Prologue: "The Era of Experimentation Is Over." — The Single Narrative Told Across Three Days
April 22–24, 2026. Las Vegas.
In front of 32,000 attendees at Google Cloud Next 2026, Google Cloud CEO Thomas Kurian opened with this declaration:
"The pilot phase is behind us. The real challenge we now face is how to deploy AI across the entire production environment of the enterprise."
The numbers back it up. Roughly 75% of Google Cloud's customers are already using AI products in their businesses, and 330 of them processed over one trillion tokens each in the past twelve months. API-based model throughput has reached 16 billion tokens per minute. This is no longer about "trying AI." It is about running AI across the entire enterprise.
But the most important message of these three days was not about model performance.
DAY 1 was a declaration — the vision of the Agentic Enterprise and the product suite to realize it.
DAY 2 was implementation — developer demos and concrete methodologies for running agents in production.
DAY 3 had no keynote at all. Zero new product announcements. The program wrapped up by noon.
At first glance, it looked like a cooldown day. But read the structure, and the "zero-announcement final day" was what completed the three-day narrative.
Technology media outlet SiliconANGLE described the essence of Google Cloud Next 2026 as "the control plane war."
What Google is pursuing is not the delivery of AI features. It is becoming the OS of the Agentic Enterprise — the foundation for running AI agents safely, affordably, and governably across the entire organization.
This article reads the structure that only becomes visible when you step back and look at all three days as one.
Chapter 1: Vertical Integration — Google's "Apple-Style" Bet
The competitive structure of AI companies has shifted significantly in recent years.
OpenAI and Anthropic deliver model capabilities horizontally via APIs. AWS lets customers choose among multiple models on its neutral Bedrock platform. Microsoft embeds Copilot into its own applications.
Only Google made a different bet.
From TPU (custom-designed semiconductor chips)
→ Gemini (foundation model)
→ Agent Platform (agent development infrastructure)
→ BigQuery / Lakehouse (data infrastructure)
→ Workspace (end-user applications)
— vertically integrating everything from the physical chip design to the Gmail and Sheets that employees use every day, all under a single architectural blueprint.
Kurian continued:
"You cannot deliver AI by just cobbling together fragmented silicon chips or isolated platforms. To unlock real value, you need a complete system."
The investment scale behind this vertical integration is staggering. Alphabet's capital expenditure is projected to grow roughly sixfold, from $31 billion in 2022 to $175–185 billion in 2026, with the majority directed at cloud and machine learning compute.
Pichai further emphasized that Google itself is "Customer Zero." Roughly 75% of newly written code inside Google is AI-generated, complex code migrations now complete 6x faster than manual efforts a year ago, and security operations center agents have reduced threat mitigation time by over 90%.
Google is not selling AI developed in a research lab. It is offering the same AI it has battle-tested across its own operations, development, and security workflows.
The implication for business leaders:
The AI adoption decision is shifting from "which model to use" to "which integrated stack to ride." The era of deploying individual generative AI tools at the department level is ending. Choosing a platform with a coherent design philosophy — from chip to application — will define a company's long-term competitiveness.
Chapter 2: The Inference-Only Chip — A Historic Fork
One of the most technically significant announcements across the three days was the design philosophy behind the 8th-generation TPU.
For the first time, Google released two distinct chip variants with explicitly separated purposes. TPU 8t (Training) is specialized for the model training phase. TPU 8i (Inference) is specialized for inference.
Why does this matter? Training a model is a one-time event. But inference — the process where AI agents analyze data, make judgments, and execute actions in daily operations — runs perpetually. In an era where agents continuously run inference loops in the background, inference cost dominates the total cost of enterprise AI operations.
TPU 8i triples the on-chip ultra-fast memory (SRAM) to 384MB compared to its predecessor, virtually eliminating the latency from loading data from external memory (the memory wall).
Google also announced that a cluster of 96 NVIDIA B200 GPUs on GKE (Google Kubernetes Engine) achieved one million tokens per second in inference throughput — compared to 22,000 tokens per second on a previous 4x H100 GPU configuration.
