Google Splits TPU Line Into Training and Inference Chips as China Tightens Domestic Stack

Google announced Wednesday it will divide its tensor processing unit architecture into separate chips for training and inference, ending years of dual-purpose design as the company seeks to challenge Nvidia's dominance in specialized AI workloads. The eighth-generation TPU will debut later this year in two variants, with the training chip delivering 2.8 times the performance of November's seventh-generation Ironwood processor at the same price, while the inference model offers an 80 percent performance gain.

The move reflects a broader industry shift toward task-specific silicon as AI agents and real-time applications demand faster response times. Amin Vahdat, Google's senior vice president and chief technologist for AI and infrastructure, said in a blog post that the company determined "the community would benefit from chips individually specialized to the needs of training and serving" as AI agents proliferate.

Google's new inference chip, the TPU 8i, relies on 384 megabytes of static random-access memory per processor, triple the amount in Ironwood. The design choice mirrors technology used by Cerebras, an AI chipmaker that filed to go public earlier this month, and aligns with Nvidia's upcoming Groq 3 LPU hardware, which the company acquired through a $20 billion deal with chip startup Groq in March.

(Google remains a major Nvidia customer but positions TPUs as an alternative for enterprises using its cloud services. The company did not benchmark its new chips against Nvidia's offerings.)

The announcement arrived the same week Chinese AI startup DeepSeek launched its V4 model with full support for Huawei's Ascend chip line, marking a strategic pivot from the Nvidia processors the Hangzhou-based company previously relied on. Huawei confirmed its AI supernode systems will support the new model across its entire high-performance product line, though industry observers expect the rollout to scale more effectively later this year as newer Ascend chips reach volume production.

The DeepSeek-Huawei collaboration underscores China's accelerating push for technological self-sufficiency in AI infrastructure. Washington began restricting China's access to advanced AI chips made by U.S. firms in 2022, prompting Beijing to prioritize domestic alternatives. DeepSeek has acknowledged using Nvidia chips in the past but has not disclosed which processors trained its latest model. A White House memo seen by the Financial Times reportedly accuses China of engaging in "deliberate industrial-scale campaigns" to replicate U.S. frontier AI systems through data distillation, claims the Chinese Embassy in Washington dismissed as "baseless allegations" and "pure slander."

Most major technology companies now pursue custom semiconductor development for AI, seeking efficiency gains and specialized capabilities. Apple has embedded neural engine components in iPhone chips for years. Microsoft announced a second-generation AI chip in January. Meta said last week it is working with Broadcom to develop multiple versions of AI processors and this week committed $5 billion to Anthropic, partly to purchase Amazon's AI chips. None have displaced Nvidia, which continues to dominate the market for high-performance AI accelerators.

The parallel hardware announcements from Google and the DeepSeek-Huawei partnership illustrate diverging approaches to AI infrastructure. Western companies are fragmenting workloads across specialized processors to maximize performance and cost efficiency within existing supply chains. China is vertically integrating domestic chips and models to create a localized stack insulated from geopolitical disruption. Both strategies respond to the same constraint: growing demand for AI compute is outpacing available capacity, driving companies to optimize every layer of the technology stack.