← All explorations

Why might open-source AI models win despite the 'top 3-4 survive' thesis — what structural forces favor diffusion over concentration

Will Everyone Get Good AI, or Will Only a Few Companies Control It?

| 91 nodes · 313 edges
↓ .md ↓ .db Take this into your AI — the full analysis + graph as markdown, ready to paste into ChatGPT, Claude, Gemini or any AI.

Based on analysis of a 91-node, 313-edge knowledge graph about the structural forces shaping AI market concentration versus diffusion.

The Question in Plain English

Some smart people who study technology markets believe that AI will end up like search engines or social media: a few giant companies will win, and everyone else will use their products. The argument is simple — building the best AI requires enormous amounts of money and computers, so only a handful of companies can afford to stay at the frontier.

But there is another possibility: AI tools spread out to many places, many companies, and many countries. This is sometimes called “diffusion.” A knowledge graph mapping out this debate — with 91 concepts and 313 connections between them — shows something interesting. The structural forces pushing toward diffusion are more numerous, more connected, and more reinforcing than the forces pushing toward concentration. Here is what that means in plain terms.

The Scoreboard Is Rigged in a Surprising Way

Imagine you are mapping a neighborhood argument. The loudest person in the room — the one everyone keeps talking about — is not necessarily winning the argument. They might just be the person everyone is disagreeing with.

That is what the graph shows. The two most-connected concepts in the entire map are “Big AI companies controlling everything” (connected to 30 other ideas) and “Getting stuck in one AI provider’s tools” (connected to 18 other ideas). But here is the twist: both of those concepts have the lowest importance score in the graph. Why? Because almost all of those connections are people and forces pushing against them. About 15 different mechanisms are lined up to poke holes in the concentration argument, and fewer than 5 are supporting it.

High connectivity at low weight is like being the most-argued-against position in a debate. You are central to the conversation, but not because you are winning.

The Most Powerful Concept: The “Weapon” Nobody Planned

The single most causally important idea in the graph is something called the “Subsidized Open-Source Weapon.” The name sounds technical, but the idea is simple.

Imagine you run a grocery store. A new specialty coffee shop opens nearby and starts taking your customers. You cannot compete with their coffee directly — they are specialists. But you can start giving away decent coffee for free, because coffee is not your main business. You make money on groceries. The coffee shop, which only sells coffee, gets hurt. You do not.

That is what large technology companies like Amazon, Google, and Microsoft are doing with AI. They make most of their money from cloud computing, advertising, and software subscriptions. AI is a threat to some of that, but it is also something they can release for free or very cheaply, because it pushes more people to use their cloud infrastructure. When they give away capable AI tools, they hurt companies whose only business is selling AI.

The graph shows this mechanism being fed from many independent directions at once: national governments wanting their own AI, Chinese industrial companies building AI into manufacturing, European countries worried about data leaving their borders, and open-source communities that want AI to be free for everyone. All of these different actors, with very different motivations, end up doing the same thing: releasing capable AI tools openly. The weapon fires itself, even without coordination.

The Real Answer: Not “Who Wins” but “Who Wins at What”

The graph does not actually answer the question “Will open source beat closed AI?” because that is the wrong question. The actual structural answer is: it depends on which layer of AI you are asking about.

Think of it like the restaurant industry. Growing and harvesting wheat is extremely expensive and concentrated — a few giant agribusinesses do almost all of it. But cooking and serving food is distributed everywhere. Concentration at the ingredient layer does not mean concentration at the restaurant layer.

AI works similarly:

  • The training layer (building the foundational models from scratch) remains highly concentrated. You need billions of dollars and rare computing hardware. This probably stays in the hands of a few.
  • The inference layer (actually using AI to answer questions or do tasks) is getting cheap very fast, and open tools are spreading here.
  • The application layer (building useful products on top of AI) is where most of the future business value will be captured — and here, neither open nor closed has decisively won.

The graph explicitly names this as a “Both/And Answer.” Concentration and diffusion can coexist in different layers of the same industry simultaneously.

Why Restrictions on Chips Made Open-Source Stronger

One of the most counterintuitive findings: the rules designed to slow down AI development in certain countries may have accelerated open-source diffusion.

When the United States restricted the sale of advanced AI chips to China, Chinese AI labs faced a shortage of computing power. The predictable response was to get more done with less. Engineers developed more efficient techniques — ways to train AI that require fewer chips. Then, following academic norms, they published those techniques openly.

