Google accelerates its transition to 'action-oriented AI'… focusing on TPU 8 as the core, betting on a unified agent platform

Businesses using artificial intelligence(AI)

The market’s focus is rapidly shifting. Analysis indicates that we are moving from the stage of answering questions and generating content into an era of “agentic AI” that carries out real execution tasks and assists decision-making. Google Cloud CEO Thomas Kurian(Thomas Kurian) emphasized that this shift is not merely about adding features—it demands a redesign of the overall architecture of infrastructure and software.

At the recent “Google Cloud Next 2026” conference, Google also showcased the next generation TPU 8, its data and AI platform, and its agent platform. The core message is clear: from semiconductors to cloud infrastructure, from the data stack to AI models and then to applications, everything must be integrated into a single unified system in order to run large-scale “action-oriented AI” stably. Market commentary believes Google has officially gone all in in the race to lead in “full-stack AI.”

TPU 8 Takes the Stage—At Bottom, a Contest for a “Unified Platform”

On the surface, the most eye-catching release is TPU 8. Although Google says the new chip has made significant improvements in performance and scalability, industry attention is not on semiconductor performance alone, but on its strategic significance. Because TPU is the core foundation that enables Google to operate its AI services faster and at lower cost.

However, this is hard to view as a direct head-on confrontation with NVIDIA. Many developers and enterprises still rely on NVIDIA’s CUDA ecosystem, and Google is not aiming to exclude it; instead, it is moving forward by broadening choices. In other words, TPU is more like Google’s method of tightly integrating hardware and software to pursue differentiation—not a weapon to replace NVIDIA.

Industry research organizations and analysts point out that the real focus of this release is not TPU 8 itself, but how Google ties it into a narrative around the data platform, cutting-edge AI models, and agent platforms. This means Google is beginning to string together semiconductors—data—AI models—and task execution into a smooth, coherent structure.

From “SaaS” to “Service-Oriented Software”

This transformation is also shaking the framework of the existing software industry. In the past, the shift from on-premises deployment to software as a service(SaaS) changed the way software is delivered and operated; now, analysis holds that software is evolving into the “service-oriented software” stage where it directly produces real business outcomes.

At the core of this shift are AI agents. The problem is that if agents are confined to the systems of individual departments, their value will be very limited. While it is possible to automate simple repetitive tasks, it is difficult to deliver performance improvements across the entire company—such as shortening the time from hiring to onboarding, or removing bottlenecks from quoting to collections.

Ultimately, enterprises need an “intelligent layer” that connects multiple types of data and business systems. Google’s release of “Knowledge Catalog(Knowledge Catalog)” can be seen as the beginning of this direction. Its structure is designed to place an enterprise’s overall analytical data and operational data within the same context, helping AI understand “what is happening.”

Data Platform Competition Now Expands into “Digital Twins”

In the industry’s view, the mature stage of data platforms is shifting from simply generating reports to building “enterprise digital twins.” Digital twins are digital representations that reflect, in real time, the people, assets, processes, and activities within an enterprise. If AI agents are to judge and act based on an enterprise’s real-time state rather than inaccurate fragmented information, they need this kind of structure.

In the early stage, the focus was on departmental data and reporting systems. Then, with platforms such as BigQuery, Snowflake, and Databricks, self-service analytics environments were broadened, but each department still often had its own “data truth.” The next phase is to more directly and realistically reflect events and operational data, thereby modeling enterprise activities more convincingly.

Salesforce and SAP are also working in this direction, but Google—by integrating BigQuery, Spanner, and the metadata layer—uniquely has a data platform that can directly compete with Snowflake and Databricks among hyperscale cloud providers, and it has received recognition. This time’s agent strategy is an extension built on top of that data foundation.

The Key to the Spread of Agents Is “Secure Execution”

The most difficult part of enterprise AI is connecting the flexibility of generative AI with the rigor of enterprise systems. AI is good at generating text and proposing ideas, but real business operations must have clearly defined rules, clear permissions, auditable processes, and well-defined accountability. Therefore, the industry believes that for agents to play a role, it is necessary to lay a “deterministic execution layer” on top of “creativity.”

For example, even when an agent executes a goal, it must also define under what conditions certain behaviors are permitted, what conditions must be met before and after execution, and how results should be recorded. Only with this kind of structure can “AI that can be run safely” be created—not merely “smart AI.”

In this process, concepts such as enterprise knowledge graphs, a layer of behavioral rules, real-time digital twins, and autonomous operations platforms become important. In short, this means AI must go beyond the level of referencing Excel files and dashboards, be able to explore the enterprise’s actual state and relationship networks, and take action within a framework of rules.

Google’s Advantages and Limitations Are Equally Clear

Google has also made meaningful progress in areas such as metadata extraction, data lineage management, knowledge graphs for unstructured documents, and multi-step agent evaluation. In particular, its optimization feature that compiles agent failure cases and proposes directions for improvement has been viewed as going beyond a simple demo—an attempt to move into the “agent operations” stage.

However, challenges are also numerous. The biggest challenge is integrating the same entity that is scattered across different systems. For example, if the “customer” object is defined differently in CRM, finance, customer service, and logistics systems, it is hard for AI to understand it as a single unified entity. Some viewpoints suggest that having only data-quality rules and business glossaries is not enough; it also needs rules that can express the real business processes.

Another challenge is capturing the “why” from human experts. Google is strengthening features that demonstrate how agents arrive at conclusions, but just that alone is difficult to replace the standards of judgment used by skilled employees. Because exceptions that cannot be explained by rules, conflicting priority order, and context-based judgments still, to a large extent, rely on human experience.

Coding Agents Opens the Next Round of Competition

Another battleground in the competition among agent platforms is “coding.” Industry believes that the fastest route to general-knowledge-worker agents is through coding agents. Because agents must interact with the external world, they ultimately need to call various tools, and in that process, the ability to write, modify, and execute code becomes key.

Anthropic’s Claude Code and OpenAI’s Codex are typical examples. Google did not push a standalone coding product to the forefront; instead, it integrated it into its broader platform.