DeepSeek is genuinely commoditising GenAI. Delivering a useful resource at negligible or low cost is the quickest way to make it ubiquitous, and change consumer behaviours. Amazon is still one of the best examples of that. DeepSeek states that its earlier V3 model cost only $5.6 million to build – a small fraction of what OpenAI spent on its GPT models – although this may not include its total development and training costs. And OpenAI claims to have evidence that GPT-4 outputs were used to develop DeepSeek models, which is prohibited in its terms of service. Whether or not that can be proved, DeepSeek was only founded in 2023, so its developers will have been able to reverse engineer and build on OpenAI’s various iterations – illustrating that first mover advantage doesn’t last long in the world of GenAI, which is a lesson for legal too. 

The knock-on effect was almost instant. The emergence of a Chinese low-cost, open source model that was built without cutting edge chips rattled the market and chip maker Nvidia’s shares dropped 17%, losing $589 billion in value in a single day. However, they recovered by 9% the following day. Microsoft shares slipped 4.5% based on disappointing cloud computing forecasts and investor concerns about cheaper competition from China. At one point more than $1 trillion was wiped off Nasdaq as the market reacted to the activities of what is (coincidentally?) a hedge-fund owned company. The relevance to lawtech is that investors are significant market influencers because their confidence, or otherwise, determines value.

While there has been a lot of speculation about DeepSeek’s direct impact on legal, obviously nobody’s using it yet, because it’s brand new. OpenAI had a three-year head start, and while law firms have been quick to take-up GenAI, this is partly because OpenAI’s GPT models are integrated into Microsoft 365 and familiar legaltech platforms, and law firms use them to build bespoke internal tools that keep data ring-fenced and secure behind their firewall. There are a few LLM agnostic legal platforms, which may integrate DeepSeek and Alibaba too. And although concerns have already been expressed about sending data to China, as Computing editor Tom Allen explained succinctly, you can get around that by downloading a copy of the entire model and running it locally – i.e. inside the firewall. The same risk management and data security and privacy considerations apply to all hosted systems and tools. 

One big potential advantage of open source/low cost GenAI models for lawtech – and lawtech start-ups – is that they open up opportunities in the consumer space for AI to support consumer legal services and access to justice without raking up huge costs. Cost-effective AI can also support the justice system, improving efficiency and reducing backlogs. Here too, China is in the vanguard. AI systems are already used in Chinese courts to streamline judicial processes so that the courts can handle more cases.