Azure, LinkedIn, Xbox—a set of platforms reaching over a billion users. Together, they captured roughly 69% of the generative AI market as of late 2023. OpenAI's models powered both its own consumer offerings and Microsoft's enterprise products, creating a combined position that would have been difficult for either company to achieve independently.

Other incumbents hold comparable distribution advantages. Google integrates AI capabilities into Search, Gmail, Android, and Chrome. Meta deploys AI across Facebook, Instagram, and WhatsApp. Amazon embeds intelligence into AWS and its commerce infrastructure. These platforms are not merely large—they are structural fixtures of daily life for hundreds of millions of people. AI features delivered through these channels reach users without any deliberate adoption decision on their part, which is a form of market penetration no startup can replicate through product quality alone.

The resulting asymmetry is stark. A company with a technically superior model must spend heavily on marketing, sales, and distribution to acquire users one at a time. An incumbent with an adequate model and massive existing distribution deploys it to a captive audience and wins much of the market by default. Switching costs compound the effect: once users integrate an AI capability into their workflows—autocomplete in their email client, AI assistance in their office suite, recommendations in their shopping interface—the bar for switching to an alternative is high, regardless of whether the alternative is better.

Network Effects: The Enduring Moat

Among all the competitive mechanisms shaping AI markets, network effects are arguably the most durable and the least subject to disruption by cost reductions or technical breakthroughs.

A network effect exists when a product or service becomes more valuable as more people use it. Social networks exhibit this property; so do marketplaces, operating platforms, and developer ecosystems. AI systems generate network effects through several distinct pathways. More users produce more behavioral data, improving model performance and attracting additional users. More developers building on a platform create more integrations and capabilities, increasing the platform's overall value. More third-party applications in an ecosystem raise switching costs, because users moving to a competitor would lose access to the tools they depend on.

AI agents—autonomous systems that perform multi-step tasks on behalf of users—may intensify these dynamics further. When agents can interact with and delegate to other agents, the value of any individual agent increases with the number of other agents it can collaborate with. A rich ecosystem of interoperable agents creates emergent value that isolated, technically superior agents cannot replicate. This architectural property means that the first platforms to establish large agent ecosystems could develop advantages that become progressively more difficult to displace, regardless of later entrants' technical capabilities.

The combined effect of network effects, distribution scale, and compute infrastructure creates what some researchers describe as winner-takes-most dynamics: not necessarily a single monopolist in every AI market, but a small number of large players capturing a disproportionate share of value in each segment, with a long tail of smaller firms competing for the remainder.

The table below summarizes how these four competitive mechanisms compare in character, trajectory, and who primarily benefits.

The Counterarguments

Not everyone accepts that AI markets are inexorably concentrating, and the counterarguments deserve serious consideration.

New training techniques and the broader availability of cloud capacity have already enabled well-resourced startups to train competitive large models, or to fine-tune open-weight models for specific tasks with modest resources. The growing emphasis on smaller, more specialized models—designed for narrow domains or optimized for efficient deployment—means that not every application requires competing at the frontier. In many cases, a well-designed specialist model can match a general-purpose frontier model for a particular task at a fraction of the cost, which opens space for focused competitors.

The open-source movement presents a more structural challenge to concentration. Meta has released the LLaMA family of models; Mistral, EleutherAI, and others have contributed powerful open-weight models that any developer can use, modify, and deploy without licensing fees. If foundation models become effectively commoditized—widely available and roughly equivalent in core capability—competition could shift toward applications, user experience, and vertical services, where incumbency advantages may be less durable and barriers to entry lower.

Compute costs themselves have fallen by several orders of magnitude over the past decade. If that trajectory continues, the capital required to train competitive models could erode substantially. And competition within each layer of the AI stack remains genuine: cloud providers engage in active price competition, model developers race to release improved versions, and application companies fight intensely for user attention. This is not a settled oligopoly in the mold of legacy utilities—it is a fast-moving competitive environment, and the leaders of today may not be the leaders of tomorrow.