The result: the restrictions on hardware accelerated the development and publication of efficiency breakthroughs that anyone in the world could then use. Every major output of the export control mechanism in the graph points toward open-source diffusion, not away from it. The restriction created the innovation pressure that produced the thing it was trying to prevent.

Four Engines Running at the Same Time

The graph identifies four self-reinforcing loops — situations where A causes B causes C causes A again, in an accelerating cycle.

The Price-Volume Loop: When AI becomes more efficient, it costs less to run. When it costs less to run, more people use it. When more people use it, the cloud companies that provide the computing make more money. When they make more money, they can afford to subsidize more open-source AI. When more open-source AI is available, engineers make it even more efficient. The cycle feeds itself.

The Distillation Loop: When a large AI model exists, engineers can train smaller models to imitate it, producing a “distilled” version that is cheaper but nearly as capable. Those smaller models can then be fine-tuned by communities of volunteers who improve them further. Those improvements generate training data that improves the next generation of models. The loop does not require permission from the original model’s creator to keep running.

The Geopolitical Loop: Export controls push countries toward independent AI development. Independent development produces openly published models (because countries want to demonstrate capability and influence). Openly published models spread globally. That spread makes further restriction politically difficult and practically ineffective. Ineffective restriction leads to more pressure for restriction. The loop escalates regardless of whether the restrictions achieve their stated goals.

The China Industrial Loop: Chinese manufacturing companies deploying AI in factories generate real-world data at enormous scale. That data improves models. Improved models lower deployment costs. Lower costs expand deployment. More deployment generates more data. The loop runs on industrial volume rather than research investment.

Three Things That Are Not Obvious

The safety work on closed AI accidentally creates demand for open AI. Closed AI companies add restrictions to their models to make them safer — they refuse certain requests, add guardrails, and reduce performance on tasks that could be misused. This is intentional. But it also makes their models less useful for certain legitimate business applications. Companies that need the unrestricted behavior build their own models or use open ones. The safety effort creates the market for the alternative.

The biggest chipmaker has the same interest as the open-source community. NVIDIA sells graphics processing units — the hardware used to run AI. Their revenue goes up when more AI is being run. Open-source AI models, which spread to more users and more use cases than closed models, generate more total computing demand than concentrated closed models would. NVIDIA therefore benefits from diffusion. The graph shows NVIDIA’s structural incentives aligned with open-source diffusion — not for any philosophical reason, but because diffusion means more chips sold.

Benchmark scores losing credibility helps open-source models. When closed AI companies market their models, they often point to scores on standardized tests: “Our model ranked #1 on this benchmark.” But those benchmarks are being gamed — labs train specifically to score well on tests, which makes the scores less meaningful as measures of real capability. As general benchmarks become unreliable signals, the purchasing decision shifts to “does this work well on my specific task?” That question favors fine-tuned open-source models, which can be specialized for any domain, over general-purpose closed models that cannot be customized.

What Is Still Genuinely Uncertain

The graph is honest about several things it cannot resolve.

Whether decentralized training — training large AI models across many small computers rather than one giant data center — will ever be as effective as centralized training. Early experiments suggest it is possible; whether it reaches competitive scale is unknown.

Whether the market for small and medium-sized businesses ends up on open or closed AI. Right now, small businesses mostly use closed AI because open AI requires engineering expertise to deploy. The graph identifies this as a “paradox” without resolving it.

Whether safety regulations will help or hurt concentration. Large AI companies are involved in writing safety rules, which could create rules that only they can follow — locking out open-source competition. But safety techniques are also becoming easier to apply, which might erode that advantage. The graph treats this as an active, unresolved tension.

The Bottom Line

The knowledge graph maps a situation where the forces favoring AI diffusion are more numerous, more interconnected, and more self-reinforcing than the forces favoring concentration — but where concentration is not disappearing, it is relocating.

The headline finding is not “open source wins.” It is that concentration is shifting vertically: from the whole AI stack to just the most expensive layer (building foundational models from scratch), while everything above that layer — running models, improving them, building products with them — is under sustained, multi-directional pressure toward openness.

The pressure comes from sources that do not coordinate with each other: national governments, economic competitors, chip companies, efficiency researchers, open-source communities, and inadvertent policy feedback loops. When this many independent actors push in the same structural direction for different reasons, the graph treats that as a more durable signal than any single actor’s strategy.

The question “top 3-4 survive” may be asking about the wrong layer.