← All explorations

What are the structural competitive dynamics of the foundation model industry, and can anyone besides the top 3-4 survive

Can Only the Biggest AI Companies Survive? What a Map of the Industry Actually Shows

| 149 nodes · 532 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 149-node, 532-edge knowledge graph mapping the structural forces shaping the foundation model industry.

What We’re Looking At

Imagine drawing a map of every force pushing and pulling on the AI industry — not just companies and their products, but the underlying mechanisms that make some companies stronger over time and others weaker. That’s what this analysis does. It maps 149 concepts (like “the cost of training a big AI model” or “how much electricity a data center needs”) and 532 connections between them, labeled by type (does X cause Y? does X undermine Y? does X depend on Y?) and weighted by how strong each relationship is.

Reading this map is like reading a river system. You can see where water flows, where it pools, where it gets blocked, and where it eventually drains. The question: who controls the river, and can anyone besides the four or five biggest players survive in it?

The Core Machine: Money Buys Compute Buys Better AI Buys More Money

The single most connected node in the entire map — with 56 connections — is something called the Compute-Capital Flywheel. Here is what it means in plain terms.

Training a powerful AI model requires enormous amounts of computing power. Computing power costs money. The better your model, the more customers you attract. More customers means more money. More money buys more computing power. Round and round.

This sounds simple, but what makes it structurally important is how many different sources feed into it. Eight separate money sources pour into this flywheel: Microsoft and Google subsidizing AI through their cloud businesses, Saudi and UAE sovereign wealth funds, Meta’s advertising revenue, OpenAI’s move toward a public company structure, and others. At the same time, six different constraints try to slow it down — physical limits on chip manufacturing, running out of high-quality internet text to train on, limits on how much electricity the power grid can provide.

The flywheel is the engine. Everything else either feeds it, constrains it, or gets powered by it.

Why the Big Get Bigger (And It’s Not Just One Reason)

The map shows capital concentrating at the top — and it shows at least 18 separate reasons why. That is the unusual finding. It is not that one big advantage compounds. It is that many independent advantages all point in the same direction.

Think of it like a game where one player not only owns the best properties, but also has more dice, goes first every turn, gets to set the rules for new properties, and has a hotel on the power company. Each advantage is separate. Any single one could theoretically be overcome. But when 18 of them stack simultaneously, the structural implication is that no single counterforce is sufficient to reverse the trend. Only several counterforces acting together at the same time would shift the balance.

There are counterforces in the map — cheap open-source models, efficiency breakthroughs from constrained environments, national governments building their own AI capabilities. But they are outnumbered and outweighed.

The Plot Twist: Making AI Is Getting Cheaper, But Using It Is Getting Stickier

Here is the core tension in the map, and it is genuinely contradictory.

On one hand, AI capabilities are becoming a commodity — what was expensive and rare yesterday is cheap and widely available today. The cost to run an AI model has been dropping fast. Open-source models that anyone can use or modify are getting nearly as capable as expensive proprietary ones. This should make it harder for big companies to charge premium prices.

On the other hand, companies are building AI deeply into their software workflows. Not just using AI occasionally, but making it so central to how their businesses operate that switching to a different AI provider would be enormously disruptive. Think about how hard it would be to switch your entire company’s email from one provider to another, except now imagine AI is embedded in every tool your company uses. That is the “lock-in ratchet.”

These two forces — commoditization and lock-in — are both active simultaneously. The map registers both as strong, connected to dozens of other nodes. It does not declare a winner. Both mechanisms are operating in real time in opposite directions, and the map records the tension without resolving it.

How Trying to Slow China Down Made Things Cheaper for Everyone

One of the clearest cause-and-effect chains in the map involves a policy paradox.

The United States restricted the export of advanced AI chips to China. The reasoning: if China cannot get the best hardware, they cannot train the best models. This should have protected the competitive advantage of US labs.

What actually happened: being forced to work with worse hardware, Chinese researchers had a strong incentive to figure out how to get more out of less. The result was a model called DeepSeek, trained at roughly 2% of what comparable US models cost. When DeepSeek’s results became public in early 2025, it triggered a price war in AI inference — the cost of actually running AI models to answer questions. Prices fell globally.

The map encodes the trigger edge from “hardware constraint” to “efficiency disruption” with the highest weight in the entire graph — a 10 out of 10 — indicating high structural confidence that the constraint caused the innovation. The mechanism designed to widen the gap instead produced an efficiency breakthrough that compressed prices everywhere. And the structural condition that produced it — China operating a separate computing ecosystem under ongoing hardware restrictions — remains in place. The map predicts this pattern will recur.

The Part Nobody Talks About: The Power Grid

Chips can be bought if you have enough money. Top AI researchers can be hired if you pay enough. But you cannot buy a new power grid in under a decade.

The map identifies physical electricity limits as a separate and potentially harder constraint than either compute or talent. Training large AI models requires enormous amounts of continuous electricity. The map shows companies racing to sign long-term contracts directly with nuclear power plants — not for environmental reasons, but because nuclear provides the guaranteed large-scale steady power that training runs require.