Market Structure and Concentration

The counterarguments are real, but the empirical data on market outcomes points in a consistent direction: concentration is increasing. A small number of companies controls the cloud infrastructure on which AI development depends. A comparably small set trained the most capable frontier models. A handful of companies captured the majority of generative AI market share. The largest firms are growing faster than small ones, and startups routinely face thin margins, high infrastructure costs, and structural dependence on potential competitors.

This concentration is not inherently pathological. Economies of scale produce genuine benefits: lower prices, faster innovation cycles, and the capacity to fund research that fragmented markets could not support. The rapid progression of foundation models has been driven largely by a small number of well-capitalized competitors, and that competition has accelerated the entire field.

However, the distributional implications deserve scrutiny. When a small number of firms controls critical AI infrastructure and captures most of the financial returns from AI adoption, productivity gains do not spread evenly across the economy. The advantages of scale—compute, data, distribution, and network effects—compound over time and are difficult for challengers to overcome even when those challengers are technically sophisticated. The structural dynamics favor incumbents in ways that persist through ordinary market competition and may require deliberate policy intervention to address.

How this resolves over the coming decade depends partly on technological trajectories—whether compute costs continue falling, whether open-source models continue improving, whether new architectural approaches lower the capital requirements for frontier development—and partly on regulatory choices about data access, platform interoperability, and infrastructure sharing. Neither the concentration scenario nor the open-competition scenario is inevitable. Both require active choices, by companies, governments, and researchers, to realize or to prevent.

Summary

Winner-takes-all dynamics in AI arise from the interaction of several compounding advantages rather than from any single cause. The four primary competitive mechanisms—data, compute, distribution, and network effects—have shifted in relative importance as the technology matured, with data moats weakening as general-purpose models trained on public text became competitive, and compute scale, distribution, and network effects becoming more central.

Compute infrastructure presents the highest direct barrier: training frontier models requires capital and hardware access concentrated in a small number of cloud providers. Architectural efficiency gains, as illustrated by DeepSeek, show this barrier is not absolute, but do not eliminate it. Distribution advantages—derived from existing platforms reaching hundreds of millions of users—give incumbent technology companies a structural head start in AI deployment that startups cannot easily overcome through product quality alone. Network effects, the most durable of the four mechanisms, compound over time and are the hardest to displace through technical superiority.

Economies of scale operate across all of these dimensions simultaneously, producing a flywheel in which existing size advantages generate the resources to extend those advantages further. The empirical result is measurable concentration: large firms growing faster than small ones, a few cloud providers controlling critical infrastructure, and a small number of companies capturing most of the financial value generated by AI.

Open-source models, falling compute costs, and the viability of specialized small models provide meaningful counterweights, and the competitive landscape remains dynamic. But the structural forces favor concentration, and understanding these dynamics—who controls the key inputs to AI development, and why those positions are difficult to displace—is foundational to assessing AI's broader economic effects.

Key Takeaways

• The 20,000x increase in frontier model training costs since 2017 has made compute infrastructure — not data — the primary competitive moat, concentrating AI development among cloud giants and their closest partners.
• Data moats have weakened for general-purpose AI (models trained on public web text can compete without proprietary data), but durable advantages remain for holders of genuinely unique, hard-to-replicate specialized datasets.
• Distribution has become as powerful a moat as compute: incumbents like Microsoft, Google, and Meta reach billions of users through existing platforms, allowing adequate AI features to beat technically superior products simply through scale of reach.
• Network effects — where product value grows with users, developers, and integrations — are the most durable competitive advantage and the hardest to displace through technical breakthroughs or cost reductions alone.
• Economies of scale compound simultaneously across all dimensions, creating a flywheel: scale → lower marginal costs → more investment → greater scale. Research shows the largest firms see 17% sales gains from AI adoption; the smallest see none.
• DeepSeek demonstrated that algorithmic efficiency can occasionally lower the barrier to frontier AI, but has not fundamentally altered the structural advantages of the largest compute operators.
• The empirical data on market outcomes consistently points toward concentration: large firms growing faster than small ones, a few cloud providers controlling critical infrastructure, and a small number of companies capturing most of the financial value generated by AI.

Last updated: 2026-02-25