Google came to Las Vegas with a clear message: the time for general-purpose AI infrastructure has passed. From specialized silicon to a new cross-cloud data layer, the company is moving in a new direction.
Las Vegas is known for big events, but few companies have made their case with a bigger price tag. At this year’s Cloud Next conference, Google focused on one main idea: general-purpose AI infrastructure has hit its ceiling, and the companies that succeed next will be those that recognize this.
There were more than 32,000 attendees and over 260 announcements in just three days. Behind it all, Alphabet plans to spend between $175 billion and $185 billion in 2026, almost twice as much as last year. This huge investment is central to Google’s message.
The Silicon Story
The biggest announcement was about architecture: Google introduced a split in its chip design.
For the first time, Google divided its eighth-generation TPU into two distinct chips rather than marketing a single die as the default answer for every workload. TPU 8t is built for large-scale pre-training, scaling to 9,600 chips in a single superpod and delivering up to 2.7 times better training performance per dollar than its predecessor. TPU 8i targets inference and reasoning, offering up to 80% better performance per dollar for the low-latency, mixture-of-experts workloads that agentic AI demands.
The choice to split the chips is more important than the technical details. Training and inference pose different challenges, and using a single chip for both results in worse performance at scale. Google is now turning this difference into a product and encouraging enterprise buyers to consider it in their own planning.
The Agent Platform
The Gemini Enterprise Agent Platform, which builds on Vertex AI, is Google’s all-in-one solution for companies that want to build, scale, and manage AI agents in real-world use.
New components include Agent Studio for no-code deployment, Agent Runtime for sub-second cold starts, and Agent Memory Bank for long-term personalization. Agent Identity assigns each agent a verifiable cryptographic ID, creating a clear, auditable trail of every action. Agent Gateway manages entire fleets from a single control point, enforcing security policies across environments. Agent Anomaly Detection flags unusual reasoning in real time.
The most important new features are Agent Simulation and Agent Evaluation. These tools test and score agents using real traffic. They tackle a key problem in enterprise AI: not just building agents, but making sure they actually do what their operators expect.
The Data Layer
The Agentic Data Cloud tackles what Google sees as the main reason companies hesitate to use agents in production. Without a reliable business context, agents can make up data connections, create fake metrics, and take wrong actions. The intelligence exists, but the foundation is often missing.
At the center is Knowledge Catalog, a universal context engine that maps business semantics across structured and unstructured data, using Gemini to autonomously generate descriptions, glossaries, and verified SQL patterns. The Cross-Cloud Lakehouse, standardized on Apache Iceberg REST Catalog, extends query access across data sitting in AWS and Azure without requiring migration.
This is a key strategic shift. Google does not expect all enterprise data to move to its cloud. Instead, it wants to be the reasoning layer that works with data wherever it is stored. Integrations with Databricks, Snowflake, Salesforce, SAP, ServiceNow, and Workday are currently in preview.
Security and Defense
The Agentic Defense stack combines Google Threat Intelligence, Security Operations, and Wiz, the cloud security platform Google acquired, into one system for detecting, preventing, and responding to threats across cloud and AI environments. The Agent Security dashboard, powered by Security Command Center, automatically maps relationships between agents and models and scans for vulnerabilities.
The Competitive Picture
At Next, Google presented a more integrated and at times more candid, stack-level narrative than usual. The company recognized AWS’s Trainium 3 and Bedrock, as well as Microsoft’s Maia and Cobalt, and Microsoft’s strong enterprise reach through Azure and Microsoft 365. Both competitors have also revealed large AI infrastructure programs.
The most telling detail, however, was not what Google said about its competitors. It was what Google did about its own silicon. The new A5X instance will be powered by NVIDIA’s Vera Rubin NVL72 platform when it becomes available later this year. For a company spending $175 billion to build the case for specialized, Google-designed chips, hedging that narrative with GPU capacity from the world’s dominant AI silicon provider is a significant signal. Google is betting on its TPUs — and keeping its options open.
Three Things Enterprise Buyers Should Watch
The new TPUs will not be generally available until later in 2026. Most production workloads will continue to run on Ironwood, which is now a generation behind the public roadmap, for the foreseeable future.
Customers who used Vertex AI agents in 2024 and 2025 will need to do significant migration work to move to the Gemini Enterprise Agent Platform. The new Agent Studio, Agent Registry, and Agent Gateway components are still developing, and Google has not yet shared comprehensive third-party benchmarks beyond its own price-performance claims.
The cross-cloud lakehouse federation features that support the data architecture are still in preview. Real interoperability will be tested not during a Las Vegas keynote, but when one of Google’s partners, such as Databricks, Snowflake, or AWS, changes a default setting that disrupts the federation. How Google responds in that situation will matter more than any announcement made this week.
The Takeaway
Google Cloud Next 2026 was not about launching new products. Instead, it presented a strategic thesis, supported by a large capital expenditure to make the case convincing. The idea is that the cost structure of agentic AI favors specialization at every layer. Enterprises that match workloads to silicon, models to context, and the data plane to where data already exists will outperform those who treat AI infrastructure as a single, monolithic commitment.
Whether this thesis proves true will depend less on what Google announced in Las Vegas and more on which preview features become generally available, which federation promises hold up in real-world use, and whether the enterprises attending this week are still having the same discussions six months from now.
