Google Builds the Agentic Enterprise Stack at Next '26

The number that matters most from Google Cloud Next ‘26 is not the $750 million partner fund or the 121 exaflops in a single TPU superpod. It is $240 billion — the backlog that Sundar Pichai announced on the first morning of the conference, a figure that had doubled in a single year and represents the largest committed future revenue in Google Cloud’s history. Pichai’s keynote framed everything that followed: Google Cloud now runs at $70 billion in annual revenue, growing 48% year over year, with $175–185 billion in planned capital expenditure for 2026 — more than any year in Alphabet’s history. The question is what that mountain of committed cash is supposed to build.

The answer, delivered across three days in Las Vegas, is the agentic enterprise: a compute and software stack purpose-built to let organizations run autonomous AI agents across business processes at scale. Google Cloud Next ‘26 produced a cluster of interrelated announcements — a renamed and substantially expanded AI development platform, two entirely new chips designed for different points in the agent lifecycle, and a $750 million partner fund to get the world’s major consulting firms building agents on Google infrastructure rather than Microsoft Azure or AWS. The bet behind every announcement is the same: whoever controls the governance layer of enterprise AI agents in 2026 controls the enterprise cloud market in 2030.

The moment Google has been waiting three years to exploit

The agentic enterprise AI market is projected to reach $9 billion in addressable spend in 2026, tiny against the hyperscaler addressable market but growing at rates that make the eventual ceiling the relevant number. Google’s strategic position heading into Next ‘26 is one of genuine tension: technically ahead in model quality and custom silicon, but significantly behind in enterprise penetration. Google Cloud holds approximately 11% of the cloud infrastructure market against AWS at 31% and Azure at 25%, and the gap has proved structurally stubborn. Enterprise relationships are sticky; the Microsoft Office bundle creates a gravitational field that Workspace has never fully escaped; and Google’s historical reputation in enterprise sales is one of technical excellence paired with organizational unpredictability.

The agentic moment represents an unusual break in that stickiness. Enterprise AI agents are not a feature you add to an existing cloud contract — they require new infrastructure, new governance tooling, new training pipelines, and new integration work with every business system in the organization. Every enterprise in 2026 is evaluating agent platforms largely from scratch. That gives Google a window it does not get in normal cloud competition, which is precisely why the company spent three full days in Las Vegas making the case that its full-stack offering — model, runtime, silicon, governance, and productivity suite, all from a single vendor — is the strongest bet for the next five years. The Google/Anthropic $40 billion investment announced days before the conference is the capital evidence behind that thesis: Anthropic’s Claude models are now first-class citizens in Google Cloud’s platform, not a competitive alternative.

The adoption numbers Pichai cited at the keynote are the best evidence that the window is genuinely open. Seventy-five percent of Google Cloud customers are now using AI products, 330 customers processed over one trillion tokens each in the past twelve months, and 35 customers have crossed the ten-trillion-token threshold — a usage tier that did not exist as a commercial reality twelve months ago. Google’s first-party models now process 16 billion tokens per minute through direct API access, up 60% from the prior quarter. These are not product-launch numbers engineered for a press release. They are the fingerprints of enterprises that have moved from AI experimentation into production-grade deployment, and they suggest the addressable market for serious AI infrastructure is arriving faster than most analyst models anticipated.

The original quantified takeaway from synthesizing the $240B backlog, 48% revenue growth, and 60% QoQ token-processing increase: Google Cloud is compounding its AI revenue base at a rate that implies it could close half the gap with Azure’s cloud market share by 2028, if the current trajectory holds. That kind of structural shift — from 11% to something approaching 17–18% — would represent the first meaningful redistribution of cloud market share in five years, and would do it not by winning generic infrastructure spend but by owning the governance layer on top of it.

Three layers of stack, one unified bet

Google’s announcements at Next ‘26 split cleanly into three layers: the platform, the silicon, and the money. Each addresses a different friction point in the enterprise agent adoption cycle, and the strategic logic connecting them is tighter than typical product launches achieve.

The Platform: Vertex AI becomes the Gemini Enterprise Agent Platform

Vertex AI has been rebranded the Gemini Enterprise Agent Platform, and the rename comes with substantive architectural changes rather than a simple cosmetic refresh. The platform absorbs Google’s Agentspace product into a unified workspace that includes an Agent Designer, an Inbox for managing agent activity, long-running agent execution for complex multi-step workflows, and a Skills library for modular agent components. The no-code Agent Designer lets non-engineers build trigger-based workflow automation; the programmatic layer handles teams with deeper technical requirements. Long-running agent support — agents that execute business processes over hours or days without human interruption — is the capability most directly relevant to enterprise use cases that the industry has been promising but rarely delivering at production scale.

