What’s changing the rules of enterprise AI adoption for IT leaders

DeepSeek’s launch of its R-1 AI model in January has clearly upset the AI apple cart. More than $1 trillion was wiped off US technology stocks with frontier model developers such as OpenAI, Alphabet, and Meta caught off guard by this Chinese startup. DeepSeek claims its reasoning model is comparable in performance to OpenAI’s o1 offering, working twice as fast but at only 10% of the cost.

Despite controversy about the model’s development, DeepSeek has greatly accelerated the commoditization of AI models. This is good news and will drive innovation, particularly for enterprise software developers. For CIOs, this presents greater choice of AI products from vendors to consider and evaluate, as well as lower barriers to build AI capabilities inhouse.

Cheaper, better, faster

The story of digital innovation over the last 50 years is one of more powerful systems and lower costs. Moore’s Law has held true over this period for microprocessors, and similar patterns can be observed for other core elements of the computing revolution. For the last decade, AI also seems to have followed this trend with rapid acceleration since the launch of ChatGPT in 2022. According to tech entrepreneur and AI expert Azeem Azhar, OpenAI’s GPT-4, which launched in March 2023, cost $36 per million tokens to run while DeepSeek offers similar performance for $0.14 — 250 times cheaper.

The proliferation of open-source AI models — more than 1 million are currently listed on the Hugging Face portal — is driving innovation particularly at the application end. DeepSeek has simply ratcheted up this trend an order of magnitude. We can expect attention to shift this year from model developers to those building business applications harnessing this low-cost environment for innovation.

Unlike Microsoft and the PC operating system, Google search, and Meta’s social media, it doesn’t look like any single model developer is going to have a controlling interest in shaping the evolution of AI in the near-term. Co-founder and CEO of IDPartner Systems Rod Boothby sees OpenAI as the dominant model developer so far, offering only unit innovations, whereas Microsoft, Meta, and Google have system innovations that exploit network effects to benefit all system users. Software developers and users have benefited from a stable and widely used PC operating system; website operators can optimize their sites for the dominant search engine; and Google, Instagram, and Facebook users can connect to and follow millions of other subscribers. “Until and unless OpenAI creates a forum where people can interact and profit, it’ll likely remain one model supplier among many, which can easily be replaced,” says Boothby.

Building for the enterprise

As model costs fall and the value from AI migrates up to the application layer, enterprises are going to have even greater choice in business solutions, either from third parties or those developed inhouse. For CIOs with access to the right resources, building applications internally is now a more realistic proposition. This becomes increasingly attractive in the context of complex business processes that may be unique to enterprises. As the costs of running models fall to near zero, the ROI equation shifts dramatically. According to Forrester Research, the ability to run hyper-efficient models like DeepSeek locally on PCs opens up a new era of edge intelligence, which businesses can deploy across organizations.

“The real value in AI isn’t just in building bigger models, but innovating on top of them and in implementing them efficiently,” says Devesh Mishra, president of CoreAI at digital transformation specialists Keystone. “Companies that pair foundation model advancements with deep business and operational expertise will lead the next phase of AI-driven ROI.”

This deep understanding of industry verticals and their specific issues and needs will define success for many vendors as they increasingly compete with inhouse development teams. The potential of AI to transform many business processes makes many off-the-shelf solutions irrelevant.

This will be particularly true for AI agents, according to Azhar, who argues that when a system such as DeepSeek can process 250 tokens per second, real-time AI is a real possibility. In that environment, AI agents would be able to communicate with each other 50 to 100 times faster than humans. The potential for massive efficiencies and cost savings will put greater pressure, and risk, on CIOs to evaluate this technology across all business functions and in conjunction with suppliers and customers. Gartner’s AI roadmap is a useful starting point as it brings together technical, human, data, and strategic factors to help leaders begin this process.

Over a decade ago Marc Andreessen, partner at VC firm Andreessen Horowitz and co-founder of Netscape, said software was eating the world, in the context of digital transformation. The extent to which many business processes and transactions are now handled by software, often in the background, supports his hypothesis. However, progress has been slower than many predicted, Andreessen included. The rise of AI in its different forms over the last several years coupled with the dramatic launch of DeepSeek’s R-1 model promises a major shift in this journey.