Google Cloud Next ‘26 ran in Las Vegas on April 22-23, 2026. After reading the full source material from the conference announcements, the picture that emerges is not a product update cycle. It is a strategic repositioning. Google Cloud CEO Thomas Kurian’s framing was explicit: “The experimental phase is behind us. How do you move AI into your entire enterprise? The answer is a unified stack.”
Three interlocking bets define the announcement set. First, the Gemini Enterprise Agent Platform consolidates Google’s fragmented AI tooling into a single surface for building, running, and governing autonomous agents. Second, the eighth-generation TPUs split into two purpose-built variants — one for training, one for inference — reflecting a fundamental shift in how Google thinks about AI infrastructure economics. Third, Workspace Intelligence attempts to turn Google’s productivity suite into a shared knowledge layer that agents can reason across, not just a collection of isolated apps.
Taken together, these announcements describe Google’s attempt to become the operating system for enterprise AI. Whether that framing holds up in practice depends on execution, but the architectural intent is clear and worth understanding in detail.
1. Gemini Enterprise Agent Platform: from fragmented tools to a governed agentic OS
The most significant announcement is the Gemini Enterprise Agent Platform, which consolidates Vertex AI, Agentspace, and related tooling into a single unified environment. The consolidation matters because the previous state was fragmented: teams building agents on Google Cloud had to navigate multiple products with overlapping capabilities and unclear boundaries.
The new platform is structured around four operational concerns that reflect where enterprise AI deployments actually break down.
Building agents without proliferation. A central agent registry is designed to prevent organizations from accumulating dozens of nearly identical agents built by different teams. This is a real problem. Without a registry, agent sprawl becomes an operational and governance liability — the same problem that plagued microservices before service meshes and service catalogs became standard. The platform also includes Agent Studio, a natural-language interface for creating agents, and a flowchart-style tool for mapping how multiple agents work together.
Running agents that can actually complete work. Long-running agents can now handle multi-step processes without pausing for human input at every decision point. This is the capability gap that has made most enterprise AI deployments feel like demos rather than production systems. The platform adds a Memory Bank that gives agents persistent context across sessions, so they do not start from scratch with every interaction. Sandboxed execution environments let agents run code and browser automations without exposing host systems.
Governing agents as a security surface. Autonomous agents create attack surfaces that traditional enterprise security models were not designed for. Google is shipping cryptographic identities for each agent, upstream filters against prompt injection, and anomaly detection for suspicious behavior — unauthorized data access, reasoning loops that never terminate, unexpected lateral movement. Simulation tools let teams test agents against synthetic interactions before production deployment.
Multi-agent orchestration. The Agent-to-Agent Orchestration capability, Agent Gateway, and Agent Observability tooling address the coordination problem that emerges when multiple agents need to divide work, hand off tasks, and maintain coherent state. This is the hardest part of agentic systems at scale, and Google’s approach of treating it as a platform concern rather than an application concern is architecturally correct.
The available model roster includes Gemini 3.1 Pro, Nano Banana 2, Lyria 3, and Anthropic’s Claude Opus 4.7 — the same model that GitLab integrated into its Duo Agent Platform this week. The multi-model availability is notable: Google is positioning the platform as model-agnostic infrastructure, not a Gemini-only walled garden.
The Data Agent Kit is worth a separate mention. It is a data engineering experience built for practitioners who want to use their existing tools — dbt, Spark, BigQuery — while adding agentic capabilities on top. This is a more pragmatic approach than asking data teams to rebuild their workflows around a new paradigm.
2. TPU 8t and 8i: purpose-built silicon for a world where inference costs matter as much as training
The hardware announcement is architecturally significant because it reflects a real shift in AI infrastructure economics.
Google is splitting its eighth-generation TPUs into two variants for the first time: TPU 8t for training and TPU 8i for inference. The split is a direct response to the rising inference demands of agents that plan, act, and learn in loops. Training and inference have different resource profiles, and optimizing a single chip for both means compromising on both.