Model access within the platform is explicitly pluralist. The Model Garden now surfaces over 200 models, including Gemini 3.1 Pro and 3.1 Flash, open models from the Gemma 4 family, and third-party frontier models — including Anthropic’s Claude Opus 4.7, which reclaimed coding benchmark leads in the weeks before Cloud Next. The inclusion of Claude is not incidental: a customer routing Claude traffic through Gemini Enterprise Agent Platform generates Google Cloud revenue even while running a competitor’s model. That same logic extends to application-layer integrations — Oracle, Salesforce, and ServiceNow agents are natively discoverable and deployable within the platform’s governed environment, enabling cross-vendor agent workflows without exiting a single oversight framework.

That governance layer, not the model quality, is the actual product being sold. Enterprise procurement decisions in 2026 are not primarily model selection decisions; they are risk management decisions. Boards and legal teams are not asking which model scores highest on reasoning benchmarks — they are asking who is accountable when the agent makes an error in a high-stakes business process. The Gemini Enterprise Agent Platform’s consolidation of agent building, deployment, monitoring, and security policy enforcement into a single surface is a direct answer to that accountability question, and it is designed to make the answer “Google Cloud.”

The Silicon: Two chips, two jobs

Google’s eighth-generation TPUs are the most structurally interesting announcement at Next ‘26, primarily because they represent the first time the company has built separate chips optimized for training and inference rather than a single architecture attempting both. The TPU 8t handles training: it delivers 3x the compute performance of the seventh-generation chip, packs 9,600 chips into a single superpod producing 121 exaflops and two petabytes of shared memory, integrates 10x faster storage access for dataset throughput, and scales near-linearly to one million chips in a single logical cluster using JAX and Pathways software. For organizations fine-tuning models on proprietary enterprise data — the most defensible form of AI customization — that scaling profile changes the economics of training fundamentally.

The TPU 8i handles inference, and its design assumptions reveal a clear architectural thesis about what agentic workloads actually require. On-chip SRAM triples to 384 MB, high-bandwidth memory rises to 288 GB — both sized to host the KV cache entirely on silicon, eliminating the round-trips to external memory that degrade latency in multi-turn agent conversations and long-context reasoning chains. ICI bandwidth doubles to 19.2 Tb/s, and a dedicated Collectives Acceleration Engine cuts on-chip latency by up to 5x. Computer Weekly’s coverage of the launch cites 80% better performance per dollar for inference over the prior generation — a claim that, if verified at production scale, would substantially compress the cost model for enterprises running continuous inference across agent-driven workflows.

The two-chip strategy signals something more fundamental than a product update cycle. It signals that Google’s chip team has accepted that training and inference have diverged enough in their computational requirements that a single-architecture compromise produces suboptimal results for both. As agents become the primary inference workload — running continuously, handling long contexts, requiring consistent latency for interactive steps — the inference chip’s memory and latency profile becomes the binding constraint on what agents can actually do in production. The TPU 8i is a $175 billion bet that this is the constraint that will define competitive AI infrastructure for the next three years.

The Money: $750M to lock in the channel

The $750 million partner fund is, at its core, a channel lock-in strategy disguised as a co-investment program. Google is embedding forward-deployed engineers alongside Accenture, Capgemini, Cognizant, Deloitte, HCLTech, PwC, and TCS — seven firms that collectively touch the majority of enterprise AI procurement decisions globally. The fund provides cloud credits, co-investment capital, and go-to-market resources, with agents developed through the program distributed through Gemini Enterprise and the Google Cloud Marketplace. An accompanying $90,000 AI Agents Challenge offers prizes and cloud credits to early-stage builders through June 5, seeding the next tier of platform-native developers.

The Next Web’s analysis frames the fund correctly: it is less about the dollar amount and more about the organizational relationships it is designed to create before Microsoft or AWS can replicate the program. A consulting firm that has already embedded Google engineers and committed to Gemini Enterprise Agent Platform for two anchor client deployments is not a firm that will easily pivot to Azure AI Foundry or AWS Bedrock AgentCore six months later. The fund’s real currency is the time those forward-deployed engineers spend inside client engagements building institutional knowledge that becomes switching-cost infrastructure. Google is buying the next two years of the consulting firms’ default recommendation — and it is doing so at a price the hyperscaler economics make trivially affordable.

The walls that could collapse this thesis

Google’s technical and capital case for the agentic enterprise is the strongest it has assembled in a decade of cloud competition. The counter-case is not primarily technical. It is organizational, political, and structural — and each element deserves serious weight before an enterprise team commits to a Google-first agent platform.

