Joachim Klement, managing director at Panmure Liberum, argues that the future of artificial intelligence (AI) may not lie in large data centers but rather in smaller language models that can operate on consumer hardware.
He cites research from Stanford University indicating that these smaller models can effectively handle over 80% of typical AI tasks, suggesting a significant shift in how AI workloads could be managed.
This transition could lead to hyperscalers, or large data center operators, facing excess capacity as value migrates towards device manufacturers and edge-chip suppliers, which are often overlooked in the AI landscape. Klement identifies Apple and Dell as potential winners in this scenario, despite their current perception as AI laggards.
He emphasizes that the economics of running AI models locally are substantially more favorable, with costs for local systems being 70-80% cheaper than those in data centers. This shift is already influencing corporate strategies, as companies like AT&T and JPMorgan adapt their AI deployments based on cost and performance.
Klement also highlights Nvidia's recent launch of its desktop AI solution, DGX Spark, as a response to potential slowdowns in data center demand. He predicts that chip designers focused on edge computing, such as Arm and Qualcomm, will benefit from this trend, while memory manufacturers will remain stable due to ongoing demand for local AI systems.
Overall, Klement believes that the industry may be overbuilding data centers and that the future of AI could increasingly favor local computing solutions