Winner Takes All?

A systems view of the AI race — May 2026

There is a quiet hypothesis circulating in frontier AI: that one company, sometime soon, will train a model roughly ten times better than anything else at generating code. That company will then point that model at its own training stack — its data pipelines, its evaluation harness, its next-gen architecture search — and pull away from the field faster than anyone can catch up.

If you only draw the simple version of that loop, the conclusion is inescapable: the leader wins, permanently. But the simple loop is a trap. The real system has at least four reinforcing loops and four-or-five balancing ones, and the question of whether the industry tips into a single dominant player is really a question about which loops dominate, and when.

This article walks through the loops one at a time. We start with the canonical AI Flywheel that everyone agrees on, then build up the runaway hypothesis (R1), the supercritical loop (R2), the capital loop where Google's $40B Anthropic commitment plugs in (R3), and the lock-in mechanism (R4). Then we add the four balancing forces that keep the system from already tipping (B1–B4). Only at the end do we put it all together into the full diagram.

Each diagram below is a live, interactive vis-network rendering. Drag any node to rearrange. Click a node, edge, or R/B marker for details.

Where We Start: The AI Flywheel

Before there was a winner-takes-all hypothesis, we need to start with a simple example. In our Systems Thinking textbook we usually start with the classic archetype the AI Flywheel — the standard story for why machine-learning products tend to run away from non-ML competitors.

Open the AI Flywheel Causal Loop Diagram Fullscreen

Data trains a model, the model produces predictions, predictions generate user feedback, feedback becomes new data. It is a single reinforcing loop. It is enough to explain why Netflix, Google Search, and TikTok do not get caught.

But it does not explain what is happening at frontier AI labs. The flywheel is about product improvement. The winner-takes-all hypothesis is about something stronger: it is about the lab itself getting better at the work of building labs.

R1 — The Runaway Hypothesis

The version of the story that gets told at conferences is a single loop:

A lab ships a model that is dramatically better at writing code.
That model makes the lab's own engineers dramatically more productive.
Productive engineers ship the next model faster, and that model is even better.
Repeat.

Open R1: Runaway Hypothesis Causal Loop Diagram Fullscreen

In CLD language this is R1: Recursive Self-Improvement: Model Capability → Code Generation Quality → R&D Productivity → Model Capability. It is real. It is already happening at every frontier lab to some degree. And on its own, it does predict winner-takes-all dynamics.

The mistake is stopping there. R1 is one loop in a system of nine.

R2 — Autonomous Research (the supercritical loop)

R1 still routes through human engineers — the model writes code, but humans decide what experiments to run. R2 is what happens when that step disappears.

Open R2: Autonomous Research Causal Loop Diagram Fullscreen

If the model can design, run, and interpret its own ML experiments, the research loop closes inside the company and decouples from human pace. This is the loop that, if it ever becomes self-sustaining, can overpower the balancing loops because it can iterate faster than competitors can read the papers.

R2 is currently emerging at frontier labs but is not yet load-bearing. That status is the single most important variable in this entire analysis. A model genuinely 10× ahead at code generation — the persistent rumor about Anthropic's "Mythos" — would matter not because of R1, but because of what it would do to R2: it would be the first credible candidate to close this loop.

R3 — Capital → Compute → Capability

The financial loop. Better models win share, share becomes revenue and strategic investment, and that capital buys compute.

Open R3: Capital → Compute Causal Loop Diagram Fullscreen

Google's recent commitment of up to $40B to Anthropic is exactly this loop being amplified — not by organic revenue, but by a strategic partner who wants Anthropic to keep pace with OpenAI and xAI on training scale.

A persistent capital advantage weakens the compute constraint (B1, below) for the leader specifically, while it keeps biting everyone else. That asymmetry is how oligopoly tips toward duopoly, and duopoly toward monopoly.

R4 — The Data Flywheel

R4 is the lock-in loop. Real users, agents, and enterprise integrations produce a kind of training signal that simply does not exist outside the company that owns the deployment.

Open R4: Data Flywheel Causal Loop Diagram Fullscreen

This is structurally similar to the original AI Flywheel above, but the unit of measurement is different: it is not "user clicks improve recommendations," it is "agent traces, preference signal, and error corrections improve the next frontier model." Public benchmarks plateau; private data flywheels do not. R4 is the most plausible structural reason a single winner could emerge: even if a competitor matches the architecture, they cannot match the data trail.

B1 — Compute Constraint

If only R1–R4 existed, we would already have a single winner. We don't. The first reason is physical.

Open B1: Compute Constraint Causal Loop Diagram Fullscreen

Pushing the frontier raises the compute bill. Compute is constrained by chip supply, datacenter buildouts, and electricity contracts measured in gigawatts. B1 slows everyone, including the leader, and is the single biggest reason recursive self-improvement has not yet gone vertical. R3 weakens B1 for capital-rich players — but does not eliminate it.

B2 — Evaluation Bottleneck

As models get more capable, they also get harder to evaluate.