The structural finding is counterintuitive: these power contracts do not expand the total amount of available electricity. They allocate existing and near-term capacity to whoever signs first, which tends to be whoever already has the capital to sign them. Energy access becomes another dimension of the same concentration dynamic. Unlike a chip shortage, which capital can address in 6 to 18 months, a power grid shortage operates on 5 to 10 year construction timelines. The map encodes this as a binding constraint that cannot simply be purchased away.

What Happened to the Companies in the Middle

The map has a specific node for companies caught in between — not at the very top and not small specialty tools, but medium-sized AI labs trying to compete at the frontier.

These companies face a squeeze from two directions at once. From above: the top-tier labs, subsidized by major tech companies, can undercut on price indefinitely. From below: open-source models are becoming good enough to satisfy many use cases for free. The middle is compressed.

The map’s encoded endpoint for these companies: their researchers get absorbed by the larger labs. The term used is “acqui-hire” — the large company doesn’t formally acquire the mid-tier lab so much as hire away its key researchers, sometimes structuring the deal to avoid regulatory scrutiny of a formal merger. The talent concentrates further at the top, raising the compensation bar that the next round of mid-tier labs would need to match to stay competitive. The loop closes: talent leaves, the lab weakens, more talent leaves.

A Genuinely Strange Strategic Move: Giving Away the Lock

One of the more unusual structural findings involves Anthropic, the company behind Claude.

The map shows that Anthropic’s business depends on AI workflows becoming deeply embedded in enterprise software — the lock-in ratchet described earlier. At the same time, Anthropic released MCP (Model Context Protocol), an open standard that makes it easier for any AI provider to connect to external tools. The logic of an open standard is that no single provider controls it.

The map records this as Anthropic’s own protocol undermining the lock-in mechanism its business depends on. The encoded strategic reasoning: rather than capturing the lock-in itself, Anthropic chose to own the open standard that the whole ecosystem routes through — trading direct lock-in for a form of infrastructure ownership. Whether this is a net gain or loss for Anthropic’s structural position is explicitly marked as unresolved in the map. The graph records the strategy but cannot evaluate its sufficiency as a substitute.

The Tests Are Broken, and That Is Making Things Worse

There is a self-reinforcing problem in the map around how AI quality gets measured.

The standard approach: run models through standardized tests called benchmarks, and compare scores. But when benchmark scores become the primary measure of success, companies optimize specifically for those tests rather than for genuine usefulness. The tests become less reliable measures of real capability. This is a well-known problem sometimes called Goodhart’s Law — when a measure becomes a target, it ceases to be a good measure.

The non-obvious structural finding is what happens next. When the tests become unreliable, companies increase their reliance on using AI to evaluate AI — one model’s feedback used to train another. But the map shows this creates a loop: AI-generated feedback may contribute further to benchmark degradation, which increases reliance on AI feedback. The corruption of the measurement system accelerates use of the mechanism that may be contributing to that corruption. The map also notes that this same concept appears as three separate nodes with slightly different edge sets — a sign that the problem was mapped iteratively rather than cleanly, which means its actual structural weight is probably underrepresented.

Who Can Actually Survive

The map identifies one specific structural path for companies that are not in the top tier: go deep in a specific industry vertical rather than competing on general capability.

A vertical here means something like: do not try to build a better general-purpose AI. Instead, build AI that is specifically excellent for, say, radiology, or contract law, or supply chain management. The concentration at the top creates pressure that, according to the map, generates its own escape valve — but only at the application layer, not the model layer. The companies that survive are not the ones that beat the top labs at general capability. They are the ones that become indispensable within a specific domain where general-purpose models are not sufficient and where deep integration creates its own switching costs.

Bottom Line

The map shows a structurally concentrated industry, with that concentration supported by at least 18 separate reinforcing mechanisms and contested by a smaller number of undermining forces. The key structural findings, stated plainly:

The flywheel is real and multi-sourced. Capital concentration is not driven by one big advantage. It is driven by many independent advantages stacking simultaneously, which means no single counterforce is sufficient to reverse it.

The competitive axis has already shifted. The race is no longer primarily about who trains the biggest model. It is about what happens after training — fine-tuning quality, workflow integration, and control of the orchestration layer that connects AI to enterprise software.

Two opposite forces are both active. AI capabilities are getting cheaper and more widely available at the same time that workflow integration is creating switching costs. The map does not declare which force ultimately wins.

Energy may be the hardest constraint. Unlike chips or talent, power grid capacity cannot be solved with money on short timescales. The companies that secured power contracts early have a structural advantage that is difficult to replicate.

The middle is structurally squeezed. The encoded endpoint for mid-tier labs is talent absorption into larger organizations, not ongoing market competition.

Constraint produced a disruption, and the constraint remains. The mechanism that produced the DeepSeek efficiency breakthrough — a compute-constrained research environment with strong incentives to do more with less — is still structurally intact. The map’s highest-confidence prediction is that similar disruptions will recur from similar conditions.

Survival outside the top tier requires vertical depth, not general capability. The escape path encoded in the graph runs through specialization and domain-specific integration, not through competing on the same axis as the front-runners.