Google split its eighth-generation TPU into a training chip (8t) and an inference chip (8i). Nvidia lists the Groq 3 LPX inference accelerator in full production and introduced Vera, a CPU it describes as built for AI agents. Alibaba's T-Head introduced the Zhenwu M890 for both training and inference and announced the V900.
Specialization follows workloads: training is bursty and compute-heavy, inference is continuous and latency-sensitive, and agents add orchestration that suits CPUs. The competitive implication is that customers can mix suppliers by job, which weakens single-vendor lock-in. Performance claims are the vendors' own.
Source: CNBC