Learn more about Enabled Intelligence.
By Peter Kant, CEO, Enabled Intelligence
Long before DeepSeek made headlines, Enabled Intelligence understood that super high quality AI data can dramatically cut costs and improve AI performance.
In January 2025, the Chinese start-up DeepSeek sent shockwaves through the AI community – and the stock market – when it announced it had developed a large language model (LLM) that could compete with LLMs from Google, Meta, and OpenAI — but at a fraction of their compute cost.
This was a huge deal because it highlighted how incredibly expensive these state of the art AI models were becoming. Many of the leading LLMs spent tens of millions or even hundreds of millions of dollars simply to train the LLM on all the data needed make accurate inferences. For example, Stanford University estimates Google Gemini spent $191 million in compute costs for model training. And even after they are trained, many of these models need thousands of chips and giant amounts of electricity to operate in vast data centers.
The inherent costliness of frontier AI models like ChatGPT may have been great news for chip manufacturers or data center developers, but it had serious downsides for end users of AI, especially for customers in the Defense and Intelligence Community with limited budgets. Not only were these customers priced out of the market if they wanted to develop an LLM specifically for their requirements, they also could not realistically deploy AI capabilities via edge devices, to the field, to the warfighter.
American Expertise + High Quality Data = AI That Helps the Warfighter
In the last several weeks, experts have begun to unravel DeepSeek’s claims. And the consensus is that while some of the company’s claims are overhyped, a large part of their success is due to something that Enabled Intelligence has known used for a long time: high quality data is the key to lowering AI compute costs, and the key to making AI of practical value to the warfighter. When EI develops a solution to automatically detect threats in satellite imagery, we don’t use low-cost foreign gig workers. We use highly trained US-based geospatial data analysts (many of whom are US veterans) to precisely annotate and prepare the data used in the AI model. We do this because this upfront investment in American expertise reduces downstream errors, makes computation faster and cheaper, and produces tools that reliably help the US military and IC find threat and targets faster and more cost-effectively. It allows our government customers to deploy and use AI technology in the real world, in the field, to solve real problems.
And at the end of the day, isn’t that what matters?
/
