January, 2025 is a hectic month for AI development. The biggest development is Deepseek’s shock, which brought down $1T US stock value, although the Deepseek-R1 model was actually released in late December, 2024. As the previous article[1] mentioned, foundation models can be US-based models or China-based models. Foundation models can also be closed models or open models. Deepseek’s shock involved both these two dimensions. But besides these two dimensions, there is an extremely important third dimension, i.e., cost or economics dimension. DeepSeek training cost (or the last training run) is $6M, while many state of the art models training cost is a few $10M’s [2] to $1B ($1B may be a tactic to scare away competitors, but obviously it backfired). It is clear now that the training cost can definitely be cheaper, much cheaper, and at least 10x cheaper. When something is 10x cheaper , it is a big change and people will try it and adopt it. As we build and empower AI shopping agents, we cheer on models which are more powerful, more widely adopted, and cheaper. Lower cost means better ROI opportunities for AI application builders such as AI shopping agent builders.

For AI shopping agent builders, we care more about what is the cost of utilizing foundation models and what is ROI after we deploy AI shopping agents in the production. To utilize foundation models, we have two ways. One is to call the foundation model APIs directly. The other is to deploy an open model ourselves. For the foundation model API calls, i.e. inference, DeepSeek APIs are cheaper, much cheaper, and at least 10x cheaper. Check the latest pricing on these model providers’ websites. There may be a price war, which is good for AI shopping agent builders. Many people try to deploy open models themselves on cloud or local servers to further save the utilization cost. Last week, we saw people deploy the DeepSeek model on a $6000 local server. On January 31, we even saw people deploy the DeepSeek model on a $2000 local server [3]. The entry point for AI shopping agent builders is getting lower and lower. More people can join to build AI shopping agents, which is exciting news!

Lower cost of utilizing foundation models is great news for AI shopping agent builders. But it is not enough. We have to make a real business impact when we deploy AI shopping agents in the production. If the core value of the AI shopping agent we build is to automate the shopping tasks, the AI shopping agent must save money than the previous non-AI way, i.e., the AI shopping agent must bring positive ROI in order to get wider adoption. If the cost of calling foundation model APIs alone is bigger than hiring a human to do the shopping tasks or using RPA (Robotic Programming Automation) to do the shopping tasks, the AI shopping agent will definitely not be widely adopted. We have to optimize the cost of calling the foundation model APIs. There are many possible ways to optimize the cost of calling the foundation model APIs. Here are three ways. The first is to call cheaper but decent quality foundation models such as DeepSeek-R, which is at least 10x cheaper. The second is to optimize the input tokens before we make API calls[4]. The third is to generate code once for all when we make foundation API calls. Then when we plan to make foundation API calls next time, execute the generated code instead of calling foundation model APIs to fulfill the task[4].

[1] https://www.nubestore.ai/articles/cloudstore-ai-a-tool-for-ai-shopping-agents/
[2] https://darioamodei.com/on-deepseek-and-export-controls
[3] https://www.youtube.com/watch?v=Tq_cmN4j2yY
[4] https://www.zyte.com/webinars/efficient-web-scraping-with-llms/