Microsoft’s Copilot+Azure stack entered the enterprise AI market three years before Google’s current agentic push, and first-mover advantage in enterprise software is not easily displaced. The Microsoft 2026 Release Wave 1 rollout embeds autonomous agent behavior across Dynamics 365, Power Platform, and Microsoft 365 Copilot — expanding AI into existing Microsoft licenses that most enterprise customers already pay for. An enterprise CIO comparing Google’s $750M partner fund against a Microsoft program that extends existing O365 and Azure spend without new procurement is comparing a new budget line item against a checkbox on a renewal. The friction asymmetry favors Microsoft in the short term even when Google’s technical architecture is objectively superior. Gemini Enterprise saw 40% QoQ growth in paid monthly active users in Q1 — impressive growth, but from a base of 8 million paid seats across 2,800 companies that Microsoft’s Copilot already exceeds by a wide margin.

AWS’s Bedrock AgentCore, launched in the same week as Cloud Next, reframes the competitive landscape accurately: all three hyperscalers are now shipping managed agent infrastructure simultaneously, and the differentiation between platforms is currently modest relative to their marketing claims. AgentCore’s managed harness and Bedrock’s 31% cloud market share mean AWS customers — which is to say most of the Fortune 500 — have a credible agent platform without changing their primary cloud provider. Google is asking enterprises to either embrace multi-cloud agent architectures or consolidate onto a cloud that, for most large organizations, is currently a secondary or tertiary provider. That is a harder sell than the capability story implies, and the $750M partner fund is the explicit admission that it requires financial incentives to overcome.

Google’s own enterprise execution history supplies the sharpest counterpoint. The Virtualization Review’s analysis of Cloud Next ‘26 notes that the pattern of Google rebranding its enterprise AI platform — AI Platform, then Vertex AI, now Gemini Enterprise Agent Platform — echoes a consistent pattern of strategic repositioning that creates adoption uncertainty among enterprise buyers who prize roadmap stability over technical novelty. The rename from Vertex AI is technically defensible given the architectural changes, but it is the fourth major branding shift for Google’s enterprise AI offering in five years. Procurement teams that built security review and compliance documentation around “Vertex AI” now face update cycles on contracts, audit reports, and internal architecture documents. Minor friction, but minor frictions compound into deal delays.

The broader structural risk is that the “agentic enterprise” remains largely aspirational in mid-2026. The usage numbers Pichai cited are real, but the vast majority of enterprise AI spending this year still funds chatbots, summarization, and retrieval-augmented search — not autonomous multi-step agents running unsupervised for hours. OpenAI’s GPT-5.5 super-app ambitions and Anthropic’s expansion into managed agent deployments are both chasing the same emerging demand, which means the competition for the enterprise agent layer is now a four-way race — with two companies that bring no cloud infrastructure relationship inertia to the fight. When OpenAI and Anthropic can route enterprise agents through any cloud via API, Google’s platform advantage requires consistently superior economics and governance to justify the lock-in.

The operator’s playbook for the agentic transition

The practical question for enterprise architecture teams is not whether Google Cloud Next ‘26 matters — it does — but whether this announcement cluster should change procurement timing, vendor allocation, or infrastructure commitments in the next six months.

The $240 billion backlog and 48% growth rate are the most credible signals in the keynote because they represent signed commitments, not product announcements. Organizations evaluating Google Cloud as a primary or secondary AI infrastructure provider should treat that growth rate as evidence of vendor viability, not vendor dominance. A $70 billion run-rate business growing at 48% is a fundamentally different risk profile than it was three years ago, when questions about Google Cloud’s enterprise commitment were reasonable. On the current trajectory it is not going away, and the Anthropic capital relationship gives it a model-provider depth that neither AWS nor Azure can currently match at comparable scale.

The TPU 8i’s 80% performance-per-dollar claim deserves independent verification before it drives infrastructure decisions. Google’s published benchmark methodologies historically favor workload profiles that match their own architecture. Organizations running heterogeneous agent frameworks across multiple models should request customer reference architectures from Google’s engineering team and validate the performance numbers against representative internal workloads — particularly KV cache utilization for long-context agent conversations — before committing to TPU-based inference at scale. DeepSeek V4’s recent 98% price advantage over proprietary frontier models is a useful forcing function: if open-weight inference at commodity prices continues to compress margins, the value proposition of Google’s custom silicon must be realized in performance-per-dollar, not just raw throughput.