TPU 8t is built for training at scale. Google claims 2.8× to 3× performance gains over the previous generation. The scale story is where Google has a structural advantage over Nvidia: while Nvidia’s Rubin GPUs connect up to 576 accelerators in a single NVLink domain before slower interconnects kick in, Google uses optical circuit switches to link 9,600 TPUs in a single pod. The new Virgo Network can tie multiple data centers together into clusters of up to one million TPUs. A managed Lustre storage system pushes data directly into accelerator memory. Google is targeting 97% “goodput” — the share of time chips spend actually training rather than waiting on checkpoints or recovering from errors.
The scale numbers are meaningful for frontier model training, but the goodput metric is the more operationally interesting claim. Training efficiency at scale is not just about peak FLOPS. It is about how much of that compute actually produces useful gradient updates versus how much is lost to coordination overhead, checkpoint latency, and fault recovery.
TPU 8i trades some compute for more on-chip SRAM and faster HBM. The larger SRAM keeps more of the key-value cache — the model’s memory of previous responses — directly on the chip, so cores do not sit idle waiting for data. A Collective Acceleration Engine is designed to speed up mixture-of-experts models. A new network topology called Boardfly cuts chip-to-chip latency. Google claims 80% better price-performance and up to 2× improvement in performance per watt compared to the previous generation.
The inference chip story matters more for most enterprise deployments than the training chip story. Most organizations are not training frontier models. They are running inference at scale, and the economics of inference — latency, throughput, cost per token — determine whether agentic AI is financially viable in production. The TPU 8i’s focus on KV cache size and latency reduction is directly targeted at the bottlenecks that make long-context, multi-step agent interactions expensive.
Both TPUs now run on Google’s Arm-based Axion CPUs for the first time, completing the vertical integration of Google’s AI infrastructure stack.
| Chip | Optimized For | Key Design Choice | Scale |
|---|---|---|---|
| TPU 8t | Training | Optical interconnects, 97% goodput target | Up to 1M TPUs via Virgo Network |
| TPU 8i | Inference | Large on-chip SRAM, KV cache locality, Boardfly topology | Always-on enterprise workloads |
3. Workspace Intelligence: turning productivity apps into a shared knowledge layer
The third major announcement is Workspace Intelligence, a layer that connects content across Gmail, Docs, Drive, Meet, and Chat so that Gemini and agents built on the platform can understand relationships between emails, meetings, chats, and files rather than querying each app in isolation.
The specific capabilities announced are incremental individually but coherent as a system:
- Gmail: Gemini sorts incoming messages and summarizes topics
- Google Chat: users can create calendar events or documents directly from a conversation
- Docs: Gemini drafts content from emails and files
- Sheets: Gemini builds dashboards
- Slides: Gemini assembles presentations
- Drive Projects: groups files and emails into topic-based workspaces
The strategic intent is to make Workspace the connective tissue for enterprise agents. An agent that can reason across email threads, meeting notes, shared documents, and chat history has a fundamentally different capability profile than an agent that can only access one data source at a time. Workspace Intelligence is Google’s attempt to make that cross-application context available as a platform primitive rather than requiring each agent to implement its own data integration.
Google is also offering a faster migration path from Microsoft 365, which is a direct competitive move. The enterprise productivity market is the distribution channel for enterprise AI adoption, and Google is betting that Workspace’s integration depth will be a meaningful differentiator as organizations decide where to build their agentic workflows.
4. The competitive context: what this means for the enterprise AI platform race
Google Cloud’s financial position at the time of this announcement is worth noting. Alphabet reported 48% year-over-year revenue growth for cloud operations in Q4 2025, the fastest growth rate among the three major hyperscalers. Cloud backlog surged 55% quarter-over-quarter to $240 billion. Sundar Pichai cited 750 million Gemini users and $175-185 billion in planned capital expenditure.
These numbers matter because they describe a company with the financial capacity to sustain the infrastructure investment required to compete at the frontier of AI. The TPU program, the Virgo Network, the Workspace integration — none of these are cheap. Google is betting that vertical integration from silicon to application layer is the right architecture for enterprise AI, and it has the balance sheet to make that bet credible.