The implication for business leaders:
The dramatic reduction in inference cost translates directly to lower agent usage fees. The economic premise for enterprises to run AI agents as "pay-per-use digital labor" around the clock has now been established. The calculus shifts from "AI is expensive, so use it sparingly" to "running AI agents full-time is cheaper than headcount."
Chapter 3: The Language That Agents Speak Has Been Decided
For AI agents to truly function inside enterprise systems, they need a way to communicate and coordinate with each other. At Google Cloud Next 2026, two "common languages" were formally established.
The first is ADK (Agent Development Kit) 1.0, now generally available. ADK is an open-source framework for building AI agents, with official support for Java, Go, Python, and TypeScript. The Java and Go support is particularly significant — it means agents can be directly integrated into existing enterprise development pipelines.
ADK 1.0 also introduces "event compaction." When an agent runs a task over several days, conversation history and logs accumulate until they hit the model's context window limit. Event compaction dynamically summarizes and compresses older history while preserving recent information, enabling agents to maintain effectively unlimited long-running sessions.
The second is A2A (Agent2Agent) Protocol 1.2. A2A is an open standard protocol that allows agents built on different vendors and frameworks to autonomously discover each other's capabilities, communicate, and delegate tasks. It is already operational across 150 organizations, with support from Salesforce, SAP, Workday, Atlassian, and ServiceNow.
While Anthropic's MCP (Model Context Protocol) connects agents to data, A2A connects agents to agents. Google fully supports both.
The implication for business leaders:
What breaks down cross-departmental data silos is no longer human coordination. Agents communicating directly via standard protocols and automating business processes across organizational boundaries — this changes organizational design itself. The concept of "cross-departmental collaboration" will shift from human meetings to autonomous agent communication.
Chapter 4: Killing Data Gravity
A problem that has plagued enterprise IT for years: data gravity. Once petabytes of data accumulate on AWS or Azure, the high egress fees and physical transfer times imposed by cloud providers make it virtually impossible to apply superior AI models from another cloud. Data becomes immovable.
Google's answer: Cross-Cloud Lakehouse. Built on the open-standard Apache Iceberg format, it executes queries directly against data stored in AWS S3 or Azure Data Lake Storage — with zero data copying. Queries travel over dedicated private networks instead of the public internet, dramatically reducing transfer costs.
Also noteworthy is Knowledge Catalog. Traditional data catalogs were metadata tools that tracked where data lived. Knowledge Catalog attaches real-time semantic context — what this data means in a business context — and feeds it to AI agents. It functions as the agent's "memory" for autonomous decision-making.
Smart Storage in GCS automatically tags and vectorizes unstructured data (PDFs, images, audio files) the moment it is uploaded to Google Cloud Storage, eliminating the need for manually built vectorization pipelines.
The implication for business leaders:
The world where data engineers spend weeks building ETL pipelines is becoming obsolete. Instruct an agent in natural language — "Compare recent customer behavior data on AWS with campaign data on Google Cloud" — and the agent autonomously generates the optimal query plan. The shift from "moving data" to "analyzing data where it lives" has profound practical implications for Japanese and global enterprises running multi-cloud strategies.
Chapter 5: 22 Seconds — The Collapse of the Security Timeline
The most shocking data point across all three days was about security.
According to Google's latest M-Trends 2026 report, the time from an attacker's initial system compromise to handing off access to secondary attackers for ransomware deployment or data exfiltration has collapsed from 8 hours to just 22 seconds over the past three years.
22 seconds. Far too short for a human security analyst to receive an alert, interpret it, and initiate incident response.
Francis deSouza, President of Security Products at Google Cloud, stated plainly:
"The AI era demands a new security era. Human analysts cannot keep pace with AI-driven attacks."
