Thomas Kurian, CEO, Google Cloud
Google Cloud used its annual Las Vegas conference this week to make a single, expensive argument to enterprise buyers. Running agents at production scale requires specialization at every layer of the stack, from silicon to data to security, rather than one general purpose architecture stretched across every workload.
The announcements at Next 2026 spanned eighth generation Tensor Processing Units, a new Gemini Enterprise Agent Platform, an Agentic Data Cloud standardized around Apache Iceberg, and an Agentic Defense stack combining Google Threat Intelligence and Security Operations with Wiz. Behind all of it sits Alphabet's planned 2026 capex of $175 billion to $185 billion, nearly double last year. The question for CXOs is no longer whether Google can compete with AWS and Microsoft on AI infrastructure. It is whether the specialized stack Google is selling delivers enough operational benefit to justify migrating off the multi vendor architectures most enterprises have already settled into.
The most consequential announcement at Next was architectural. Google split its eighth generation TPU into two distinct chips. TPU 8t handles large-scale pre-training, scales to 9,600 chips in a single superpod with 2 petabytes of shared high bandwidth memory, and delivers up to 2.7 times better training performance per dollar than the seventh generation Ironwood. TPU 8i targets sampling, serving and reasoning workloads, ships with 384 megabytes of on chip SRAM versus 128 megabytes on TPU 8t, and combines a Collectives Acceleration Engine that Google says reduces on chip collective latency by five times with a new Boardfly topology that reduces network diameter and improves latency for communication intensive workloads. Google claims up to 80 percent better inference performance per dollar than Ironwood, particularly for low latency large mixture of experts workloads.
The split matters because Google is now productizing separate training and inference paths rather than marketing a single TPU generation as the default answer for both. Mixture of experts models, long context reasoning and millions of concurrent agents have made inference the dominant cost center, and the bottlenecks differ from training. TPU 8i optimizes for keeping key value cache resident on silicon and crushing all reduce latency during decoding. TPU 8t optimizes for raw compute throughput and memory bandwidth across thousands of chips. Forcing one die to do both requires tradeoffs that hurt unit economics at agent scale.
The Agentic Data Cloud is the second leg of the strategy. Google evolved Dataplex Universal Catalog into a Knowledge Catalog that maps business semantics across structured and unstructured data and uses Gemini to autonomously generate descriptions, glossaries and verified SQL patterns. Integrations with Palantir, Salesforce Data360, SAP, ServiceNow and Workday are in preview. The Cross-Cloud Lakehouse extends Google's lakehouse strategy by standardizing on Apache Iceberg REST Catalog and using Cross-Cloud Interconnect to enable query access across data in AWS and Azure without wholesale migration. Bidirectional federation with Databricks Unity Catalog, Snowflake Polaris and AWS Glue is also in preview.
The strategic intent is clear. Google is conceding that enterprise data will not move to a single cloud and is positioning Google Cloud as the query and reasoning layer over data that lives elsewhere. This inverts the historical hyperscaler playbook of pulling data in. It also addresses the most common reason enterprises hesitate to put agents into production. Without trusted business context, agents hallucinate joins, invent metrics and execute incorrect actions. Knowledge Catalog and the cross cloud lakehouse are designed to ground them, though the most important federation features remain in preview rather than general availability.
Google reported that 330 customers each processed more than one trillion tokens over the past 12 months, with 35 customers crossing the 10 trillion token mark. First-party models now serve more than 16 billion tokens per minute via direct API use, up from 10 billion the previous quarter. Anthropic remains the marquee TPU customer and one of Google's most important cloud accounts.
The competitive picture is more complex than the keynote suggested. AWS counters with Trainium 3 in UltraServer configurations and the Bedrock model marketplace. Microsoft has Maia, Cobalt and the Fairwater data center program plus deep enterprise distribution through Azure and Microsoft 365. Both have disclosed sizable AI infrastructure programs of their own. What Google did at Next was provide an unusually integrated stack-level narrative spanning chips, networking, data, models and security, backed by a capex number that signals the buildout is funded. Nvidia Vera Rubin NVL72 will also be a core part of Google Cloud's portfolio through the new A5X instance, which means Google is hedging its silicon strategy with GPU capacity even as it markets a TPU first narrative.
Three constraints deserve attention. First, Google says TPU 8t and TPU 8i will be generally available later in 2026. Until then, the capacity that matters for most production workloads is Ironwood, which is now generally available but already a generation behind the public roadmap.
Second, the Gemini Enterprise Agent Platform is a consolidation of Vertex AI rather than a clean break. Customers who built on Vertex AI agents in 2024 and 2025 will face migration work, and the new Agent Studio, Agent Registry and Agent Gateway components are still maturing. Migration effort, agent observability, and identity remain harder problems than the keynote demos implied, and Google has not yet published comprehensive third-party benchmarks beyond its own price performance claims.
Third, the cross cloud lakehouse depends on Apache Iceberg becoming a true neutral standard. Databricks, Snowflake and AWS each have commercial reasons to keep their catalog implementations differentiated. The bidirectional federation features that make the architecture viable are currently in preview. Real interoperability will be tested when one of those vendors changes a default that breaks Google's federation.
For CXOs the takeaway is not that Google has won the agentic enterprise. It is that the cost structure of agentic AI rewards specialization at every layer, and Google is investing ahead of that conclusion with an unusually detailed public stack-level narrative. Operational simplicity may still favor standard GPU fleets and existing multi cloud data tools for many enterprises. But buyers who treat AI infrastructure as a single monolithic vendor commitment will pay more than buyers who match workload to silicon, model to context and data plane to where the data already lives.
The practical action is to audit current AI workloads against the bifurcation Google is betting on, and to separate generally available capabilities from preview features when scoring vendor proposals. Training spend, inference spend and agent orchestration spend are increasingly distinct line items with different elasticity to vendor choice. Locking long-term commitments to a single architecture before that separation is fully understood is the most expensive mistake available in 2026.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。