The AI boom, which has propelled rapid advancements in large language models, multimodal AI, and agentic systems over the past few years, is now facing one of its most difficult obstacles – a severe shortage of critical hardware components, particularly high-bandwidth memory (HBM) chips and other specialised semiconductors. Demis Hassabis, the Nobel Prize-winning CEO of Google DeepMind and one of the most influential figures in the global AI landscape, has now delivered a candid warning in a recent interview with CNBC.
According to Hassabis, the unprecedented demand for compute and memory resources, fueled by hyperscalers, cloud providers, startups, and enterprises racing to integrate and scale generative AI, has outpaced the industry’s manufacturing and supply capacity, far more dramatically than we know. Even leading players with significant vertical integration, such as Google with its custom Tensor Processing Units (TPUs), remain vulnerable to bottlenecks at key points in the supply chain.
Supply chain “choke points” hamper AI scaling
Hassabis described the situation as a significant “choke point” for AI advancement. “The whole supply chain is kind of strained,” he said, noting that even companies with in-house chip designs like Google’s Tensor Processing Units (TPUs) remain dependent on a limited number of suppliers for key components. “It still in the end actually comes down to a few suppliers of a few of the key components,” Hassabis explained. “So anywhere where there’s some kind of constraint on the capacity, there’s a sort of choke point,” he warned.
He emphasised that the demand for AI infrastructure far outstrips current supply capabilities. Google, in particular, is seeing “so much more demand” for its Gemini models and other AI offerings than it can currently fulfill. “Yes, I think that’s constraining a lot of the deployment for sure,” Hassabis stated. “We see so much more demand… than we can serve.”
The constraints extend beyond commercial rollout to the research domain. “It does constrain a little bit the research,” he added. “You need a lot of chips to be able to experiment on new ideas at a big enough scale that you can actually see if they’re going to work.” This limitation is more nuanced when testing new AI architectures or training runs that require thousands of accelerators operating in concert.
Memory shortages drive up costs and slow progress
The memory chip shortage, particularly high-bandwidth memory (HBM) essential for training and running large language models, has become a broader industry bottleneck. Major players, including Google, Meta, and OpenAI, are all grappling with tight supplies from the handful of dominant manufacturers, such as Samsung, Micron, and SK Hynix. This scarcity has driven up costs dramatically and slowed progress in scaling AI systems to the next generation of capabilities.
Despite these challenges, Hassabis pointed to Google’s relative advantages, including its custom TPU development and long-term supplier relationships, which provide some insulation compared to competitors fully reliant on third-party hardware. However, he acknowledged that physical infrastructure limits are now acting as a real brake on the pace of AI innovation and adoption worldwide.