Google's answer is Agentic Defense — delegating security operations themselves to AI agents. Three new security agents — Threat Hunting, Detection Engineering, and Third-Party Context — compress manual analysis that typically takes 30 minutes down to 60 seconds. The existing Triage and Investigation agent has processed over 5 million alerts in the past year.
AI-APP (AI Application Protection Platform), integrating Wiz technology acquired for $32 billion, autonomously protects AI applications across multi-cloud environments with Red (attack simulation), Blue (threat identification), and Green (auto-remediation) AI agent teams working in concert.
And Code Mender — Google's direct answer to Anthropic's Claude Mythos. Code Mender autonomously identifies software vulnerabilities, proposes fixes, and rewrites code — fully automated. As Kurian put it: "Defense must also be AI."
The implication for business leaders:
Security has shifted from a "cost center" to an "AI-vs-AI warfare department." Hiring more human analysts will not beat 22 seconds. The CISO's role is irreversibly shifting from managing people to governing a fleet of AI agents. And this is not just a security department issue — for any enterprise running AI agents across all business processes, agent identity management, permissions governance, and behavior auditing become board-level concerns.
Chapter 6: The Japan Signal — "Labor Shortage" as the Greatest Accelerant
On DAY 3, during the Partner Summit, a session titled "Japan GTM: Unlocking the Scaled Opportunity Together" focused on the Japanese market. Yumi Ueno, Google Cloud's Japan partner business lead, emphasized that Japan's rapid demographic shift and severe labor shortage are, paradoxically, functioning as the greatest accelerant for AI agent adoption.
Google positions this structural reality as an "Opportunity."
Concrete proof points from DAY 3: NTT Integration won the "2026 Google Cloud Partner of the Year" award for public sector DX in Japan. NTT DOCOMO and NTT DATA engineers presented a zero-trust architecture running agents in closed environments without VPNs on Cloud Run. Thales demonstrated encryption and key management solutions fully compliant with Japan's APPI, FISC security standards, and My Number Act.
The partner ecosystem investment is massive: Google announced a $750 million partner funding program across Accenture, Deloitte, Capgemini, NTT DATA, and others.
The implication for business leaders:
For Japanese enterprises that can no longer cover operations with human labor, AI agents are not an efficiency tool. They are digital labor itself. Delaying adoption is now synonymous with deepening the labor crisis. What the Japanese market demands is not "using generative AI" but end-to-end agentification of core business flows — order processing, infrastructure control, customer service, security operations.
Conclusion: The Axis of Competition Has Shifted from "Intelligence" to "Governability"
Looking across all three days, one structural shift becomes clear.
The axis of AI competition has irreversibly moved from "which model is smartest" to "how do you run AI safely, affordably, and governably across the entire enterprise."
DAY 1 declared the vision. DAY 2 demonstrated the implementation. DAY 3 closed the operational design loop. The absence of a keynote on DAY 3 was itself the message: the subject is no longer new models — it is operational governance.
What Google presented is the OS of the Agentic Enterprise: inference-optimized hardware (TPU 8i), open cross-vendor protocols (ADK / A2A / MCP), a foundation that destroys multi-cloud data silos (Cross-Cloud Lakehouse), and autonomous defense against 22-second cyber attacks (Agentic Defense / Wiz / Code Mender) — all tightly vertically integrated.
Choosing the right model means nothing without the design for governance.
The Information summarized Google Cloud Next 2026's theme as a shift from "last year's model strength to this year's focus on making models actually usable in the enterprise."
This structural shift applies to every enterprise worldwide. AI adoption has moved past the stage where it can be stopped at PoC. The gap between enterprises that have a design for safely governing AI agents in production and those that do not will now widen rapidly.
"The era of experimentation is over."
Originally published in Japanese on note.com on April 25, 2026.
Satoshi Yamauchi — AI Strategist & Business Designer at Sun Asterisk | Founder & CEO, Leading.AI
Open-source bilingual AI strategy books (14 titles, 10,000+ unique readers in 35 days): github.com/Leading-AI-IO
Top comments (0)