Microsoft's CTO Aims to Decouple AI Workloads from NVIDIA and AMD GPUs

Microsoft Aims to Ditch NVIDIA and AMD GPUs for In-House Chips In a significant shift in its datacenter strategy, Microsoft's CTO Kevin Scott revealed plans to transition the majority of its AI workloads from Graphics Processing Units (GPUs) to custom-built accelerators. Speaking at a fireside chat on Wednesday, Scott emphasized the importance of performance per dollar for hyperscale cloud providers like Microsoft. Timeline and Details Microsoft has been purchasing large quantities of GPUs from both NVIDIA and AMD, but Redmond's leaders now want to switch to in-house chips for most datacenter workloads. The transition is driven by a focus on cost-effectiveness, as Scott acknowledged that NVIDIA had offered the best price-performance up until this point. Microsoft is reportedly developing its second-generation Maia accelerator, which will be released next year and promises improved compute, memory, and interconnect performance. Immediate Impact The implications of this shift are significant, as it could lead to increased efficiency and reduced costs for Microsoft's datacenter operations. However, the transition may also raise concerns about vendor lock-in and the potential for reduced innovation in the GPU market. Background Context Microsoft has been investing heavily in custom chip development, with its Cobalt CPU being one of several projects underway. The company's focus on in-house acceleration is part of a broader trend towards specialized computing hardware, driven by the growing demand for AI and machine learning workloads. What Happens Next As Microsoft continues to develop its custom accelerators, we can expect to see increased competition in the datacenter market. Other cloud providers may follow suit, leading to a shift away from traditional GPU vendors like NVIDIA and AMD. The long-term implications of this trend will be worth watching, as it could have far-reaching consequences for the tech industry and beyond. *This story is developing. Information compiled from Hardware reporting.*