Google Cloud launches two new AI chips to compete with Nvidia
Google's 8th-gen AI chips split into training and inference models, promising major cost and performance gains.
Google Cloud has unveiled its eighth-generation Tensor Processing Units (TPUs), marking a strategic split into two specialized chips: the TPU 8t for training AI models and the TPU 8i for running inference. The company claims these custom chips deliver substantial improvements, including up to 3x faster training speeds and an 80% better performance-per-dollar ratio compared to previous generations. A key architectural advancement is the ability to link over 1 million TPUs into a single cluster, enabling unprecedented scale for massive AI workloads while aiming to reduce both energy consumption and customer costs.
Despite these advancements, Google is not declaring war on Nvidia. The launch represents a strategy of supplementing, not replacing, the GPU giant's hardware. Google Cloud will continue to offer Nvidia-based systems, including plans to provide Nvidia's upcoming Vera Rubin chip later this year. In a notable collaboration, the two tech giants are even working together to enhance software-based networking technology called Falcon, aiming to make Nvidia systems run more efficiently within Google's cloud infrastructure.
This move mirrors similar efforts by other hyperscalers like Amazon and Microsoft to develop in-house AI silicon. The long-term play is to capture more enterprise AI workloads on their respective clouds, potentially reducing reliance on external chip suppliers over time. However, as analyst Patrick Moorhead noted, past predictions of custom chips dethroning Nvidia have not materialized; the chipmaker's market cap now approaches $5 trillion. For now, Google's growth as an AI cloud provider may ultimately drive more business for all chipmakers, including Nvidia, as the overall market expands.
- Google Cloud launched two new 8th-gen TPUs: the TPU 8t for training and TPU 8i for inference, promising up to 3x faster training and 80% better performance per dollar.
- The chips can scale to over 1 million TPUs in a single cluster, offering massive compute power for large AI models with improved energy efficiency.
- This is a supplement, not a replacement, for Nvidia; Google will still offer Nvidia GPUs and is even collaborating with them to improve networking software (Falcon) in its cloud.
Why It Matters
Cloud customers get more powerful, cost-effective AI compute options, increasing competition and potentially lowering the barrier to training and deploying large models.