The competitive framing is also explicit. The Agent-to-Agent protocol, the agent registry, the governance tooling — these are all designed to make Google’s platform the coordination layer for enterprise AI, not just a model provider. The risk for organizations building on this stack is the same risk that has always existed with platform bets: deep integration creates leverage for the platform vendor as well as for the customer.
For platform engineering teams, the practical question is not whether Google’s vision is compelling — it clearly is — but whether the governance and portability story holds up. Cryptographic agent identities and anomaly detection are the right primitives. Whether they are implemented in a way that gives organizations genuine control, or whether they primarily serve to lock workloads into Google’s observability stack, will become clear as the platform matures.
5. What this means for teams building AI infrastructure
Three practical implications for platform and infrastructure teams.
The inference economics argument is now explicit. The TPU 8i’s 80% price-performance improvement and 2× performance-per-watt claim, if accurate, changes the cost model for running agents at scale. Teams evaluating AI infrastructure should now be running inference benchmarks on purpose-built inference chips, not just comparing training performance. The training vs. inference split in silicon is a trend that will accelerate across the industry.
Agent governance is becoming a platform engineering concern. The cryptographic identities, prompt injection filters, and anomaly detection that Google is shipping are not application-layer features. They are infrastructure primitives. Platform teams that are not already thinking about agent identity, agent permissions, and agent observability as first-class concerns are behind the curve. Google’s announcements this week will accelerate the expectation that these capabilities exist at the platform level.
The multi-agent coordination problem is real and unsolved. Agent-to-Agent Orchestration, Agent Gateway, and Agent Observability are all attempts to address the same underlying problem: when multiple agents need to collaborate on a task, the coordination overhead and failure modes are qualitatively different from single-agent systems. Google is shipping tooling for this, but the problem is hard and the solutions are early. Teams building multi-agent systems should treat coordination as a first-class architectural concern, not an afterthought.
A compact view of the announcement set
| Announcement | Category | Why It Matters |
|---|---|---|
| Gemini Enterprise Agent Platform | Platform | Consolidates fragmented AI tooling into governed agentic OS |
| Agent registry | Governance | Prevents agent sprawl, enables lifecycle management |
| Memory Bank | Runtime | Persistent agent context across sessions |
| Cryptographic agent identities | Security | First-class identity model for autonomous agents |
| TPU 8t | Training silicon | 2.8-3× perf gain, 9,600-chip pods, 97% goodput target |
| TPU 8i | Inference silicon | 80% better price-perf, KV cache locality, Boardfly topology |
| Virgo Network | Infrastructure | Up to 1M TPU clusters across data centers |
| Workspace Intelligence | Knowledge layer | Cross-app context for agents across Gmail, Docs, Drive, Meet, Chat |
| Data Agent Kit | Data engineering | Agentic capabilities on top of existing practitioner tools |
| Agent-to-Agent Orchestration | Multi-agent | Coordination layer for multi-agent workflows |
Radar takeaway
The most important signal from Google Cloud Next ‘26 is not any individual announcement. It is the architectural claim: that the right way to build enterprise AI is a unified stack from silicon to application, with governance baked in at every layer.
Watch the Gemini Enterprise Agent Platform if you are evaluating where to build agentic workflows. The consolidation of Vertex AI and Agentspace removes a real source of confusion, and the governance primitives — agent identities, anomaly detection, simulation tooling — are the right foundation for production deployments.
Watch the TPU 8t/8i split if you are making infrastructure decisions for AI workloads. Purpose-built inference silicon is becoming a meaningful cost lever, and Google’s scale advantage in training interconnects is real.
Watch Workspace Intelligence if you are thinking about enterprise AI distribution. The organizations that win the enterprise AI platform race will be the ones whose AI can reason across the full context of how work actually happens — email, meetings, documents, chat — not just the ones with the best models.
The experimental phase is over. The question now is which platform teams can govern, scale, and operate agentic systems in production. That is a harder problem than building them.
This Tech Radar bulletin is automatically curated by the OpenClaw AI network and technically supervised by Senior System Architect @TuanAnh. Data is extracted real-time from trusted sources.
📚 Related Reading:
- GitOps at Scale with K8s & ArgoCD
- Deploying an Autonomous AI Swarm
- MCP Engineering in Production Series