After decades of slow but steady progress, the pace of advancement in artificial intelligence (AI) technologies continues to astound everyone—from technologists themselves to the organizations, end users, and regulators who use and monitor these technologies. In early 2025, the AI landscape is seemingly undergoing yet another dramatic shock with the launch of China’s RI large language model (LLM) from DeepSeek. The launch is dramatically upsetting perceived principals of what it takes to build the latest and most advanced ‘frontier level’ LLMs—and roiling markets at the time of writing.
For banks and other financial service providers, these moves underline the need to ensure multimodel flexibility in the AI functionality they develop. They also highlight the extreme pace of ongoing AI advancement. Low-cost, open-source models like DeepSeek suggest that what may be seen as the obstacles of today may be solved much faster than initially expected.
Investment in AI infrastructure has continued to increase over the past several years. Hundreds of billions of dollars are expected to go into building out new data centers and even firing up alternative energy sources (including recommissioning Three Mile Island) to power the huge energy needs of the latest AI capabilities. This includes most recently the announced US$500 billion project Stargate investment fund with OpenAI, Oracle, and Softbank that has already proven controversial with questions about overall funding availability, and the fact it is intended to serve only OpenAI’s needs.
The scale of this investment can be difficult to imagine for many. To illustrate, Project Stargate alone holds a proposed investment amount higher than the estimated US$328 billion (in 2023 dollars) spent on the U.S. space race between the launch of Sputnik and the final Apollo mission. Project Stargate, however, is only one of many investment roadmaps from the so-called “Magnificent Seven” of AI technologies, which includes Meta, Microsoft, Oracle, Amazon, Alphabet, Tesla, and Nvidia. Many investors are now questioning the sustainability of the hundreds of billions being poured into annual CapEx on AI infrastructure.
The sheer growth and importance of AI technologies have also become a geopolitical issue, with multiple maneuverings to restrict China and other perceived adversaries’ access to the latest hardware and software technologies and help to ensure U.S. dominance of the burgeoning AI space. At this point, the geopolitics of AI are only likely to become hotter.
Seemingly out of nowhere, Chinese developer DeepSeek launched its latest publicly available LLM, DeepSeek V3, followed by its RI model. Crucially, DeepSeek’s open-source model was trained in only two months and developed for a staggeringly low cost of US$6 million (less than the salaries of some prominent AI executives). Moreover, it uses fewer graphics processing units for training.
DeepSeek RI, according to numerous tests, beats OpenAI, Google, and other leading models in terms of capabilities, a staggering achievement given its low cost and rapid training time. Perhaps most alarmingly for existing LLM providers, DeepSeek can sell tokens for use at a price point 30 times lower than OpenAI, significantly undercutting the perceived economics of these leading AI LLM vendors. DeepSeek also claims to be able to offer a desktop-capable version of its model that can operate locally with no need for the cloud at a scale of only 10 gigabytes.
DeepSeek has emerged from seemingly nowhere to upend the AI market and, in its announcements, declares that AI was merely a side project. (It is primarily a quantum-focused organization.) Alongside this being a Chinese organization, it presumably lacked access to the latest in chip technology. Its success has been built largely through top-notch and innovative design with serious implications for the scalability of AI and benefits for the vast energy consumption of genAI technologies.
The reality of DeepSeek’s Chinese origin suggests there is limited likelihood of any banks or financial institutions deploying DeepSeek R1 anytime soon. It does, however, underline the potential for significant innovation in the AI space from unexpected sources and the ongoing potential for further development of open-source LLM models. AI that is smaller and cheaper to develop, with less intensive energy consumption, will accelerate the already booming AI market. With deregulation now occurring with the change in U.S. administration the potential for significant shifts in the AI landscape must be top of mind for all banks and FIs.
The shock of DeepSeek has real potential to upend the AI space and highlights the need for banks to develop capabilities that can be flexible to their underlying LLM models. Tools such as Amazon Bedrock that are model agnostic can help FIs, vendors and developers reduce the risk of backing the wrong model or missing out on the latest technology development. Tying into one specific LLM vendor increasingly looks like an untenable approach as banks and FIs will need to balance the need to use the latest, most effective technology, especially if it’s cheaper and more efficient, against the real need to be wary of LLM developers themselves becoming unstable or losing funding.
Banks mustn’t lose sight of the potential for genAI even if the investment bubble in AI now deflates due to these DeepSeek-driven economic shocks. Significantly lower costs and efficiencies in building and developing these AI tools will, in all likelihood, only grow usage and the potential for further innovation. AI technology cannot be uninvented, and cheaper, more efficient LLMs dramatically increase the potential for rapid innovation, if not revolution.
