Build models in weeks, not months. Use Google’s training stack to speed up development and testing without sacrificing performance.

Develop LLMs 36% faster and squeeze up to 97% productivity (Goodput) out of every accelerator using TPU 8t together with software co-designed with Google DeepMind and integrated with open source frameworks - from Pathways to Pallas (training), and Ray. We also know that one size doesn't fit all, so we partner closely with NVIDIA to deliver the latest GPUs, including A5X instances based on NVIDIA Vera Rubin NVL72. 

Use Gemini platform with BigQuery to train models on proprietary data 16X faster by combining your data estate, ML development and accelerators in one place. Both are powered by AI Hypercomputer, whether you use G4 VMs or Ironwood TPUs. 

Run GPU-based simulations on DeepMind’s MuJoCo-Warp, up to 100X faster than standard MuJoCo. Then simulate impossible, risky, or expensive edge cases using synthetic media from Veo, Genie and Nano Banana, or ingest petabytes of real-world sensor data in BigQuery. Learn more about building physical agents on Google Cloud here.

Get validated model profiles plus fully-integrated Google and open software to boost application responsiveness with less complexity and waste.

Use integrated inference technologies to deliver useful, responsive services to customers. Cut time-to-first-token by 71% with GKE Inference Gateway, serve up to 120k tokens per second using llm-d for disaggregated serving, and load models 5X faster using Rapid Cache and TPU 8i to keep your working memory exactly where it’s needed.

Deploy classical ML models 70% faster using one of 200+ models available on Gemini platform, using on your choice of TPU or GPU, including A5X VMs (NVIDIA Vera Rubin) and TPU 8i when they become available later this year.

Serve swarms of agents securely in GKE Agent Sandbox, provisioning up to 300 sandboxes per second while instantly pausing and resuming as needed, so you never pay for agents sitting idle.

Use any framework or accelerator across hybrid and multicloud environments with automated cluster maintenance and management fit for exascale.

TorchTPU removes the TPU learning curve for developers by providing native PyTorch support, so you can use the best available accelerator without complex code rewrites.

Based on open source Kubernetes, GKE gives you multicloud portability with enterprise scale, supporting up to 130,000 nodes while integrating natively with Gemini platform and Google Distributed Cloud for hybrid deployments.

Every accelerator on AI Hypercomputer is supported by cluster director capabilities, including a pre-deployment bill of health, 360 degree observability dashboards and always-on health checks.

Connect services across clouds without laggy connections using Cross-Cloud Network, a networking backbone trusted by over 65% of the Fortune 100 which moves over 27 exabytes of data per month.

Our flexible consumption models give you multiple ways to schedule and reduce the cost of accelerators. Save up to 91% on batch or fault-tolerant jobs with Spot VMs, up to 50% on jobs with a flexible start date using Dynamic Workload Scheduler, and up to 50% off when you sign up for committed use discounts.

Push the limits of performance and use energy responsibly as you scale on the infrastructure foundation trusted by Google and frontier AI Labs

Google Cloud supports 9 out of 10 top AI labs and 70 percent of funded AI startups. By deploying on AI Hypercomputer, you’re using data centers that reliably processed over 100 billion tokens for nearly 350 customers in December 2025 alone.

Google Cloud’s data centers, including AI Hypercomputer, deliver industry-leading energy efficiency, with six times more computing power per unit of electricity than five years ago. This enables our 8th generation TPU to deliver 80% better price-performance and 20% more energy-efficiency than the previous generation.

We are committed to paying for 100% of the power our data centers use and any new infrastructure costs directly driven by our growth. Partner with us to ensure that as your AI ambitions scale, local households and businesses don’t foot the bill. In the coming years, we will fund new power and infrastructure to serve our models, and continue investing in alternative energy sources like advanced nuclear, geothermal and long-duration storage.

Our Titanium architecture’s custom Titan chips deliver a verifiable hardware root of trust and zero-trust security. Independent analysis from cloudvulndb.org shows that our systems experience up to 70% fewer critical vulnerabilities than other leading clouds.

Analyst insights

Tutorials

Documentation