There are times we must pause and reconsider our assumptions about technology. The internet, the iPhone, the Cloud—each of these transformative innovations challenged conventional wisdom. And each, in its own way, changed everything.
Now it’s DeepSeek, the open-source reasoning model developed by a Chinese AI startup of the same name, that is disrupting the resources required to build and train AI models. The January release of DeepSeek R1 forced many experts to reassess what they knew—or thought they knew.
Some have questioned whether DeepSeek R1 might have been distilled from other models, but that hasn’t muted the enthusiasm. DeepSeek’s surprising efficiency, performance, and bargain-basement cost suggest we don’t necessarily need data centers the size of a football stadium to train AI.
That insight had a near-immediate and sharply downward impact on the Nasdaq and some AI stocks, causing Nvidia’s market cap to plummet by nearly $600 billion in a single day. Yet, the initial panic might have been an overreaction. The second-day angle is that, in many ways, DeepSeek’s breakthroughs bode well for the tech industry at large and for AI training and development.
And why not? The tech sector has always put a premium on resource efficiency, engineering innovation, and performance. And that’s where DeepSeek has changed the game.
Call it the DeepSeek effect: Pilot projects can get underway faster and more affordably. Private language models that combine LLMs with retrieval augmented generation (RAG) become more feasible. Startups and entrepreneurs may find it easier to pursue new ideas.
LOWERING THE BARRIER TO INNOVATION
DeepSeek’s advances are especially relevant for businesses that are downstream from the AI companies developing frontier models. In other words: most everyone. DeepSeek claims to have created R1 at a fraction of the investment—86.6% lower inference costs, by one estimate—that has gone into more mainstream models like OpenAI’s ChatGPT.
Cheaper AI training is a reason for long-term optimism because it lowers one of the primary barriers to entry—the upfront expense. At lower cost, more businesses will be able to integrate large language models (LLMs) and generative AI into their own environments and applications.
In one example among many, AWS is now talking about how to use its Amazon SageMaker AI with DeepSeek for fast-emerging agentic AI use cases. So, DeepSeek may now be used to build next-generation AI agents.
The economies of lower-cost AI models are welcomed by technology vendors, too, because many are also consumers of LLMs. For example, Informatica uses multi-step reasoning models in the development of our AI engine, CLAIRE AI. Anything that lowers our operational costs translates into benefits we can pass along to customers.
PROMISE OF THE OPEN PLAYBOOK
DeepSeek is an open-source model, so it can be modified and redistributed. That’s a net positive because we know from experience that the open philosophy—be it in software, hardware, or computing architectures—enables innovation to thrive through the efforts of like-minded contributors in burgeoning ecosystems.
As an open-source model, DeepSeek R1 can give organizations greater control over their AI projects, such as the ability to fine-tune the model on proprietary datasets. In our case, that means tweaking these models for data management purposes, making them more accurate, and aligning with our strategy of using open small language models and LLMs in our AI-powered technology stack.
The open playbook also means these models may eventually be deployed in more kinds of environments and industries. Not long after DeepSeek R1 was released, BYD, a Chinese manufacturer of electric vehicles, revealed it’s using DeepSeek AI in its driver-assist system. Also, distilled (or scaled-down) versions of DeepSeek R1 can run on edge devices, hinting at new and different LLM deployment models.
Even with so many exciting possibilities, a word of caution is necessary: Businesses should be careful in how they initially proceed with DeepSeek. That’s generally true with any emerging technologies, but especially here given potential privacy and geopolitical security concerns and the prospect of regulation.
For organizations that want to experiment with Deepseek, it’s important to have the right safeguards in place. One way to do that is through collaboration with trusted partners who have strong governance and security practices.
DEMOCRATIZATION OF AI MODELS
Where is all this headed? One possibility is less reliance on resource-hungry foundational models—something sure to appeal to CIOs and CTOs so they can focus more on outcomes than costs.
As you assess and select technology partners, it’s critical to pick a partner that enables you to have flexibility and not get locked in. There are many great LLMs out there, and business and technology decision makers are the ultimate authorities on which are best for their purposes.
But I do believe that the kind of innovation and affordability that DeepSeek represents are promising signs of the direction the industry is heading. Not only is AI adoption poised to accelerate because of this trend, but so are the business benefits.