Since the introduction of its latest model, Chinese startup DeepSeek has taken the world of artificial intelligence by storm.
Its innovations could have finance leaders rethinking how much they need to spend on AI tools and development, one expert told CFO Brew.
“I think the huge revolution is not how good it is and how good the outputs are…but just the dramatic cost reduction because of how they use a different kind of architecture,” Conor Grennan, chief AI architect at the NYU Stern School of Business and founder of consulting firm AI Mindset, told us.
Major disruptor. DeepSeek claims to have made its large language model (LLM) for dirt cheap relative to its competitors. The startup wrote in a technical report that it only cost about $5.6 million to train its model.
But that number doesn’t tell the whole picture. DeepSeek noted in the report that its estimate leaves out research and experimentation costs. Some estimate that DeepSeek’s costs for hardware alone likely exceeded $500 million, CNBC reported. Also, OpenAI said it’s investigating whether DeepSeek “inappropriately” trained its model on data generated by ChatGPT, NBC News reported.
However, it does offer some advances. Grennan said DeepSeek is a cheaper alternative to other AI models because it uses a machine-learning technique called mixture of experts (MoE). What does that mean? Simply put, DeepSeek uses only the areas of its memory banks it needs to solve a question.
Using MoE “is fascinating but also technically hard to do, which is why more organizations don’t use it,” Grennan said. “They’ve kind of cracked the code on how to do that.”
And as Grennan pointed out in a LinkedIn post, DeepSeek using less computing power to solve problems means it uses less energy. That’s a good thing for pocketbooks and the planet.
But it doesn’t entirely solve AI’s energy problem, Grennan told us. He said it’s likely that as AI becomes leaner and cheaper, “the more expansively it will be used,” meaning “we’re going to need those energy solutions more than ever.”
David and Goliath(s). OK, so DeepSeek is cheaper and it has some neat innovations. But is it really a good substitute for leading US-based AI technologies? Grennan suggested looking at what the biggest thinkers in AI said about it.
News built for finance pros
CFO Brew helps finance pros navigate their roles with insights into risk management, compliance, and strategy through our newsletter, virtual events, and digital guides.
In a post on X, OpenAI CEO Sam Altman wrote DeepSeek has “an impressive model, particularly around what they’re able to deliver for the price.” Dario Amodei, CEO of Anthropic, wrote in a blog post that DeepSeek “appears to come close to the performance of state of the art US models on some important tasks.” He also gave kudos to DeepSeek’s developers for some “genuine and impressive innovations.” DeepSeek held its own when Ars Technica pitted it against OpenAI’s $20-per-month and $200-per-month models in a series of test prompts.
There’s an important caveat with DeepSeek. It follows Chinese censorship regulations, meaning that users can’t rely on it to answer certain questions about the country and its history.
Decisions, decisions. Censorship issues aside, it’s clear that DeepSeek is legit. But what does all this mean for business?
In an interview with CNBC, SAP CFO Dominik Asam said the introduction of DeepSeek is “very good news for us” and that “competition [and] innovation on that front is extremely helpful for creating better products for our customers.”
“What we focus on is really leveraging the context of customer problems we want to solve,” Asam said. “It’s about the data, connecting the data to these foundation models, and we try to petition the software solution in a way that if innovation comes we can quickly capitalize on it and basically upgrade our systems. The more supply, the more choice we can use on our AI hub, the better for us.”
Finance leaders need to ask what kind of AI capabilities their organization needs, then determine which offering best suits them, Grennan said.
“If you have a business that’s answering questions on a chatbot about a product or helping people write copy or brainstorm…then do you really need the Ferrari of large language models? I don’t think so,” he said.
Forward looking. DeepSeek’s model has greater long-term implications than simply being a cheap AI tool. Grennan said he liked the descriptor “Sputnik moment” to encapsulate what it will do to the AI industry.
“I think that everybody’s going to get more efficient very, very quickly,” he said. “We’ve already seen that, if we look back a year, year and a half, it’s way cheaper even now, and the models are more powerful.”