DeepSeek should make us start again on our AI assumptions

by Kinase Tech Team

DeepSeek has shocked Nvidia shareholders, AI chatbot users, other Chinese AI companies, and now tech opinion writers around the world with its R1 model and App Store conquering app. Why has it caused shock, rather than becoming just another milestone in a year of AI advancements?

Firstly, a lot less investment was needed to develop R1 (potentially only $6M vs hundreds of millions for OpenAI’s o2), and a lot less money is needed to run R1. This is why DeepSeek has disrupted not just Silicon Valley, Nvidia’s share price, and US tech stocks, but also Chinese competitors such as ByteDance and Alibaba who have scrambled to release new models in the last few days (with claims to outperform DeepSeek in their press releases). Alibaba’s cloud unit also announced price cuts of up to 97% on the use of its AI models.

So the impact has been to expose incumbents to a sharp new pressure - in the US and China alike. OpenAI supercharged the current US and global AI fixation with ChatGPT - and have since raised the drawbridge with various paid tiers, while US trade restrictions on Nvidia H100 chips going to China was meant to raise the drawbridge on potential competition by restricting the tech they could use. Both defensive moves have now been challenged - and have in fact created the right environment to be disrupted, by creating the ideal disruptive position to be filled: that of a new, cheaper, fully functional model which could be shown to be ‘just as good’. DeepSeek have now occupied that position with predictable results.

Lee Davies, AI Engineering Director at our partner agency Fresh Egg, notes that there are multiple new opportunities opening up: “DeepSeek is undeniably a very powerful tool, but the big shift here is how it will change access and control over AI resources. The fact that DeepSeek R1 can run on a $6K PC (and in a few years likely a simple laptop), significantly lowers the barrier to entry. Furthermore, being an open-source model (like Meta's Llama), it can be self-hosted and customised better by individuals and organisations.”

DeepSeek’s interruption wipes the slate clean. Perhaps now the AI world of agreed upon benchmarks, open source code and publicly available articles will slowly be forced to close down under pressure from governments - perhaps slowing down AI development into silos. The fact that the power of sophisticated models can be replicated and their costs undercut shows us just how nascent the modern AI phase of development is. This means that the companies and the technology which really will drive a breakthrough is a lot less clear than was thought. Silicon Valley and its investors thought they were surging ahead, building the unassailable foundation for artificial general intelligence. But actually we don’t yet know what the breakthroughs will look like and what size of company will deliver them; how cutting edge their chips will need to be, and what size of investment they will have driving them.

We also don’t know what will characterise the path ahead for AI - years of disruptions and undercutting; a definitive break through that begins a new phase coming soon; or a bubble bursting and an AI winter, indeterminable in length.