It was recently revealed that Microsoft has been working on its own AI chip called “Athena” since 2019. The initial release will be based off of TSMC’s 5nm process with subsequent generations using the latest TSMC nodes. Athena chips are reportedly designed to specialize in the training of LLMS and related software. However, Microsoft is the slowest in the AI chip race with Google (TPU v4) and Amazon (Trainium and Inferentia) already having their own offerings on the market.
Why AI Accelerator Chips Pose no Threat to NVIDIA
All the hyperscalers are still using NVIDIA’s GPUs for their servers to offer AI or ML-related services. AWS and Google’s AI accelerator chips augment NVIDIA’s GPUs to provide the compute that businesses require for LLM training or AI analytics. Beyond AI workloads, NVIDIA GPUs are still the go-to for VM, data analytics, and high-performance computing use cases
Currently, NVIDIA is still the market leader in AI chips with a 88% market share, according to John Peddie Research. Demand for the A100 and H100 GPUs is exceeding supply as companies require increasing amounts of GPU processing power for LLM training. As the hyperscalers continue to develop and improve their AI chips, no doubt they will be able to wrangle some market share from NVIDIA in the AI space. However, AI chips are ultimately ASICs that are specialized in AI workloads and unable to compete with NVIDIA’s GPUs in any other use cases.
And we don’t expect NVIDIA or AMD to sit back and watch the hyperscalers lead the AI space. It's more than likely that both companies will also increase the performance of their existing GPUs or even launch their own AI ASIC product. We expect competition among the tech giants to increase in the coming years which will ultimately decrease the overall cost of AI workloads and hardware, enabling the emergence of more AI startups and maybe new innovations.
Comments