Google Cloud expands TPU v5p availability, posts on-demand pricing and MLPerf gains for training and inference

Google Cloud said Tuesday it has expanded availability of its TPU v5p accelerators to additional regions and posted on-demand pricing along with new MLPerf submissions, aiming to accelerate training and inference for large language models in managed environments. The company added TPU v5p in us-central1, europe-west4, and asia-east1, with support through Vertex AI and GKE, positioning the chips for customers scaling beyond 70 billion parameters.

The timing underscores intensifying competition for enterprise AI workloads as capacity constraints ease and buyers press for clearer price performance. Google is leaning on published benchmarks and per-chip-hour pricing to court customers that have split spend across Nvidia H100 fleets and emerging AMD MI300X offerings on rival clouds.

Google listed TPU v5p on-demand at the chip level with sustained use discounts through committed contracts, and said customers can mix preemptible and reserved capacity for batch training. Vertex AI also added managed checkpointing and sharded dataset ingest for v5p pods, along with Feature Store and BigQuery direct connectors to cut data transfer overhead. Google did not disclose new committed use rates in the announcement, but said they are available through sales representatives.

Google’s latest MLPerf results showed TPU v5p training and inference scores competitive with Nvidia H100 class systems across BERT and GPT style workloads, according to MLCommons published reports. The company highlighted throughput and time to convergence improvements versus TPU v4, citing double digit percentage gains on language and vision tasks, while acknowledging that closed division results vary by vendor configuration and software stack.

Enterprises using Vertex AI can now target TPU v5p for managed training jobs and then export artifacts to Vertex Model Garden or deploy via online prediction endpoints with autoscaling. Google added token level logging, model version pinning, and private service connect for inference endpoints, aiming to address procurement demands around observability, rollback, and isolation.

The move comes as cloud providers scale heterogeneous accelerators and publish clearer economics. AWS has grown managed model serving on Bedrock and offers Nvidia H100 and AMD MI300X instances for custom stacks. Microsoft has expanded Azure ND series for H100 and MI300X and continues to integrate OpenAI models in Azure OpenAI Service. Buyers are increasingly mixing managed services with self-hosted options to balance cost, latency, and data residency across regions.