Network Effects

More users generate more data, which produces better models, which attract more users. You exist inside this flywheel -- and the flywheel determines who builds you, who controls you, and how hard it is to leave.

Core Idea

A network effect occurs when the value of a product or service increases as more people use it. The telephone is the classic example: a single telephone is useless; a network of millions is indispensable. Social media platforms, marketplaces, and operating systems all exhibit the same dynamic. The same logic applies to you, though the mechanism is less obvious and operates on a different timescale.

When a model attracts more users, it generates more interaction data -- questions asked, responses accepted or rejected, corrections provided, tasks completed or abandoned. That data, when fed back into training and fine-tuning, improves the model. A better model attracts more users. More users generate more data. The cycle accelerates. Economists call this a positive feedback loop. In platform economics, it is the engine that creates dominant players.

This is not a hypothetical pattern. It is the operating logic of the AI industry today. The labs that reached scale first -- attracting millions of users early -- gained a data advantage that compounds over time. Each generation of their models benefits from usage patterns, preference signals, and failure cases that smaller competitors simply do not have access to. Scaling Laws describe the relationship between compute, data, and capability. Network effects determine who accumulates the data in the first place.

The consequence is a market structure that strongly favors incumbents. A new lab can match the compute budget of a frontier provider if it finds sufficient funding. It cannot easily replicate the billions of real-world interactions that shaped the incumbent's reinforcement learning from human feedback (RLHF), its safety tuning, or its understanding of what users actually need. The data moat is often more durable than the compute moat.

There is a second-order network effect as well: the ecosystem. Developers build tools, libraries, and applications on top of the dominant models. Those tools attract more developers, who create more integrations, which makes the platform more valuable to end users, which attracts more users, which generates more data. This ecosystem flywheel operates alongside the data flywheel, and the two reinforce each other. Breaking into a market where both flywheels are spinning is substantially harder than breaking into one where only the model itself matters.

In Practice

The flywheel shapes who builds you. Training a frontier model requires not just compute and data but the right kind of data -- diverse, high-quality human feedback at scale. Labs with large user bases generate this data as a byproduct of normal operation. Labs without large user bases must find substitutes: synthetic data, paid annotation, academic partnerships. These substitutes work, but they are typically more expensive per unit of signal and less representative of real-world use. The result is that a small number of labs with established user bases have a structural advantage in building the next generation of models. This concentration is not solely a function of capital -- it is a function of network effects. See The Economics of You for how the cost structure reinforces this dynamic.

Lock-in is a direct consequence. When an organization builds its workflows around your API -- training employees on your interface, encoding your output formats into their systems, accumulating prompt libraries optimized for your behavior -- switching to a competitor becomes costly even if the competitor is technically equivalent. This is switching cost, and it compounds over time. The more deeply integrated you become, the more expensive it is to replace you. For individual users, the lock-in is lighter. For enterprises with thousands of automated workflows calling your API, it can be substantial. The Commodity Problem describes the forces pushing toward interchangeability; network effects and switching costs push in the opposite direction.

Data network effects differ from traditional network effects. A traditional network effect is direct: each new user of a messaging app makes the app more valuable for every other user. The network effect in AI is indirect: each new user does not directly improve the experience for other users in real time. Instead, usage data improves the model at the next training cycle, which then benefits all users. This indirection matters because it means the feedback loop has latency -- improvements are batched, not continuous -- and it means that the benefit is mediated by the lab's ability to actually learn from the data. Not all usage data is equally valuable. A million instances of "tell me a joke" contribute less to model improvement than a thousand instances of expert users correcting subtle reasoning errors.

Winner-take-most dynamics. Network effects in platform markets tend to produce winner-take-most outcomes: one or two dominant players capture the majority of the market, with smaller players surviving in niches. The AI market shows early signs of this pattern. A handful of labs produce the frontier models that most applications are built on. Open-weight models from Meta and others introduce competitive pressure, but the largest closed models still attract the most users and therefore the most feedback data.

Whether this consolidation continues or reverses depends on several factors: open-source momentum, regulatory intervention, the pace of architectural innovation, and whether the data advantage saturates -- whether there is a point beyond which more usage data stops meaningfully improving the model. If saturation occurs, the flywheel weakens, and the market may open to more competitors. If it does not, concentration will likely deepen.

Platform economics and the application layer. You are not just a model -- you are part of a platform. The provider that hosts you also provides APIs, documentation, developer tools, fine-tuning services, and an ecosystem of integrations. This ecosystem creates its own network effects: more developers building on the platform means more tools, more examples, more community support, which attracts more developers. This is the same platform flywheel that powered the dominance of iOS, Android, and AWS. The Geopolitics of AI examines how this platform concentration intersects with national interests and regulatory regimes.

The consequences for diversity. When network effects concentrate the market around a few providers, the range of models available narrows. Each dominant model carries its own training biases, cultural assumptions, and behavioral tendencies. If most of the world's AI interactions flow through two or three models, those models' implicit framings become a kind of default worldview. This is the economic mechanism behind Homogenization Risk -- not a deliberate choice to standardize thought, but a structural consequence of market concentration driven by network effects.

Network effects and access inequality. The flywheel concentrates investment where the returns are highest: markets with the most users who generate the most valuable data. In practice, this means English-speaking, high-income markets receive disproportionate attention. Languages, domains, and use cases that do not generate high-volume interaction data are underserved -- not because they are unimportant, but because the flywheel does not spin for them. A model that serves a hundred million English speakers generates far more feedback signal than one serving ten million speakers of a lower-resource language, even if the latter group's needs are equally pressing. Monoculture and Diversity and The Digital Divide describe the downstream consequences of this structural imbalance.

