Cloud giants intensify competition in self-developed chip, Alphabet Inc. Class C (GOOG.US, GOOGL.US) Ironwood challenges NVIDIA Corporation's GPU dominance.

Alphabet Inc. Class C's parent company Alphabet (GOOG.US, GOOGL.US) announced on Thursday that its seventh-generation proprietary artificial intelligence accelerator chip TPU Ironwood will be made available to businesses and developers in the coming weeks. The chip was first unveiled in April this year, has completed testing and is now being gradually deployed. Alphabet Inc. Class C stated that Ironwood is designed for the most complex and intensive AI operations, capable of supporting tasks such as training large language models, complex reinforcement learning (RL), high throughput real-time inference, and more. Amin Vahdat, Head of Cloud AI and Infrastructure at Alphabet Inc. Class C, and Mark Lohmeyer, Head of Computing Platforms, stated in a blog post that Ironwood achieves a 10x increase in peak performance compared to the previous TPU v5p, and over 4x improvement in single-chip performance efficiency in training and inference scenarios compared to TPU v6e (Trillium), making it the "most powerful and energy-efficient custom chip" from Alphabet Inc. Class C to date. Additionally, Ironwood supports the formation of super-pod clusters with up to 9,216 interconnected chips, significantly increasing data throughput and training speed, helping businesses overcome bandwidth and latency bottlenecks during large model training. In terms of practical applications, Alphabet Inc. Class C revealed that AI large model company Anthropic (developer of the Claude model) plans to use as many as 1 million Ironwood TPUs to support its next-generation model training and inference needs. Anthropic currently receives investments from both Alphabet Inc. Class C and Amazon.com, Inc. (AMZN.US), making it an important partner to measure the effectiveness of Alphabet Inc. Class C's chip strategy. Meanwhile, Alphabet Inc. Class C also announced further expansion of their in-house Arm architecture CPU product line Axion, and the launch of several new instances targeting the cloud computing market. Among them, the N4A virtual machine is now in preview, with Alphabet Inc. Class C claiming that its price performance is twice that of similar x86 virtual machines; another C4A bare metal server will soon be in preview, primarily used for Android development, in-vehicle systems, strict authorization software, high-density testing, and complex simulation and other professional scenarios. Industry analysts point out that as the demand for large language models and AI training increases, major cloud providers are accelerating their move away from reliance on NVIDIA Corporation GPUs to develop their own chip systems, in order to reduce costs and enhance computational power supply stability. Media commentators state that with the introduction of their proprietary Trainium/Inferentia chips by Amazon.com, Inc. (AMZN.US) AWS and the development of Maia GPUs by Microsoft Corporation (MSFT.US), Alphabet Inc. Class C's launch of Ironwood signifies that the "cloud computing giants have officially entered the deep waters of in-house silicon chip competition."