Google Cloud Next '26 ran for three days, and the biggest signal boiled down to one sentence: Vertex AI is no more; the new name is Gemini Enterprise Agent Platform.
It may sound like just a rename, but Google has reorganized its entire product line, hardware roadmap, and even the way it sells cloud services around the core concept of "agents."
What Vertex AI has become
Gemini Enterprise Agent Platform is not a simple reskin of Vertex AI; it has been rewritten as a full-stack platform specifically designed to run agents.
It integrates more than 200 models, including Google's own Gemini 3.1 Pro, Flash Image, Lyria 3, and Gemma 4, and also directly hosts Anthropic's Claude Opus, Sonnet, and Haiku. This is a subtle but significant move: Google has invested $40 billion in Anthropic while positioning Claude as an equal option alongside its own Gemini in an enterprise product.
According to Google, the platform's components cover the entire agent lifecycle:
- Agent Studio — Development environment
- Agent Development Kit — SDK
- Agent Runtime — Runtime
- Agent-to-Agent Orchestration — Multi-agent collaboration
- Agent Gateway — Traffic and access control
- Agent Identity — Identity authentication
- Agent Registry — Registry
- Agent Observability — Observability
- Agent Simulation / Evaluation — Simulation and evaluation
Those familiar with this terminology will recognize that Google has essentially ported the last decade's cloud-native stack (Service Mesh, API Gateway, Identity, etc.) directly to agents. In short: treating agents as a new form of service to manage.
TPU is now split into two distinct lines
On the hardware side, the eighth-generation TPU was officially announced, but instead of a single "general-purpose" chip like its predecessors, training and inference have been split into two product lines.
TPU 8t (Training):
- Single superpod scales to 9,600 chips
- Shared memory 2 PB
- Compute power 121 ExaFlops
- Per-pod compute is nearly 3x the previous generation
- Storage access speed is 10x the previous generation
TPU 8i (Inference):
- On-chip SRAM tripled
- Interconnect bandwidth doubled to 19.2 Tb/s
- 80% improvement in performance per dollar
The reasoning behind this split is clear: the hardware requirements for training and inference have diverged too much. Training demands massive shared memory for distributed optimization, while inference requires low latency, high interconnect bandwidth, and large on-chip caches. Using a single chip to satisfy both is no longer cost-effective.
Nvidia has followed a similar path (the divergence between H100/H200 and L40S), but Google has baked this distinction directly into the TPU generation naming, signaling a more decisive stance.
Redesigning the entire data center
Supporting the new TPU is a new data center architecture called Virgo Network:
- A single network connects 134,000 TPU 8t chips
- Non-blocking bandwidth 47 Pb/s
- Bandwidth available per TPU 8t is 4x the previous generation
- Unloaded latency reduced by 40%
To put 47 Pb/s in perspective: it's roughly equivalent to cramming the total bandwidth of the global internet's core nodes into a single data center.
Software layer continues to cut costs
Hardware alone isn't enough; the software layer also made significant cuts to operating costs:
- GKE Agent Sandbox: Launches 300 sandboxes per second, 30% improvement in performance per dollar
- BigQuery fluid scaling: Average cost reduction of 34%
- Lightning Engine for Apache Spark: Doubles performance per dollar
- Managed Lustre: Throughput up to 10 TB/s
- M4N virtual machine series: Running Oracle workloads, TCO reduced by over 20%
These data and orchestration layer moves target enterprise workloads that couldn't afford Vertex AI but still need to do AI.
How big is this bet?
Google Cloud Next's past few keynotes have always featured "AI," but this year was markedly different:
- Name changed — from a tooling platform (Vertex AI) to an application platform (Agent Platform)
- Chips split — training and inference separated, a product-line-level bet
- Data center redesigned — Virgo is a ground-up architectural overhaul
- Selling point shifted — from selling "AI capabilities" to selling "agent lifecycle management"
The underlying judgment is: in the future, the primary workload on the cloud will no longer be code written by humans, but tasks run by agents. If this judgment is correct, all cloud vendor products need to be redesigned around agents; if it's wrong, Google has effectively bet at least two years of R&D budget on the wrong direction.
AWS hasn't made its move yet (re:Invent is due at the end of the year), and Azure's Microsoft Fabric MCP made a move last month. But Google's boldness this time lies in simultaneously re-engineering hardware, networking, platform, and commercialization models.
Anyone familiar with the cloud industry knows: getting chip generations and data center architecture wrong means a three-year minimum to recover. Google's willingness to bet this heavily shows real confidence in its agent roadmap.
Sources: Google Cloud Next '26: Gemini Enterprise Agent Platform Leads AI-Centric News (Virtualization Review); Google Cloud Next made it clear: AI is coming for everything (The Register); CocoLoop; Google Cloud Next 2026: News and updates (Google Official Blog)