The partner fund’s embedded-engineer component is the most structurally significant offer for organizations that lack internal ML operations capability. Access to Google’s forward-deployed engineers during a platform launch is the fastest path to production-ready agent architecture — and the most effective way to build institutional knowledge that does not leave when the engagement ends. Teams evaluating which hyperscaler’s agentic program to anchor on should ask specifically whether the forward-deployed engineer program is available, funded, and staffed — not just listed in the press release.

Operator checklist for teams evaluating Gemini Enterprise Agent Platform:

• Map governance requirements before model selection. The platform’s multi-model architecture is only an advantage if your security and compliance review can accommodate the resulting audit surface. Identify which regulatory requirements apply to your agent deployment — GDPR, HIPAA, SOC 2, FedRAMP — and verify which model-deployment combinations within the platform are certifiable before committing to a specific stack.

• Benchmark TPU 8i on your specific inference profile. The 80% performance-per-dollar claim is a best-case figure for workloads that match Google’s benchmark design. Test your actual token throughput, KV cache utilization, and latency requirements for long-context agent conversations against both TPU 8i and comparable GPU-based inference before making a hardware commitment.

• Treat the $750M fund as a procurement accelerator, not a selection criterion. If you are already evaluating Google Cloud for agent deployment, the embedded-engineer program is a real resource worth requesting. Do not let its existence shift your vendor decision — AWS Bedrock AgentCore and Azure AI Foundry are both credible alternatives that may fit your existing infrastructure better.

• Pilot long-running agents on actual business workflows, not benchmarks. The platform’s long-running agent capability is technically impressive but operationally unproven at enterprise scale. Start with a workflow that has clear success metrics, bounded error recovery, and human-in-the-loop review built in. Do not deploy long-running agents in high-stakes autonomous workflows — contract management, financial execution, supply chain ordering — without a full failure-mode analysis and a tested rollback plan.

• Negotiate multi-year commitments only after validating Model Garden multi-tenancy. The ability to switch between Gemini, Claude, and open models within a single platform is a significant flexibility advantage — but only if your organization can operationalize model-switching without rebuilding integration pipelines. Request a proof-of-concept demonstrating model substitution in your specific agent architecture before locking in platform pricing.

• Account for the Vertex AI → Gemini Enterprise Agent Platform nomenclature shift in security reviews. Existing Vertex AI compliance documentation, vendor assessments, and contract language will need update cycles. This is administrative overhead, not a technical barrier, but it is real time that procurement and security teams will spend before a new agent deployment can be approved internally.

The agentic enterprise era that Google is describing at Cloud Next ‘26 is arriving — the token-processing metrics confirm that the underlying demand is real, and the $240 billion backlog confirms that enterprises are making long-term commitments to Google Cloud infrastructure. Whether those commitments materialize into agent-layer lock-in for Google specifically, or whether the governance layer remains split across Azure, AWS, and independent platforms, is the open question that the next three years will resolve. Google enters that competition with the best custom silicon in the market, a credible multi-model platform with 200+ models, a $240 billion backlog proving enterprise buyers are serious, and a $750 million fund designed to make the decision easier. The gap between that offer and the gravitational pull of Microsoft’s enterprise bundle is the market structure that 2026 will begin to rewrite.

In other news

AWS launches Bedrock AgentCore at scale. Amazon shipped Amazon Bedrock AgentCore — a managed, framework-agnostic platform for deploying and operating AI agents at production scale — in the same week as Google Cloud Next, directly intensifying the enterprise agent infrastructure competition. The service handles memory management, multi-step orchestration, security policy enforcement, and infrastructure provisioning without requiring teams to manage underlying compute, and supports popular agent frameworks including LangGraph, CrewAI, and Strands (AWS Machine Learning Blog).

Microsoft’s 2026 Release Wave 1 embeds agentic AI across the Office stack. Microsoft’s spring wave of enterprise software updates deploys autonomous agent behavior across Dynamics 365, Power Platform, and Microsoft 365 Copilot — including “always-on” agents in Office applications that monitor workflows without user initiation. A new AI Agent Builder Certification for IT administrators enables enterprise teams to audit and govern agent deployments at the organizational level, launching for enterprise customers in May 2026 (Cloud Wars).

Shield AI wins Navy slot in $800M ISR competition. The San Diego defense AI startup — which raised $1.5 billion at a $12.7 billion valuation in March — was selected by the U.S. Navy to compete for up to $800 million in contractor-operated intelligence, surveillance, and reconnaissance services using its V-BAT autonomous drone system. Shield AI joins AeroVironment, Insitu, and Textron on the multi-vendor vehicle; individual delivery orders are competed rather than pre-awarded. The selection validates Shield AI’s V-BAT as Navy-qualified hardware and positions the company for recurring task-order revenue through the program’s ceiling (TechCrunch).