# DeepSeek and Monopoly Economics - Author: Kojin Glick - Published: 2025-01-29 - Canonical: https://www.kojinglick.com/blog/deepseek-and-monopoly-economics --- When a model makes headlines like this, the media’s “Breaking News” tunnel vision often obscures what’s important. Between the Mixture of Experts architecture, the Chain of Thought capabilities possessed by the R1 model, there’s a lot to discuss about the model in isolation. But DeepSeek hasn’t done anything particularly novel by itself. DeepSeek’s challenge to ChatGPT is fundamentally about the future of the AI economy, and whether the walls built by the partnership between OpenAI and the new administration are high enough to keep challengers, Chinese or otherwise, out. ### What DeepSeek _does_ DeepSeek, as a family of models, uses an architecture called Mixture of Experts (MoE), [originally pioneered by Mistral in France](https://mistral.ai/news/mixtral-of-experts/). In a Mixture of Experts model, parts of the neural network are specialized to certain tasks, and a traffic controller, sometimes called a router or a load balancer, directs certain requests to its entourage of “Expert” neural networks. When it generates new text, it’s also far more efficient because it doesn’t use the entire network in generation. ### What DeepSeek _did_ DeepSeek scares Silicon Valley, not only because of what it does, but how it got there. DeepSeek trained a 607B parameter chat model, like OpenAI’s gpt-4o, and a 70B distilled reasoning model, like OpenAI’s gpt-o1. If we assume that the unit cost of a GPU-hour was around $2, [the price tag for training them was $5.576 million](). Though we don’t know exactly how much OpenAI spends on their models, we do know that in the year that OpenAI released gpt-4o, they spent $5 billion. These discounts are passed to AI entrepreneurs in the almost 10x reduction in price to use the models for enterprise purposes. It is cheaper, by an order of magnitude, to build your AI product idea on DeepSeek than it is on ChatGPT. Some of this cost-reduction can be explained by the efficiency of training many domain experts rather than one massive oracle, but the shock in Silicon Valley and Wall Street reveals the emperor’s lack of clothing. | Model | Input Tokens\* (per million) | Output Tokens (per million) | | :---------------------------------------------------------------------: | :--------------------------: | :-------------------------: | | [deepseek-chat](https://api-docs.deepseek.com/quick_start/pricing/) | $0.07 \- $0.27\*\* | $1.10\*\* | | [deepseek-reasoner](https://api-docs.deepseek.com/quick_start/pricing/) | $0.14 \- $0.55 | $2.19 | | [gpt-4o](https://openai.com/api/pricing/) | $1.25 \- $2.50 | $10.00 | | [gpt-o1](https://openai.com/api/pricing/) | $7.50 \- $15.00 | $60.00 | \* The two prices represent the difference in cost of inputting a token that the large language model is already familiar with, vs a token that is new to the large language model. \*\* Expected cost after the promotional pricing that ends Feb 8\. ### The price of monopoly is the entire market Wall Street and Silicon Valley are being faced with an important crossroads: continue pursuing monopoly economics in AI and lose the wholesale market to China, or prioritize their existing bottom line by nurturing a robust domestic AI economy. The companies that make up “Silicon Valley” act predictably when they search for new markets to capture. When distributed systems became the standard computation paradigm in the 2010s, Silicon Valley was able to capture the market by turning the Cloud into economies of scale: branded data-centers, billions in broadband investment, and bulk priced access to their walled gardens. After OpenAI launched ChatGPT in 2022, investors predicted that AI adoption would follow Cloud adoption. Huge losses are expected in these endeavors, since the real reward is the capture of the market. Since the typical threat to the monopolistic corporation is the antitrust enforcement of the government, Sam Altman and OpenAI have been hard at work trying to capture regulators through [the Stargate Project](https://openai.com/index/announcing-the-stargate-project/). Between aligning the federal government’s pocketbook with their long term ambition to own and operate the infrastructure of AI in the United States, the current administration and OpenAI are making a bet that the future of AI is the same thing as the future of OpenAI. If DeepSeek can operate at 10% the cost of OpenAI, where is the use for OpenAI, and OpenAI-branded datacenters? At the core of the Stargate Project is the assumption that AI entrepreneurs are only going to use OpenAI services to build the next generation of AI tooling. But that assumption can not stand a 90% cheaper competitor. Unless OpenAI is willing to reduce potential revenue by price matching, doubling down on the OpenAI monopoly is tantamount to handing a majority of the AI infrastructure market to China. In order to keep AI entrepreneurial dollars in the United States, OpenAI and the new administration must start investing in open infrastructure. This consists of open research and a clear, simple path to competitiveness for smaller players.