Google Cloud Next 2026: What You Need to Know About AI Costs

Google Cloud Next 2026: The Recap

Feet may have recovered, hydration hopefully the same. No surprise, Google Cloud Next ’26 was madness. We’re talking hundreds of announcements, demos, and partner showcases compressed into a few short days (okay, they felt long at the time!). The keynote alone covered new TPUs, an enterprise agent platform, cross-cloud data architecture, and a $750 million partner fund.

Now that we’re all back home and rested, we’ve taken a step back from the noise. Three themes defined Next ’26 when it comes to costs:

• Agentic AI has crossed from concept to production, and the numbers prove it
• Usage-based pricing is getting sharper across models, compute, and services, which fragments cost attribution further across systems that weren’t designed to talk to each other (of course pricing strategies are far from perfect and will be a discussion that continues throughout the year)
• Data is becoming portable across clouds, but harder to govern financially

Together, they point to the same underlying dynamic: it’s getting easier to build and deploy AI, and harder to maintain financial clarity over what it costs to operate at scale.

The Scale Numbers Are Not Incremental

Google reported that nearly 75% of its cloud customers are now using AI products. More specifically, 330 customers processed more than one trillion tokens each over the past year. And across the platform, models are now handling more than 16 billion tokens per minute via direct API, up from 10 billion in the previous quarter.

We’re no longer talking about incremental adoption. Those numbers clearly reflect a very real change in consumption happening across enterprises with meaningful budgets, active SLAs, and expectations about ROI. It wasn’t too many months ago that AI was considered “experimental.” Now it’s everywhere: production infrastructure, touching core operations, customer-facing workflows, and internal systems.

Agentic AI Has a Different Cost Problem

One of the biggest announcements at Next ’26 was the Gemini Enterprise Agent Platform, Google’s push to make agentic AI production-ready at scale. The platform covers the full lifecycle of agent development: build, deploy, observe, optimize. That’s meaningful progress, but the economic implications of agentic AI are something many organizations haven’t fully worked through yet.

Traditional AI workloads are relatively predictable: a batch job, an inference call, a scheduled pipeline. Agents operate differently. They make repeated, smaller model calls, chain together multiple services, run continuously in the background, and trigger downstream actions across systems.

The cost profile that results is harder to reason about:

• Continuous rather than batch consumption: usage doesn’t end when a job completes
• Non-linear scaling: more agents doesn’t mean proportionally more cost in any predictable way
• Distributed attribution: a single agent action may touch five services, each with its own pricing model

Google’s Agent Platform includes observability tooling designed to provide standardized logging across agents, tools, and API handoffs. Now enterprises have observability to tell them what happened, but there’s still a gap when it comes to what it cost, who owns that cost, or whether it’s generating the return it should. 

Beyond Token Pricing: The Full Cost Stack 

Google’s compute story at Next ’26 centered on its eighth-generation TPUs, framed around unit economics: more performance per dollar, more performance per watt. That framing signals how Google expects AI cost to evolve at the infrastructure level.  
 
In an agent-driven world, the cost isn’t driven by a few large model calls. It’s driven by thousands of smaller decisions: which model to invoke, how often, how efficiently each inference runs, how much data moves in the process. 

Token pricing used to be the thing people watched, but that’s no longer sufficient. 

AI cost is now a combined function of: 

• Inference efficiency at the infrastructure level 
• Model selection and routing 
• Orchestration design and agent architecture 
• Storage and data movement 
• API chaining and third-party service fees 

You’re not managing a line item anymore. You’re managing a system and systems have emergent cost behaviors that individual components don’t predict. 

The Cross-Cloud Lakehouse: Less Lock-In, More to Govern 

The Cross-Cloud Lakehouse didn’t get the loudest reception at Next ’26, but it carries significant implications that extend beyond the data team. It’s Google’s move to standardize on Apache Iceberg and enable zero-copy data access across AWS and Azure environments. 

The stated benefits are straightforward: data can stay where it lives, queries can run across clouds, and migration costs can be avoided. Google also announced compatibility with a wide ecosystem of data platforms including Databricks, Snowflake, Salesforce, SAP, and others, positioning itself as a cross-environment control plane rather than just a cloud provider. 

But zero-copy access across environments doesn’t make the architecture simpler, it makes it more flexible. There’s a difference. You still need: 

• Governance and policy enforcement across environments
• Identity and access management across systems 
• Cataloging, federation, and performance tuning across distributed data 
• Cost attribution when the same data is being queried from multiple places 

The Cross-Cloud Lakehouse reduces friction for AI adoption. It also adds surface area for cost, latency, and governance issues. Organizations will accept that tradeoff, but they need to go in with clear eyes.  

$750 Million and a Clear Direction 

Google also announced a $750 million AI fund to accelerate partner-led agentic AI adoption, a clear signal of where investment is flowing and how urgently the ecosystem is being built out. 

Speaking of which: Mavvrik was included in Google’s roundup of companies building the agentic future alongside names like Notion and Gamma, and we were the only FinOps-focused vendor in that group. That placement reflects where the conversation is heading. As more enterprises move agents into production, the question of how to build and run them is being answered. The question of how to govern the economics of doing so is still catching up. 

The Pattern Across Next ’26 

If there’s one through-line to all of this, it’s this: Google is reducing friction for AI adoption and the underlying system is getting more complex. 

Google is building a powerful platform with a robust ecosystem for enterprises to run AI. This approach makes it easy for customers to purchase adjacent solutions, like AI and cloud cost governance through Google Cloud Marketplace.  Enterprises now have the infrastructure, and like any infrastructure, it needs to be governed financially, not just technically. 

The old models for doing that don’t hold anymore. Traditional FinOps assumes centralized workloads, predictable usage patterns, and clear service boundaries. None of those assumptions survive contact with agentic AI at scale. 

What enterprises need is the ability to see cost across the full stack, spanning cloud, GPU compute, inference, agents, SaaS, data platforms, and on-prem, attributed down to the team, customer, or feature level, and connected to the outcomes that spending is supposed to generate. 

A billing dashboard only accomplishes one piece of the whole. AI cost governance connects the rest. 

What Google Cloud Next ’26 made clear: AI is becoming infrastructure. How enterprises govern the economics of that infrastructure, not just how they run it, will define if their AI spend is a managed investment tied to outcomes they can account for.  

Mavvrik is an AI Cost Governance platform built for the full AI stack  cloud, on-prem, Kubernetes, GenAI services, SaaS, and agentic workloads. See the Mavvrik platform in action here.

Alexa Abbruscato

Certified FinOps Practitioner & Content Strategist

Alexa is a certified FinOps Practitioner and FOCUS Analyst who translates complex concepts into language that resonates across engineering, finance, and procurement.