AI Companies Shift Focus to Efficient Models Running on Fewer Chips

AI Companies Embrace Efficient Models That Run on Fewer Chips

Nearly two months after the viral success of China's DeepSeek prompted a reckoning over how much tech companies spend to develop artificial intelligence systems, some leading AI firms are embracing a less-is-more approach. On Thursday, Toronto-based Cohere Inc. is set to announce a new model called Command A that can carry out complicated business tasks while running on just two of Nvidia Corp.'s AI-focused A100 or H100 chips. That's significantly less than the number of chips required for some large models -- and also less than what DeepSeek's system is thought to need.

AI companies embrace efficient models that run on fewer chips

Nearly two months after the viral success of China's DeepSeek prompted a reckoning over how much tech companies spend to develop artificial intelligence systems, some leading AI firms are embracing a less-is-more approach. On Thursday, Toronto-based Cohere Inc. is set to announce a new model called Command A that can carry out complicated business tasks while running on just two of Nvidia Corp.'s AI-focused A100 or H100 chips. That's significantly less than the number of chips required for some large models -- and also less than what DeepSeek's system is thought to need. A day earlier, Alphabet Inc.'s Google unveiled a new series of its Gemma AI models that it said can run on a single Nvidia H100 chip. Both companies said their models rivaled or outperformed DeepSeek's most recent AI system on certain tasks. AI companies, including Google, continue to invest billions of dollars in data centers, chips and talent to develop their most cutting-edge models and support widespread adoption of them. But the recent launches also hint at an industrywide push to make AI software that can run as efficiently as possible. While some of these projects, including Cohere's new model, were set in motion before DeepSeek's most recent launch, the frenzy around the Chinese upstart may only accelerate this broader effort. In January, DeepSeek released open-source AI software that rivaled models from OpenAI and Google -- and was purportedly built at a fraction of the cost of competitors' models. DeepSeek's team came up with some simple but key innovations, such as finding ways to get more use from the computer chips they had access to.Some AI executives argue the immense attention on DeepSeek is a positive for the industry, indicating it may be cheaper to roll out advanced AI systems than was commonly thought. "I think the service DeepSeek did was make the market aware of how inefficient many players have been," Aidan Gomez, Cohere's co-founder and chief executive officer, said in an interview. "It was a very healthy wakeup call." Cohere said it began work on the new model eight months ago, before DeepSeek upended the tech world. For Cohere, which focuses on business uses of AI and was most recently valued at $5.5 billion, there's an added benefit to this approach. Running AI models on fewer chips is "crucial" for business customers, who may have limited access to computing power. "They don't have tens, let alone hundreds, of GPUs to be able to deploy against problems," Gomez said, referring to graphics processing units. "So they need a very light and scalable form factor."