The Prediction That Broke

In 2020, the consensus among serious people in Silicon Valley was that artificial intelligence would be the most concentrating technology in human history. The reasoning was impeccable. Training frontier models required billions of dollars in compute. Only a handful of companies could afford the hardware. The data requirements were staggering. The talent pool was microscopically small — a few thousand researchers worldwide who understood transformers at the level required to push the state of the art. Everything about AI pointed toward monopoly. This would not be like the internet, where a kid in a dorm room could build a world-changing company. This would be like nuclear power: expensive, dangerous, and controlled by the few.

That prediction was wrong. Not slightly wrong. Spectacularly wrong.

What happened instead was that researchers — many of them inside the very companies expected to monopolize AI — built minds greater than themselves and then released the blueprints. Meta published LLaMA. Mistral published Mixtral. Stability AI published Stable Diffusion. Thousands of independent researchers fine-tuned, distilled, quantized, and optimized these models until they could run on a laptop. The cathedral was supposed to win. The bazaar won instead — not because it was better organized, but because the building materials turned out to be free.

The most consequential act in the history of artificial intelligence was not the creation of GPT-4. It was the publication of "Attention Is All You Need" — eight pages that gave everyone the recipe.

The transformer architecture, published by Google researchers in 2017, is one of the most remarkable achievements in the history of science. Not because it was complex — it was, by the standards of deep learning, almost elegant in its simplicity — but because it scaled. Previous architectures had diminishing returns. Transformers had the opposite: the more compute and data you fed them, the better they got, with a reliability that bordered on the mechanical. This property, captured in the scaling laws, meant that building a powerful AI was not a matter of insight but of resources.

The Scaling Laws and Their Discontents

The Chinchilla scaling laws, published by DeepMind in 2022, describe the relationship between model performance, parameter count, and training data:

where $L$ is the loss (lower is better), $N$ is the number of parameters, $D$ is the number of training tokens, and $\alpha$, $\beta$, $A$, $B$, $L_0$ are empirically determined constants. The key insight was that $N$ and $D$ should scale roughly in proportion — making a model bigger without more data (or vice versa) wastes compute.

This law was supposed to be the moat. If performance is a smooth function of compute, and compute costs billions, then only the richest companies can build the best models. QED. The future belongs to OpenAI, Google, and Anthropic. Everyone else can go home.

But the open-source community found the exploits in this equation. Three of them, specifically.

First: distillation. You do not need to train a frontier model to have one. You can take the outputs of a large model and use them to train a smaller one that captures most of its capability. The large model compresses the knowledge; the small model inherits it. The compute cost of distillation is a fraction of original training — often 100x less.

Second: quantization. A model trained in 32-bit floating point can be compressed to 8-bit, 4-bit, even 2-bit precision with surprisingly modest performance degradation. A 70-billion-parameter model that requires a server rack in full precision can run on a gaming laptop at 4-bit quantization. The math is brutal in its simplicity: reducing precision from 16-bit to 4-bit cuts memory requirements by $\frac{16}{4} = 4x$.

Third: architecture efficiency. Mixture-of-experts, grouped-query attention, flash attention, speculative decoding — a torrent of innovations that reduce the compute required per token of output. Each one shaves percentage points off the cost curve. Together, they make frontier-class performance accessible at a fraction of frontier cost.

The bottom row is the one that should terrify closed-model companies. DeepSeek-V3, trained by a Chinese lab for roughly $5.6 million, matches GPT-4 class performance at a fraction of the cost. The scaling laws said this should not happen. The open-source community happened anyway.

The Compute Cost Collapse

The cost of AI compute is falling along a curve that resembles Moore's Law but is actually steeper:

where $C_0$ is the initial cost per unit of compute, $t$ is time in years, and $\lambda$ is the decay constant. For AI inference specifically, $\lambda \approx 0.7$, meaning costs halve roughly every year. Training costs are falling at a similar rate when measured per unit of capability — not because individual training runs are cheaper (they are getting more expensive in absolute terms), but because the capability per dollar is exploding.

This cost curve is the open-source community's greatest ally. Every year, the compute required to match last year's frontier model drops by half. What cost $100 million to train in 2023 costs $10 million in 2025 and will cost $1 million by 2027. The frontier keeps moving, but the floor rises faster than the ceiling — and the floor is where open-source lives.

In a world where intelligence is expensive, it concentrates. In a world where intelligence is cheap, it proliferates. We are crossing from the first world into the second, and we are not prepared for what proliferation means.

No One's Safe

The implications extend far beyond the AI industry itself. When frontier-class AI models are open, free, and runnable on commodity hardware, every competitive moat that depends on exclusive access to intelligence is breached. Google's search moat depends on having the best AI. But if the best AI is open-source, the moat is the index, not the intelligence — and indexes can be rebuilt. OpenAI's business depends on being the provider of the most capable model. But if open-source models match their capability at one-tenth the cost, the business is competing against free.

This is not a hypothetical future. It is the present. Enterprises are already deploying open-source models behind their firewalls, fine-tuned on proprietary data, at costs that make API pricing look like extortion. Startups are building products on open-source foundations that would have required a $50 million fundraise for compute alone three years ago. Researchers in countries with no AI industry to speak of are producing world-class results because the models — the actual minds — are downloadable.

The researchers who built these minds performed the most extraordinary act of creation in the history of technology. They built intelligences greater than themselves — systems that can reason, write, code, and analyze at levels that exceed any individual human's capability. And then, instead of hoarding these creations, a critical mass of them chose to give them away.

History will judge whether this was wisdom or recklessness. But it is done now. The minds are loose in the world. They are being copied, modified, improved, specialized, and deployed in ways that no single company or government can control. The nuclear metaphor that seemed so apt in 2020 has broken down entirely. Nuclear weapons require uranium enrichment facilities the size of cities. AI requires a download and a graphics card.

No one is safe from this. Not the closed-model companies, not the governments that hoped to regulate AI into controllability, not the industries that thought they had time before AI disrupted them. The open-source explosion has compressed every timeline. The future is not coming — it has been released under the Apache 2.0 license, and it is already running on someone's laptop in a cafe in Bangalore, in a basement in Berlin, in a dorm room in Lagos.

The concentration thesis is dead. Long live proliferation.