Memory chips become a bottleneck in AI development

In an interview with CNBCHassabis acknowledges that physical challenges are holding back the widespread deployment of AI models. Large companies like Google are also being hit hard as demand for Gemini and other language models is outpacing the capacity of existing infrastructure.

“You need a huge amount of chips to test new ideas on a large enough scale to determine whether they really work or not,” Google DeepMind leader said. “The whole supply chain is very stressed.”

In the field of AI, memory chips are not simply a component but also become a strategic weapon. Mark Zuckerberg, CEO of Meta, once revealed that in addition to money, what AI researchers crave is less administrative procedures and as many chips as possible.

Even though it has the advantage of designing its own Tensor processor (TPU) to serve internal needs and rent it via the cloud platform, Google still cannot be completely self-sufficient. The key components to create a complete system still depend on a small number of partners in the market.

The memory chip market is currently dominated by the power trio Samsung, Micron and SK Hynix. These manufacturers are also trying their best to serve the huge demand from AI giants while not abandoning traditional customers in the consumer electronics segment such as laptops, gaming devices, routers…

The difficulty lies in the fact that AI companies are mainly hunting for high-bandwidth memory chips (HBM), a variant that is much more expensive and difficult to produce than memory chips used for regular personal computers. This gap in supply and demand has pushed costs to record highs, causing the prices of consumer electronics products to increase and research progress to be delayed.

Despite supply chain barriers, Google shows that it has no intention of slowing down in the race to equip technology. In its fourth quarter 2025 financial report, Google expects capital spending on infrastructure and AI chips to be about $175-185 billion in 2026.

The statement of Google DeepMind’s leader once again shows that in the AI ​​era, the winner not only has the smartest algorithm but also must hold the most hardware in their hands.