Google Cloud Debuts Agentic AI Stack and New TPUs

Google Cloud used Cloud Next in Las Vegas to showcase a broad set of product and partnership announcements focused on agentic AI, new silicon, and enterprise security. Per TechCrunch, Google announced two new eighth-generation tensor processing units, the TPU 8t (training) and the TPU 8i (inference), with vendor-reported performance improvements including up to 3x faster training, 80% better performance per dollar, and the ability to scale to 1 million+ TPUs in a single cluster (TechCrunch). The Google Cloud blog published posts describing new agent-focused capabilities in its platform, plus security partnerships and features such as agent identities, Model Armor, and integrations with Wiz for an AI Bill of Materials (cloud.google.com).

Per Google Cloud messaging covered in company blog posts, the platform updates combine compute (the new TPUs), model and agent orchestration (a redesigned Vertex/agent surface), and security tooling (Model Armor, Agent Gateway integrations). Reuters and Fortune coverage highlight that the product announcements were accompanied by enterprise commercial activity: Reuters reports that Merck will invest up to $1 billion over multiple years to adopt Google Cloud infrastructure and Gemini Enterprise tooling, with Merck chief information and digital officer Dave Williams quoted saying, "I easily see us investing a billion over the next several years in this, in those capabilities," (Reuters).

Industry reporting frames Cloud Next as another signal that hyperscalers are pushing vertically integrated AI stacks that span chips, models, orchestration, and security. Fortune notes recent Google Cloud growth metrics including a reported 48% Q4 revenue increase to $17.7 billion and a cloud backlog cited at $240 billion by the end of 2025, situating these product pushes inside an enterprise growth narrative (Fortune). TechCrunch and The Next Web place the TPU announcements and agent tooling in the ongoing competition between hyperscalers to reduce per-workload cost and to offer differentiated managed silicon.

Google Cloud's security announcements-documented on the company blog-emphasize agent-aware defenses and partnerships with Wiz to cover AI-specific risks (cloud.google.com). Separately, reporting by The Register describes an incident where an unauthorized party accessed Anthropic's Mythos model quickly, and Google Cloud signals limited/private preview access for Anthropic's Mythos via its platform in some channels (The Register; cloud.google.com snippet). Those are separate factual items reported by the respective outlets.

Industry observers will monitor vendor claims versus independent benchmarks for the TPU 8t and TPU 8i, adoption traction for the agentic surface and Gemini Enterprise in regulated industries, and the operational security posture around third-party model integrations such as Mythos. Reporting to date includes vendor performance claims (TechCrunch) and enterprise commitments (Reuters) but not independent third-party performance validation.

Adoption decisions should be informed by independent benchmarking of new TPU pricing and throughput, scrutiny of agent orchestration APIs and governance controls, and attention to security integrations and supply-chain visibility cited in Google's blog posts. Observers may also track how enterprises contract for compute and managed models given deals like the Reuters-reported Merck agreement.