DeepSeek’s disruption is a delicious thing
What exactly is DeepSeek, the Chinese AI model that has outperformed the best Silicon Valley has to offer and this week created turmoil on the Nasdaq?
Is it a Communist Party psyop designed to shake confidence in US tech supremacy and make Trump think twice about slapping big tariffs on Chinese imports? Has High Flyer, the US$8 billion quantitative hedge fund trading company behind DeepSeek simply “distilled” the best of the US large language models to create a check knock-off? Or is it a genuine breakthrough, a better way of creating generative AI that lowers the barriers to entry and could deflate the AI bubble we are in?
Australia’s top AI expert, Professor Toby Walsh has made up his mind.
“It marks the end of US dominance of the AI race,” he said this week. “It's like running the four-minute mile.”
Even if DeepSeek cost ten times the $6 million sum the company claims it spent developing it, that’s still a dramatic efficiency gain on existing LLM tokenomics. OpenAI charges US$7.50 for one million tokens (750,000 words) for its advanced o1 model. DeepSeek charges 14c to get API access for developers for one million tokens and is free to use via the hosted app. That is a true game-changer and the open source nature of the model means it can be downloaded, modified and built on top of.
It was exhilarating this week to consider the options that open up when this technology is no longer the domain of a small handful of US companies spending hundreds of billions building data centres and nuclear reactors in the race to reign supreme in AI.
It was also a delicious irony when it emerged that OpenAI was investigating whether DeepSeek’s creators had broken the OpenAI terms of use by leaning on data from ChatGPT to train its models.
“This complaint against DeepSeek is not exactly the same as the news publishers, authors, artists, and others who have raged about US AI companies helping themselves to their work, but the premise is similar enough to elicit smirks,” Paul Smith wrote in the Australian Financial Review today.
What is the motivation for going open source with DeepSeek? It’s usually a strategy, as is the case with Meta’s Llama open-source models, to stimulate innovation and a community of users that generates something useful that can be used to make money elsewhere. In Meta’s case, it knew it would struggle to compete with OpenAI/Microsoft so wanted a community of developers to help improve its own models which is using to power AI features on its social media platforms.
High Flyer founder Liang Wenfeng claims he’s just striving for excellence in a geeky field.
“For technologists, having others follow your work gives a great sense of accomplishment,” he said in an interview last year with 36Kr.
“Open source is more of a culture rather than a commercial behaviour, and contributing to it earns us respect.”
Whatever his angle is, DeepSeek’s arrival is good news for New Zealand businesses and our tiny AI scene. Generative AI will likely commoditise faster and cost less as a result of this nimble Chinese competitor’s debut. Other companies will reverse engineer DeepSeek and copy it because let’s face it, no Western company will trust DeepSeek because it is Chinese. President Trump may well try and ban its use in the US.
But in 2025, we’ll likely see a range of new AI services emerge built on top of a new generation of high-performance AI models from many companies that cost much less to use than the existing options. DeepSeek has floated the prospect of doing cutting-edge AI with a lot fewer resources which will help spur more innovation.
With the likes of Microsoft Copilot and ChatGPT Pro costing nearly $40 a month, that’s a very good thing, promising to put powerful AI in the hands of more people. AI agents are being rolled out everywhere. The transaction costs of providing them now stand to be greatly reduced.
It also allows New Zealand to play a meaningful role in AI model development, something that appeared to be off the cards because the computing capacity required was too great.
So credit to DeepSeek, whatever its motivations, for upending the doom-loop narrative of more data centres, more energy, more everything, to feed the AI machine. The Chinese startup has shown that sometimes, less is more.
Check out this great video by Dr Michelle ‘Nanogirl’ Dickinson summing up some of the the things that gave DeepSeek its low-cost advantage.
