Google Cloud Next 2026: What TPU 8i and Agent Inference Pods Mean for Your Next

24.04.2026

7 min read

On April 22, 2026, at Cloud Next in Las Vegas, Google unveiled the eighth TPU generation, separating training and inference on the hardware side for the first time. TPU 8t connects up to 9,600 chips for training workloads, while TPU 8i bundles 1,152 chips per inference pod with triple the on-chip SRAM. Concurrently, the Gemini Enterprise Agent Platform went live. For German corporate boards, this isn’t a technology detail, but a new purchasing dimension that will appear in the next IT investment review.

The Essentials at a Glance

Inference capacity becomes a purchasing factor. TPU 8i bundles 1,152 chips per pod, three times more SRAM than Ironwood, optimized for parallel agent execution.
Agentic infrastructure becomes a commodity. Gemini Enterprise Agent Platform live since April 22. Accenture, BCG, Mars, and Merck as reference customers. The market is opening up faster than many boards plan.
Board question from now on: Inference scaling or supplier risk. Those unfamiliar with TPU 8i and A2A Protocol cannot provide a reliable commitment to productive scaling.

What Sundar Pichai announced on April 22

Google CEO Sundar Pichai presented three product components in the keynote window at 10:00 PST. The TPU 8 generation is the most visible, but not the most strategically important. TPU 8t uses a new inter-chip interconnect technology and links up to 9,600 TPUs plus two petabytes of shared high-bandwidth memory in a single Superpod. Google claims this offers three times the performance of Ironwood. TPU 8i is the inference twin: 1,152 chips per pod, triple the on-chip SRAM capacity, optimized for low latency with millions of parallel agent requests.

The second component is the Gemini Enterprise Agent Platform. It went directly into general availability on April 22. Google reported a 40% quarterly increase in paid monthly active users for the predecessor platform in the first quarter of 2026. Meanwhile, Accenture and BCG announced partnership extensions to the Gemini Enterprise Acceleration Program. Mars is using the platform for 150,000 employees. Merck had already announced a $1 billion Agentic AI alliance with Google Cloud on April 22, which we have categorized elsewhere as a template for board decisions.

The third component, often underestimated in boards, is the Agent-to-Agent Protocol. Google has positioned it as an open standard, with OpenAI and Anthropic providing interoperability signals. This creates an ecosystem that defines agent communication across providers. For IT strategy, this means lock-in arguments are shifting from individual models to the underlying infrastructure layer.

1,152

TPU 8i chips per inference pod, triple the on-chip SRAM compared to predecessor Ironwood. Goal: execute millions of competing agents with low latency cost-effectively.

Source: Google Cloud Next 2026 Keynote, 04/22/2026

What is TPU 8i? TPU 8i is the inference variant of the eighth Google TPU generation, announced on April 22, 2026, at Cloud Next. It connects 1,152 chips per pod with triple the on-chip SRAM compared to the predecessor generation Ironwood and is explicitly optimized for the parallel execution of autonomous AI agents. The separation between training (TPU 8t) and inference (TPU 8i) is the strategic signal: Google treats these two workload types as standalone products with their own pricing and scaling logic.

Three Board Questions for the Next IT Investment Review

The message to the board is not to buy a Google TPU roadmap now. It’s about knowing the vocabulary before the IT management asks about the budget. Three points should be on the agenda for the next board meeting, regardless of the company’s hyperscaler preference.

Firstly: Inference capacity as a separate purchasing factor. Until Cloud Next 2026, many IT teams have combined inference and training. Google is now separating them on the hardware side. AWS had already differentiated Trainium 2 and Inferentia 3 last year. The strategic consequence: budget plans must include two numbers, not one. Those who confuse this calculate peak capacity as base load and pay the surcharge.

Secondly: Build, Buy, or Managed. The Gemini Enterprise Agent launch dramatically shortens the time-to-deployment for agents. The board question is not “do we want to build agents,” but “do we build them ourselves, buy standard agents, or rent a managed agent lane.” The same distinction that was clarified in 2025 for managed services now reappears on the agent level. The BCG partnership with Google shows where the market is heading: bundles of platform plus integration plus change management will become the standard purchasing package.

Thirdly: Rethinking sovereignty questions. With TPU 8i, Google becomes technologically harder to catch up with. European cloud providers that rely on NVIDIA chips must compete with an integrated hyperscaler on inference prices. This is not prohibited, but it changes the economic equation. For DACH boards that need data location guarantees, the decision between EU providers and hyperscalers will become harder, not softer.

A board that has no position on inference capacity and agent governance by 2026 will make the next IT budget decision blindly. Google Cloud Next was the signal that the market is currently defining the questions before the answers are circulating.

What this specifically means for German boards

Three reflexes are useful in the coming weeks. Firstly: review your own agent roadmap. If you’re currently running pilot agents on generic LLM infrastructure, you should raise the question of productive capacity in the next steering committee. Secondly: document the contract situation with hyperscalers. What exit options are realistic two years after rollout if TPU 8i becomes the standard? Thirdly: involve compliance early on. Agent-to-agent communication raises liability questions that neither DORA nor NIS2 address cleanly. This will be refined, but not within this quarter.

The calm assessment is: Google Cloud Next 2026 was not the moment that tipped the market. It was the moment that made it clear that it is tipping. Boards that base their architecture decisions in 2026 on 2024 assumptions will be in adaptation mode in 2027. The interesting leadership task is to take a position now that will still be viable in twelve months.

Frequently Asked Questions

What is TPU 8i and why is it relevant for boards?

TPU 8i is Google’s eighth generation of Inference TPU, announced on April 22, 2026, at Cloud Next. It bundles 1,152 chips per pod with triple on-chip SRAM. Relevant for boards because inference capacity now appears as a separate purchasing category in IT budgets.

What is the Agent-to-Agent Protocol?

An open standard for agent communication between different providers, presented by Google at Cloud Next. OpenAI and Anthropic have given interoperability signals. As a result, lock-in arguments shift from the model to the underlying infrastructure.

Do we need to switch to Google now?

No. The message is not about switching technology, but rather a vocabulary check. Boards should master the distinction between training and inference capacity and take a position on agent governance. AWS and Azure will follow with their own responses.

How does this relate to Gartner’s $6,150 billion forecast?

Gartner’s 2026 IT spending forecast sees a double-digit percentage of the market running on AI. Cloud Next 2026 provides the infrastructure that makes this budget shift operational. Boards that do not have an inference line in their budget will need to follow suit over the course of the year.

What compliance questions need to be addressed now?

Agent-to-agent communication creates liability scenarios that DORA and NIS2 only partially cover. Boards should instruct their legal and compliance teams to draft contract clauses for foreign agents in their own system before the first production cases occur.

Editor’s Picks