Open-Source Anarchy: How DeepSeek Upended the AI Power Game

In January 2025, DeepSeek released an AI model you could download, modify, and run yourself. Its published results put it in serious competition with OpenAI’s o1 on several reasoning benchmarks. The Chinese company also gave developers unusually broad permission to build on it. That was a wonderfully inconvenient development for anyone hoping to collect rent on the future. DeepSeek’s R1 release announcement

The next day, OpenAI and SoftBank announced Stargate: a company intending to invest $500 billion over four years in American AI infrastructure. Its initial equity funders included Oracle and MGX. Two announcements, twenty-four hours apart: a downloadable model and a construction budget large enough to require its own foreign policy. Stargate’s original announcement

When I first wrote about DeepSeek, I saw an opening for people who would rather own their tools than petition for access to them. I still do. But the original argument got carried away in places. It confused a training bill with the cost of building an AI company, treated hardware constraints too casually, and drew the line between open and closed systems along national borders where it never belonged.

Twenty months later, there is a better case to make. It requires less prophecy and has considerably more evidence.

Start with what DeepSeek actually released. R1’s main model weights, the learned numerical parameters that make the model work, came under the permissive MIT license. The accompanying repository explicitly allowed commercial use, modification, and derivative work. Developers could take possession of a capable model and build with it outside DeepSeek’s own service. Smaller distilled models were also released, with additional licensing details depending on their underlying models. R1 repository and licensing

That freedom deserves precise language. “Open weights” is the safer description. A downloadable model does not automatically come with the complete training code and data information needed to reproduce its development. The Open Source Initiative’s definition requires those elements too. You can have substantial freedom to use a machine without receiving everything required to rebuild it from scratch. Open Source AI Definition

For a developer, the immediate advantage is practical. A permissively licensed model can become part of a product you control. With a fully local setup, your work need not pass through the original developer’s servers. You can keep using your copy when the company changes direction. You also inherit the work and expense of operating it.

That last part matters. The original suggestion that anyone with an electric bill could run the whole thing was nonsense. Permission to download a model does not supply the hardware to run it well. Different model sizes involve different costs and compromises. A free grand piano is still a problem on the third floor.

The famous $5.58 million figure needs the same sobriety. DeepSeek’s V3 technical report estimated $5.576 million for its official training run, using 2.788 million Nvidia H800 GPU-hours at an assumed rental rate of $2 per hour. The report explicitly excluded earlier research and experiments. That figure belonged to V3, the model underlying R1’s development, and was never an all-in invoice for the company or the R1 program. DeepSeek-V3 technical report

It was evidence of efficient engineering. Comparing it directly with Stargate’s multiyear infrastructure budget was accounting by bar fight. One number described a particular use of computing equipment; the other described a plan to build enormous amounts of it.

Nor did DeepSeek make advanced chips optional. That same report described a cluster of 2,048 H800 GPUs. Clever software helped the company get more from expensive hardware. It did not abolish expensive hardware. The result alone cannot tell us whether export restrictions helped or hindered China overall, much less prove that restrictions were the sole cause of the breakthrough.

The case for competition survives the bad accounting. A powerful alternative can matter without being cheap to invent, cheap to operate at every scale, or superior at every task. What matters is whether it gives the customer a credible choice.

DeepSeek kept going. Its official release history records V3.2 in December 2025, V4 in April 2026, and V4.1-Flash on September 10, 2026. The latest release added native visual understanding, and the company announced another reduction in API prices. Those are documented product developments. Its claims about beating other models still deserve the usual scrutiny of company-published benchmarks. DeepSeek’s dated release history

OpenAI, meanwhile, released gpt-oss-120b and gpt-oss-20b in August 2025 under the Apache 2.0 license. These were downloadable models designed for independent deployment. OpenAI said the smaller one could run with 16 GB of memory. The company behind Stargate was also supplying AI that people could run on their own equipment. OpenAI’s gpt-oss announcement

That makes the original geography of good and evil untenable. An American company can sell a closed service and release open weights. A Chinese company can distribute useful weights and operate a restrictive service. Judge the permissions attached to the product. The flag on the headquarters will not tell you whether you hold the keys.

The prediction of a uniform Washington crackdown also needs revising. The White House’s July 2025 AI Action Plan explicitly endorsed open-source and open-weight models, including their value to startups that want independence from closed providers. A policy document is no guarantee of future conduct, but it is evidence, and an honest update has to account for it. America’s AI Action Plan, page 4

Meanwhile, the infrastructure boom survived. In its April 2026 assessment, Epoch AI identified operating capacity at Stargate’s Abilene site and six additional American sites under construction. Its much larger capacity estimates remained projections. The distinction matters: an announcement cannot answer a prompt, and a projected data center cannot keep the lights on. Epoch AI’s Stargate assessment

We have downloadable models and construction cranes. The cranes do not disprove the value of the downloads. The downloads do not make the cranes obsolete. Cheaper ways to perform a task can make more uses of that task worthwhile, giving people reason to buy more computing power.

The uncomfortable possibility is that small businesses gain tools they could never have built while the largest companies grow richer selling the infrastructure underneath them. I would expect plenty of that. Access can spread without wealth or industrial power spreading evenly alongside it.

There is another mistake worth avoiding: treating DeepSeek as a Chinese liberty movement with a download button.

Researchers studying R1 documented censorship of politically sensitive subjects, including in locally run versions. Their work also explored ways to bypass or reduce it. Downloading the weights does not automatically remove behavior learned during training. You can take the model out of the company’s servers without taking every company preference out of the model. R1dacted: Investigating Local Censorship in DeepSeek’s R1 Language Model

A model that gets nervous about the wrong historical question is a peculiar instrument of liberation. Nevertheless, the ability to study and alter its behavior is valuable. You need not admire its makers to appreciate having fewer obligations to them.

This is where the case for open models becomes strongest. It asks less of human nature. We do not have to establish that a founder is benevolent, that a regulator is wise, or that a company’s mission statement was handed down on stone tablets. We can ask what happens when the relationship goes bad.

Can you retain the model? Move it? Modify it? Hire somebody else to run it? Keep your data on equipment you control?

Those are questions a business owner can understand without attending a conference about responsible transformation. They concern the cost of leaving.

Suppose you build an important part of your business around an AI service. A price increase, a discontinued model, or a change in permitted uses could become your problem overnight. A suitable model you can operate independently gives you another course of action. It may require technical help and money. It may perform worse on some tasks. But the choice exists, and that changes the relationship.

That gives open models a way to discipline the market even among people who never run one. A paid provider has a harder sale when customers can obtain something good enough elsewhere. That does not mean proprietary AI disappears. Reliability, convenience, support, and superior performance can all be worth paying for. Companies should have to earn that payment.

My expectation now is continued competition between those arrangements: enormous hosted systems for customers who want their capabilities and convenience, alongside increasingly useful models that individuals and businesses can possess. The balance will depend on hardware costs, model quality, licensing, and policy. No release schedule guarantees freedom.

DeepSeek gave the original argument something real to work with. The mistake was letting excitement turn an opening into a victory parade. The practical opportunity remains: more people can choose how to use capable AI, and some can choose who gets to supply it.

I do not need an AI company to promise it will love humanity forever. I would like the option to fire it.

Originally published January 23, 2025. Revised September 16, 2026.