"The Era of Pilots
Is Over" Full-Stack Bet

Over the past two years, Google Cloud has accumulated too many brands at the AI application layer: Vertex AI, Agentspace, Agent Development Kit (ADK), Vertex AI Studio, Model Garden... A frequent question from enterprise customers in sales conversations was: "What exactly is the relationship between all of these?" This question received a formal answer at Next 2026—Google renamed the above products collectively to Gemini Enterprise Agent Platform, unified under a single umbrella brand, and added Agent Observability (an AIOps layer for link tracing, cost monitoring, and anomaly detection).

The signal of consolidation goes beyond branding. Google simultaneously announced Workspace Studio—a no-code agent builder embedded in Gmail, Docs, and Sheets. Non-technical users can directly drag and drop to connect different data sources and agents within a familiar Office-style interface, invoking the A2A protocol to orchestrate cross-system workflows. This is Google's answer to competitors' challenges—Microsoft embeds Copilot in Office 365, Salesforce embeds Agentforce in CRM, and now Google embeds Workspace Studio into the productivity tool ecosystem of 3 billion users. Distribution is the deepest moat.

Google released two TPUs with distinctly different characters this time, driven by an important engineering judgment: Training and inference are two fundamentally different computational problems that should be solved with entirely different chips.

"Near-linear scaling" is the most important technical claim here—most distributed training systems encounter severe communication bottlenecks after exceeding thousands of GPUs. Google claims TPU 8t can smoothly scale to the million-chip level, with the key being the two-layer flat topology design of the Virgo network, eliminating the multi-hop latency of traditional Fat-tree architectures. TPU 8i's design philosophy resonates with Groq's LPU: make SRAM larger, reduce reliance on HBM, keep weights resident on-chip as much as possible, and eliminate memory bandwidth bottlenecks—the difference is that Google's implementation is at a larger scale with deeper ecosystem integration.

In 2025, Google, along with Microsoft, Salesforce, SAP, and others, released the Agent-to-Agent (A2A) protocol to address interoperability issues between AI agents from different vendors. The industry's reaction at the time was cautiously optimistic—a good idea, but no one knew if it would actually be adopted. A year later, the answer arrived: at Next 2026, Google announced the A2A protocol has been upgraded to v1.2, with 150 organizations having deployed A2A to production environments. Microsoft Azure AI Foundry and AWS Bedrock now natively support it, and the A2A protocol has been submitted to the Linux Foundation for stewardship, entering an open governance phase.

The most important new capability introduced in v1.2 is Cryptographic Agent Cards: each agent carries a digital identity certificate issued by a trusted authority, solving the previously biggest security pain point—you couldn't verify whether the agent communicating with you was actually the one it claimed to be. It's worth noting that A2A and Anthropic's MCP (Model Context Protocol) are not in competition: MCP solves "how agents call tools/APIs," while A2A solves "how agents delegate tasks to other agents." Google supports both protocols simultaneously in the ADK—this is a pragmatic choice and a de facto acknowledgment of MCP's status as a standard.

Agentic Data Cloud is Google's repositioning of the enterprise data layer—the core requirement is to let agents securely access data across enterprise systems without needing to move all the data to the same place. The Cross-cloud Lakehouse carries the most strategic significance: BigQuery can directly query data stored in AWS S3 and Azure Data Lake without ETL—enterprises don't need to "go all-in on Google Cloud" to use Google's AI layer, which directly undermines the logic of "migration costs" as a competitive moat.

Google's $32B acquisition of cloud security company Wiz in 2025 was the largest security acquisition in tech history, and many didn't understand the price at the time. At Next 2026, the strategic intent behind the integration was fully revealed for the first time—Agentic Defense combines Google Threat Intelligence with Wiz's cloud asset visibility capabilities, specifically addressing new security issues in the agent era: privilege sprawl, prompt injection, and cross-agent lateral movement. Wiz maintains multi-cloud neutrality, allowing Agentic Defense to cover customers' agent deployments on any cloud—not just Google's own turf.

Among all the announcements at Next 2026, one was overlooked by most tech media: Google Cloud announced the establishment of a $750M partner fund, specifically designed to support ISVs, SIs, and distribution partners in deploying Google AI capabilities to enterprise customers. This is not an ordinary market support budget, but the capitalization of a clear strategic judgment: The bottleneck for AI scaling is not the model, but distribution. In the same month, OpenAI announced DeployCo—a $10B joint venture with PE firms like TPG, Bain, and Carlyle.

Both models are answering the same question: who will help enterprises actually use AI? OpenAI chose Wall Street's PE network, betting on quickly winning deals within portfolio companies where trust relationships already exist; Google chose the partner ecosystem built up over thirty years, betting on a distribution network with broader coverage but relatively less depth per deal. The winner of this distribution war won't necessarily be the one with the best technology, but the one who can most efficiently translate AI capabilities into actionable enterprise workflows.

Thomas Kurian mentioned Google's "four-layer advantage" multiple times in his keynote: custom chips (TPU 8t/8i + Virgo, not dependent on the Nvidia supply chain), frontier models (Gemini Ultra 2 series), cloud platform (Gemini Enterprise Agent Platform), 3B-user distribution (Workspace). Each of these four layers has competitors, but no single competitor has all four—Microsoft has distribution but no custom chips and relies on OpenAI for models; AWS has a cloud platform and infrastructure but no custom general-purpose LLM or consumer-grade distribution; OpenAI has the strongest model brand but lacks chips, a cloud platform, and distribution channels (DeployCo is precisely addressing this gap); Anthropic is the most trusted by enterprises but possesses none of the four layers.

Whether Google's four-layer proposition holds up depends on whether each layer is truly "good enough" rather than merely "existing." Another detail that cannot be ignored: the exclusive cloud agreement between OpenAI and Microsoft has begun to loosen, and OpenAI has started migrating some inference workloads to Google Cloud—Google Cloud not only has sufficient TPU inference capacity but also the geographic diversity (data sovereignty compliance) and non-US market coverage that OpenAI urgently needs.