It had been believed until very recently that the key to getting better AI systems was to have more compute — more chips would allow companies to build bigger and better models. But DeepSeek’s release changed all that — the Chinese company managed to create a model with reportedly substantially fewer chips than its US counterparts, and used new reasoning techniques to improve their models instead. This caused NVIDIA’s shares to crash — if AI systems could be built without needing as many chips, it meant that companies would need much lesser compute than previously believed.
But Google DeepMind’s Demis Hassabis believes that wouldn’t necessarily be the case. “ Compute is still very much a critical part of what the infrastructure needed for AI, not only for exploring new ideas,” he said in an interview. “If you’re innovating at the frontier, you have to experiment at scale. Otherwise, the results of the experiment don’t necessarily hold at the final training run scale,” he said.
“We’re seeing incredible demand for models like Gemini 2.0 models. And you need a lot of computers. And then the final thing is, with the advent of more thinking models, what’s sometimes called thinking models or inference type models, actually, the more compute time you spend at the point of processing inference, you get more powerful and better answers. So actually, for all of those reasons, you need probably more compute than ever,” he added.
Hassabis seemed to be saying that he didn’t believe the demand for compute would go down. While it might be true that the gains from pre-training — which required large amounts of compute — were slowing down, there would still be large amounts of compute required for inference as AI models got cheaper. Also, the newer styles of AI models, such as DeepSeek R1, use reasoning, which requires much more compute at inference time. As such, Hassabis believes that the demand for compute isn’t going down anytime soon, and this could be good news for the NVIDIA stock which has been languishing since DeepSeek’s release.