Nvidia lost $589 billion on January 27, 2025. Not over a quarter. Not across a bad week. In a single trading session — the largest one-day wipeout in U.S. stock market history. No fraud. No product recall. No macroeconomic shock. A research paper did it. A lab called DeepSeek, operating out of Hangzhou, had trained a model competitive with GPT-4 for approximately $5.6 million. OpenAI's GPT-4, by widely cited estimates, cost north of $78 million. The market looked at the ratio — seven cents on the dollar — and repriced trillions of dollars in assumptions about who gets to own intelligence.
We watched the ticker that Monday morning the way you watch a building being demolished from the inside out. Clean. Fast. The old thesis evaporating floor by floor.
How did a Discord joke become an industry standard?
The origin story of open AI is not a corporate strategy memo. It is a group of researchers who were, by their own admission, partly joking.
EleutherAI formed in 2020 on a Discord server. The name comes from the Greek word for liberty. The stated goal was to replicate GPT-3, which OpenAI had described in a paper but declined to release. No funding. No institutional backing. No business plan. Somebody made a channel. People showed up.
Within two years they had built GPT-NeoX-20B, assembled the Pile (a training dataset now used across the industry), and — here is the part that still makes us shake our heads — created the evaluation benchmarks that became the standard for measuring model performance. The companies whose monopoly EleutherAI set out to break now measure their own models using EleutherAI's yardstick.
We have covered a lot of emergent phenomena in this series. This one has a specific texture. It smells like a garage band whose demo tape accidentally becomes the recording industry's reference track.
Can you build a frontier model without frontier capital?
DeepSeek said yes, and then proved it with receipts.
Their approach combined architectural innovations — mixture-of-experts routing, multi-head latent attention — with training efficiency techniques that squeezed more capability out of less compute. The technical details matter less than the structural implication: the knowledge required to replicate the approach is public. It lives in papers, code repositories, and forum threads. Each generation of open models teaches the next generation how to be cheaper.
But DeepSeek was one data point. Three former DeepMind researchers in Paris provided another. They founded Mistral AI. Within two years, all three were billionaires. Not through monopoly pricing. Through building frontier-competitive models with radically smaller teams and releasing them openly. Mistral proved that the assumed relationship between headcount and model quality was weaker than the industry believed — or wanted to believe.
Then Meta entered with Llama. Corporate motivations are always mixed, and we are not naive about Meta's reasons for open-sourcing a frontier model. Commoditizing the layer your competitors are trying to monetize is good strategy. But motivations are one thing. Effects are another. Llama has crossed 1.2 billion downloads. Roughly half of Fortune 500 companies are piloting Llama-based systems. The horse left the barn, circled the globe, and is now teaching other horses to run.
What do 2 million models actually mean?
Hugging Face hosts over 2 million models. Fifty billion cumulative downloads. Fifteen million downloads per day. We want to sit with those numbers for a moment, because they are easy to read past.
Two million models means two million experiments in what intelligence can do. Most of them small. Many of them terrible. Some of them extraordinary. The point is not quality control — it is combinatorial exploration at a scale no centralized lab could match. Alibaba's Qwen model family alone has spawned over 113,000 derivative models. Each derivative is a team or an individual who took the base, bent it to a specific problem, and often released the result back into the commons. One hundred thirteen thousand forks of a single model. That is not a platform. That is an ecosystem with its own weather patterns.
China has produced 1,509 of the 3,755 publicly tracked large language models worldwide — roughly 40%. Not through a single government program. Through hundreds of labs, universities, and companies operating independently. The geographic distribution of AI capability is already multipolar. By the time policy circles finish debating whether it should be, the question will be moot.
How fast is the cost actually falling?
Two numbers tell the story.
Training: $78 million to $5.6 million. A 93% drop. And that is the headline figure — subsequent open models have pushed costs lower still.
Inference: approximately $36 per million tokens in early 2024. Between $0.14 and $0.28 per million tokens in early 2026. Call it a thousand-fold reduction in roughly two years. Ten-x per year.
We want to translate that out of token-pricing abstraction. A startup that budgeted $100,000 for AI inference in 2024 can now get equivalent capability for under $100. An individual experimenter who could afford a few hundred API calls can now afford hundreds of thousands. The cost barrier has not lowered. For most practical purposes, it has disappeared.
We are honestly not sure the market has fully processed this. The conversation still assumes intelligence is scarce. The spreadsheets say otherwise.
Does the quality gap still exist?
MMLU — Massive Multitask Language Understanding — is the standard benchmark for comparing model capability. In 2023, the gap between the best open model and the best proprietary model was significant. Multiple percentage points that translated into real differences on real tasks.
As of early 2026, that gap is 0.3 percentage points. Three-tenths of one percent. On most practical applications, the difference between the best model you can download and run on your own hardware and the best model behind an API paywall is within measurement noise.
We want to be precise about what this does and does not mean. It does not mean open models are identical to proprietary ones. It does not mean there are no tasks where proprietary models retain an edge. It means the general capability gap — the thing that justified premium pricing and lock-in — has closed to a margin that most practitioners cannot detect in their actual work.
The argument for paying ten or fifty times more for proprietary inference now rests on brand trust, enterprise support agreements, and integration convenience. Those are real considerations. They are not a moat.
What breaks when intelligence is cheap?
This is the question we keep returning to. Not whether open AI will replace proprietary AI — that question is already being answered by download counts. The harder question is what happens to every system, institution, and business model that was built on the assumption that intelligence is scarce and expensive.
Education designed around credentialing access to expertise. Healthcare structured around specialist gatekeeping. Legal systems that charge by the hour for pattern-matching that a $0.14 API call can approximate. Financial advice priced as if insight were a rare mineral.
We are not predicting the collapse of these systems. We are noting that their economic foundations have shifted, and that the shift happened faster than most participants in those systems have registered. The gap between what intelligence costs and what institutions charge for it is now measured in orders of magnitude. Gaps like that do not persist indefinitely. They close — sometimes gradually, sometimes all at once.
We do not know the timing. We know the direction.
One thing we feel confident saying: the extractive playbook for AI — concentrate capability, charge for access, create dependency — assumed that training costs would create natural monopolies. That assumption has been falsified by data. Not by argument. By DeepSeek's $5.6 million receipt, by Hugging Face's two million models, by an inference cost curve that makes Moore's Law look leisurely.
Intelligence is a commodity now. Not as a forecast. As a measured fact, with fifteen million downloads a day to prove it.
What people do with cheap, abundant intelligence — that is the open question. And it is a question about wisdom, not technology. We will be watching.
Episode 10 of the Superpuzzle Developments series. Intelligence has been democratized — not as a political project but as an economic inevitability. The cost collapsed. The quality converged. The question now is not who has access to intelligence, but who has the direction to use it well.