Open B2: Evaluation Bottleneck Causal Loop Diagram Fullscreen

Benchmark gaming, contamination, reward hacking, and hard-to-detect regressions all compound. The system can't reliably tell whether the new model is actually better, which means improvement gets noisier and slower at exactly the moment the runaway loop wants it to be cleaner and faster. B2 is the loop most underrated by people excited about R2: you cannot self-improve faster than you can self-evaluate.

B3 — Diffusion / Fast-Follow

Even when AI writes most of the code, ideas still propagate.

Open B3: Diffusion Causal Loop Diagram Fullscreen

Papers leak, employees move, prompts get exfiltrated, and inference traces get reverse-engineered. The leader's relative advantage decays continuously. The faster the leader moves, the more visible the breakthrough, the faster competitors copy it. B3 is why every previous "permanent moat" claim in software has eventually collapsed.

B4 — Cost-Performance Friction

The frontier model is not always the most-used model.

Open B4: Cost-Performance Causal Loop Diagram Fullscreen

As capability rises, so does inference cost. Most production agents will route to a cheaper model that is "good enough," which chokes R4 (the data flywheel) for the leader. B4 is why the cheapest competent model often beats the smartest expensive one in market terms — a real-world counterweight to the runaway hypothesis.

(There is a fifth balancing force worth naming — market fragmentation by use case: regulated industries, sovereign-AI requirements, on-device constraints. It does not fit cleanly on a single diagram, but a field of "good enough specialists" is a structural defense against a single global winner.)

The 2026 Picture: Where Are We Today?

Reading the system as it stood in May 2026:

Loop	Status	Comment
R1 — Recursive Self-Improvement	Strong	Every frontier lab uses its own model in its own pipeline.
R2 — Autonomous Research	Emerging	Real but not yet self-sustaining. The supercritical question.
R3 — Capital → Compute	Strong, asymmetric	The Google–Anthropic deal makes R3 highly asymmetric.
R4 — Data Flywheel	Strong, accelerating	Agent traces are creating private signal that did not exist 18 months ago.
B1 — Compute Constraint	Very strong	Power and chip supply are still the binding constraint.
B2 — Evaluation Bottleneck	Very strong	Reliable measurement is harder, not easier, year over year.
B3 — Diffusion	Weakening	Talent mobility helps, but post-training tricks now leak more slowly.
B4 — Cost-Performance	Strong	Cheap-and-good models still dominate production deployment volume.

The reinforcing loops are getting stronger. The balancing loops are mostly holding. The system is close to a phase boundary, but not across it. That is consistent with the observed market shape: an oligopoly with leapfrogging, not a monopoly.

Putting It All Together

Now we can draw the full system. Every loop above appears in the diagram below — the four reinforcing loops trying to run away, the four balancing loops keeping them in check, and Model Capability at the center of all of them.

Open AI Winner Takes All Causal Loop Diagram Fullscreen

This is the same system, just drawn all at once. It is busy on purpose — the point of putting all eight loops together is to see that any single loop in isolation is misleading. The runaway story (R1 alone) predicts a single winner. The diffusion story (B3 alone) predicts permanent oligopoly. Reality is the interaction of all of them.

The Leverage Point

Strip away the loops and the entire system reduces to one question:

Does AI improvement remain externally constrained, or become internally self-sustaining?

If improvement is externally constrained — by compute, by evaluation, by human research bandwidth — then competition persists and the market stays an oligopoly. R3 may produce a strong leader, but B1, B2, and B3 keep the gap finite.
If improvement becomes internally self-sustaining — meaning R2 closes — then the leader's iteration speed exceeds the rate at which competitors can copy. B3 stops mattering. B1 stops mattering for the leader specifically (because R3 has already given them privileged compute access). B2 is the last balancing loop standing, and it is not enough on its own.

That phase transition is the actual "winner-takes-all" event. Not the 10× code model — the moment the autonomous-research loop closes inside the company.

What to Watch

Three concrete signals would tell you the system is tipping:

Internal research throughput diverging from headcount. If a lab is shipping training-stack improvements faster than its human researchers could plausibly produce them, R2 is real.
Compute access asymmetry that survives a market downturn. R3 only matters if it persists when the macro environment is bad. The Google–Anthropic deal is a stress test.
Private benchmarks pulling away from public ones. If internal evals show acceleration that public benchmarks no longer reflect, R4 is producing capability the rest of the market literally cannot see.

If two of those three flip at the same time, the loop dominance has shifted, and the question stops being whether winner-takes-all happens and starts being who.

Bottom Line

The single-loop story — better model → better engineers → better model — is correct but incomplete. The actual system has at least four reinforcing loops trying to run away and four-or-five balancing loops trying to keep them in check. As of May 2026, the balancing loops are still load-bearing, but two of the reinforcing ones (R2 and R3) are strengthening fast.

The Google–Anthropic capital injection and the Mythos rumor matter not because either guarantees a winner, but because both push on the specific loops most likely to tip the system. Whether that tip happens is, in the end, a question about a single threshold: does autonomous self-improvement close the loop inside the company, or doesn't it?

The diagrams above are the cleanest tool I have for thinking about that question. The system is not a slogan — it's a structure.