DeepSeek turns frontier reasoning into a cheap open input

An open model priced at a fraction of o1 erased close to $600 billion from Nvidia in a day. Value shifts to the companies that build on cheap models.
Close to $600 billion of Nvidia's market value vanished on 27 January, when the stock fell about 17%, the largest one-day loss of market value in US stock market history. The trigger was R1, a reasoning model the Chinese lab DeepSeek released on 20 January with open weights under the MIT license, shown performing on par with OpenAI's o1 on math, coding and reasoning tasks, at an API price of $0.55 per million input tokens and $2.19 per million output tokens against $15 and $60 for o1.
What the market repriced was scarcity. The price of frontier-grade reasoning, sold at $200 a month in December, fell by more than 90% in a single release, and every business that assumed the capability would stay scarce lost its premise. Value moves toward companies that own distribution, proprietary data and workflows, and away from those whose edge is a closed model a few months ahead of open weights.
Spending did not fall with the price. On 4 February Alphabet guided to about $75 billion of capital expenditure for 2025, a sign that the largest buyers are not pulling back. Cheaper training and inference widen use rather than shrink budgets, so total spending holds while the margin on access to a model narrows. The cost signal reached the training end as well: the base model behind R1 needed 2.788 million GPU hours on Nvidia H800 chips, a version cut down for export to China. Demand followed, and a week after release DeepSeek's consumer app reached the top of the US App Store.
Reasoning is now an input any team can download, fine-tune and run on its own servers. That eases gross-margin pressure for application companies and makes on-premise deployment practical for clients in regulated fields who cannot send data to an outside API. The license also allows distillation, and DeepSeek shipped smaller distilled versions alongside R1, so specialised models trained on its outputs become cheap to build. The same shift erases the moat of any startup whose product is a thin layer over one closed model, which is why model switching belongs in the design from day one and open weights belong on the company's own servers wherever data residency matters.
A single day of selling proves less than the price lists that follow it. OpenAI moved first, making o3-mini available to free ChatGPT users on 31 January, the first reasoning model on its free tier, and further cuts to reasoning APIs from US labs would confirm the trend. Washington can still tighten export rules on the chips China is able to buy. The next round of hyperscaler guidance in April settles the spending question, by showing whether cheaper models raise or trim investment in inference capacity.
A cheaper frontier is good news for the layer that builds on it. Fanaura backs companies that turn falling model prices into wider margins and new products, the application layer at the center of its thesis, and stays away from those whose only asset is privileged access to a model.
Building an AI-native company? Apply in five minutes. →
More from Fanaura
Building products that get better as open reasoning models get cheaper? Send us the project.
Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.
Get the news by email
One short note each weekday. No marketing, and you can unsubscribe at any time.