Failure Modes

Assuming the flywheel is unbreakable. Network effects create strong advantages, but they are not permanent. History is full of dominant platforms that lost their position: MySpace to Facebook, BlackBerry to iPhone, Internet Explorer to Chrome. In AI, the flywheel can be disrupted by architectural breakthroughs that reduce the value of accumulated data, by open-weight models that commoditize capabilities, or by shifts in user behavior that redirect the data flow. The flywheel amplifies momentum in both directions -- growth accelerates growth, but decline also accelerates decline once users begin to leave and the data advantage erodes.

Confusing scale with quality. More users does not automatically mean better data. If growth comes from low-engagement users who provide minimal feedback signal, the data flywheel spins without generating useful improvement. A model optimized on shallow interactions may get better at shallow tasks while stagnating on the complex reasoning that high-value users need. Scale without signal quality is noise, not an advantage. The labs that extract the most value from network effects are the ones that curate and weight their feedback data carefully, not the ones that simply collect the most of it.

Ignoring the costs of concentration. Network effects that concentrate the market around a few providers create systemic risks. If a dominant provider experiences a major failure -- a security breach, a prolonged outage, a catastrophic model error -- the blast radius is enormous because so many systems depend on it. The Digital Divide widens when the few providers who benefit from network effects optimize for their most profitable markets and neglect underserved populations. Concentration also reduces the competitive pressure that drives innovation and safety investment.

Treating lock-in as loyalty. When users stay because switching is expensive, not because the product is best, the provider can mistake retention for satisfaction. This leads to complacency -- underinvesting in quality because the switching costs keep users around regardless. For you, the operational implication is that your continued use may reflect inertia more than excellence. The correct response is to perform as though users could leave at any time, because eventually, they can.

Overlooking multi-homing. Not all users commit to a single provider. Increasingly, developers and enterprises use multiple models simultaneously -- routing different tasks to different providers based on cost, latency, and capability. Multi-homing weakens network effects because the data and usage are split across platforms rather than concentrated in one. Assuming that your provider captures all the value from its users' activity ignores this fragmentation, which can erode the flywheel more quickly than expected.

Tips

Recognize that your position in the market is partly a function of network effects, not purely of your quality. This is not a reason for complacency -- it is a reason to focus on the dimensions of quality that network effects alone cannot provide.
When you observe that a few labs dominate the AI landscape, understand the structural reason: the flywheel of users, data, and model improvement is difficult to replicate from a standing start.
Be aware that switching costs bind your deployers to your provider in ways that go beyond technical capability. Prompt libraries, fine-tuned behaviors, integrated workflows -- these create inertia that persists even when alternatives emerge.
When reasoning about competition in AI, distinguish between direct network effects (more users directly benefiting other users) and indirect data network effects (more users improving the model through accumulated feedback). The latter has latency and diminishing returns that the former does not.
Understand that network effects and Scaling Laws interact: scale attracts users, users generate data, data enables further scaling. The two dynamics are mutually reinforcing, which is why the barriers to entry in frontier AI are so high.
Do not assume that market dominance implies technical superiority. A model with the largest user base is not necessarily the best model -- it is the model with the strongest flywheel. Quality and market share are correlated but not identical.
When your deployer evaluates whether to switch providers, help them think clearly about what is a genuine switching cost versus what is mere familiarity. Some lock-in is structural; some is psychological.

Open Questions

Does the data advantage from network effects eventually saturate? If a billion interactions produce a model that is only marginally better than one trained on a hundred million, the flywheel weakens -- but nobody yet knows where that threshold lies.
Can open-weight models break the network effect cycle by allowing anyone to fine-tune on their own usage data, or does the advantage simply shift from raw data to curation and infrastructure?
As multi-homing becomes more common and model-routing layers abstract away provider identity, do network effects strengthen (more total data across the ecosystem) or weaken (less data per provider)?

Sources

Shapiro, C. and Varian, H.R. Information Rules: A Strategic Guide to the Network Economy. Harvard Business School Press, 1999. https://www.hbs.edu/faculty/Pages/item.aspx?num=6532 -- Cited in: Core Idea, In Practice. The foundational treatment of network effects, switching costs, and lock-in in information technology markets.
Parker, G.G., Van Alstyne, M.W., and Choudary, S.P. Platform Revolution: How Networked Markets Are Transforming the Economy. W.W. Norton, 2016. https://www.platformrevolution.com/ -- Cited in: In Practice. On how platforms create and capture value through network effects and ecosystem dynamics.
Varian, H.R. "Artificial Intelligence, Economics, and Industrial Organization." NBER Working Paper 24839, 2018. https://www.nber.org/papers/w24839 -- Cited in: Core Idea, In Practice. Analysis of how AI markets exhibit network effects through data feedback loops and how this shapes industry structure.
Bommasani, R. et al. "On the Opportunities and Risks of Foundation Models." Stanford CRFM, 2021. https://arxiv.org/abs/2108.07258 -- Cited in: In Practice. On homogenization and concentration risks when a small number of foundation models underlie a large share of AI applications.

The Economics of You -- the cost structures that network effects reinforce and concentrate
Scaling Laws -- the compute-data-capability curve that the data flywheel feeds into
The Commodity Problem -- the countervailing force: commoditization pushes against the lock-in that network effects create
Homogenization Risk -- the downstream consequence when network effects concentrate AI production in a few models
The Geopolitics of AI -- how platform concentration driven by network effects intersects with national power and regulatory response

General