The DeepSeek AI now available. (Photo illustration by Anthony Kwan/Getty Images)
A single announcement about theChinese AI DeepSeek wiped out $1 trillion in stock market value this January. But behind the dramatic headlines lies a fundamental movement in artificial intelligence that will reshape how businesses operate.
The rise of DeepSeek illustrates a significant trend in AI’s evolution—not because it matched the capabilities of industry leaders like OpenAI, but because it achieved similar results at a fraction of the cost. For business leaders, this marks AI transforming from an interesting toy to a widely accessible tool.
The cost revolution does not surprise us who have followed technology trends for years. Like many technological advances before it, AI has been steadily becoming more efficient on a quality-adjusted basis. What makes DeepSeek significant is how dramatically it demonstrated this trend to the market. Note the qualifier above: “on a quality-adjusted basis.” A few years ago, high costs led to AI that was interesting but not very good. With even higher costs, the quality improved. Now AI companies will move in two directions. One will be spending more to improve quality, while the other will be spending less with acceptable quality.
The early development of large language models (LLMs) focused on improving quality. Early models were interesting and promising but not particularly useful. Great improvements in quality came from training the models with more data, which was quite expensive.
To see the path for improvement, let’s take a quick excursion into the guts of any artificial intelligence model. There are a huge number of mathematical relationships with adjustable settings called parameters. The model starts with random parameters and automatically tests alternative values for the parameters by evaluating the models answers. This adjustment continues until the model hits a target performance level. It takes a long time because improving one parameter means that all of the other parameters need to be re-evaluated.
When the models were not very good, improving quality was the top goal. Now, however, the models’ quality is high enough to be useful, so other goals are addressed. One goal is cheaper performance. The model could be trained to reduce the number of parameters by giving less attention to minor details and omitting the use of parameters that provide little value to the overall results. In essence, the developer would look for changes that would maintain some target level of quality with fewer calculations. The implications for businesses extend far beyond the immediate stock market reaction.
Another development path underway is reducing hallucinations, the LLM’s tendency to simply make up stuff. When the computer scientists were developing the models, they would ignore the occasional hallucination. Now that people are actually using LLMs for real world tasks, reducing hallucinations has become a high priority.
For an historical analogy, consider automobiles. In the 1890s, just getting a carriage to move without a horse was a great achievement. Then Henry Ford worked on cost. Others later worked on features, then reliability, then again more features. Priorities shift over time. With AI, many of the things that we complain about are issues that developers are working on.
As noted by economist John Cochrane, the real winners won’t be AI producers but rather the businesses that figure out how to leverage increasingly affordable AI. For another historical analogy, he wrote, “The profit, and ultimate benefit, of railroads was not so much in the railroad itself, but in the wheat fields of Kansas.” The benefit of AI will be to businesses in mundane, ordinary sectors of the economy.
For business leaders, this cost reduction creates both opportunities, which will become imperatives. As AI costs fall, we’re likely to see what some call the Jevons Paradox in action: As the cost of using AI drops, total spending on AI will actually increase. Although some writers view this economic idea as a lost relic, anyone with an old textbook on the principles of economics can see that a drop in price causes an increase in spending when the elasticity of demand is greater than one. That is, when the quantity demanded is relatively sensitive to the price.
As individual AI operations become cheaper, businesses will find more and more applications for the technology, potentially increasing their overall AI expenditures while getting significantly more value from each dollar spent.
DeepSeek is not the cause of change, but an illustration. Businesses need to start preparing for a world where AI continues to improve with lower costs. The competitive advantage will shift from simply having access to AI to using it creatively and effectively. Smart executives will start identifying areas where previously cost-prohibitive AI applications might soon become viable, and begin developing strategies to implement them.
