AWS
Microsoft Azure
Google Cloud
Short answer
The most important Cloud Next ’26 updates for FinOps teams are the ones that change billing baselines or commitment decisions. Agent-related services introduced additional billable components, BigQuery fluid scaling changed autoscaling billing behavior, new VM families affect CUD coverage, and newer TPU systems changed AI unit economics.Teams should re-baseline actual spend before renewing commitments rather than treating provider performance claims or product announcements as direct savings forecasts.
What changed around Cloud Next ’26
Cloud Next ’26 took place April 22 to 24, 2026. Several of the announcements affect decisions that FinOps teams may still be working through months later, while other infrastructure lifecycle changes became relevant after the event.Separating the original announcement from the action it creates today helps teams focus on financial exposure rather than announcement volume.
| Date | Update | FinOps significance |
|---|---|---|
| April 22–24, 2026 | Google Cloud Next ’26 | New AI infrastructure, agent platform capabilities, VM families, and data-platform changes entered the planning cycle. |
| April 22, 2026 | BigQuery fluid scaling announced as GA | Eligible autoscaling reservations can use per-second billing without the default one-minute minimum. |
| September 15, 2026 | NVIDIA P100 reached end of support | Remaining P100 workloads require migration to supported accelerator options. |
| August 1, 2027 | T4 and P4 scheduled end of support | Teams still using these GPUs have a future migration deadline to include in infrastructure planning. |
Gemini Enterprise Agent Platform changes AI cost planning
Google introduced Gemini Enterprise Agent Platform at Cloud Next ’26 as the evolution of Vertex AI for building, scaling, governing, and operating enterprise agents.For FinOps teams, the more immediate issue is billing. According to Google Cloud’s Vertex AI release notes, Sessions, Memory Bank, and Code Execution began charging on January 28, 2026.
Google Cloud’s current Free Tier documentation includes the first 180,000 vCPU-seconds and 360,000 GiB-seconds of Agent Runtime on Gemini Enterprise Agent Platform per month. Teams should verify current thresholds against the Free Tier documentation before forecasting production workloads.
The cost model is broader than runtime alone. An agent workload may combine:
Agent Runtime
Sessions
Memory Bank
Code Execution
Model inference
Storage
Network usage
APIs and supporting Google Cloud services
A practical FinOps approach is to calculate both total spend and a workload-level unit metric such as cost per completed workflow, cost per request, or cost per successful automation.
TPU 8t and TPU 8i separate training and inference
Google split its eighth-generation TPU architecture into two systems aimed at different AI workloads.TPU 8t is designed for large-scale model training. Google says a TPU 8t superpod can scale to 9,600 chips and provide 121 exaflops of compute.
TPU 8i is designed for inference and reinforcement-learning workloads. It includes 384 MB of on-chip SRAM and 288 GB of HBM.
According to Google’s Cloud Next ’26 AI infrastructure announcement, TPU 8t delivers nearly 3x the compute performance of the prior generation, while TPU 8i provides up to 80% better inference performance per dollar.
That 80% figure should not be treated as an 80% reduction in an individual company’s inference bill.
Actual economics depend on factors such as:
Model architecture
Batch size
Accelerator utilization
Latency requirements
Throughput
Network usage
Supporting infrastructure
Commitment coverage
BigQuery fluid scaling changes the baseline
BigQuery fluid scaling was announced as generally available at Cloud Next ’26 on April 22, 2026. Google describes the feature as providing true per-second billing and reports up to 34% cost savings for autoscaling workloads. (Google Cloud)When enabled for an eligible autoscaling reservation, fluid scaling removes the default one-minute minimum for autoscaling-reservation billing and bills autoscaled capacity per second with no minimum duration.
This distinction is important. Billing granularity should not be interpreted as a guarantee that BigQuery capacity itself will resize every second.
The savings figure is also not a guaranteed result for every environment.
The impact depends on workload shape. Short, bursty workloads can respond differently to the removal of a minimum billing duration than workloads that sustain autoscaled capacity for longer periods.
For FinOps teams, the commitment implication matters as much as the immediate billing change.
If a BigQuery commitment was sized using an older run rate, that historical baseline may no longer represent future eligible spend after billing behavior changes.
C4N and M4N change CUD planning
C4N and M4N expand Google Cloud’s Compute Engine portfolio for network-intensive and memory-intensive workloads.The important FinOps issue is commitment compatibility.
Google’s resource-based CUD documentation explains that hardware commitments apply to specific eligible machine series and regions.
That means moving a workload from an older series to C4N or M4N does not automatically mean the existing resource-based commitment follows it.
Before migrating, compare:
Current machine series
Existing CUD type
Target machine series
Region
Remaining commitment term
Eligible spend after migration
For a deeper explanation of how commitment types differ, see our GCP Committed Use Discounts guide.
GKE and infrastructure updates need workload testing
Cloud Next ’26 also introduced changes across GKE, Axion N4A, networking, and agent infrastructure.These announcements can improve price-performance, but better technical efficiency does not automatically equal lower total spend.
A workload can run faster while costing more per hour. Higher node density can reduce infrastructure consumption, but only if the application and scheduler actually make use of that capacity.
The same principle applies to networking improvements. Removing infrastructure overhead may reduce compute demand, but teams should measure actual node requirements before adjusting commitment coverage.
For major infrastructure migrations, benchmark the workload first and change the financial baseline second.
GPU lifecycle dates now affect planning
NVIDIA P100 GPUs reached end of support on Google Cloud on September 15, 2026.Google documents the retirement through its P100 end-of-support guidance.
Google Cloud has also announced that NVIDIA T4 and P4 GPUs are deprecated and are scheduled to reach end of support on August 1, 2027.
The cost decision should not be reduced to choosing whichever replacement GPU has the lowest hourly price.
Before migration, compare:
Application compatibility
Performance requirements
Regional capacity
Availability
Per-workload cost
Migration effort
Existing commitments
Commitment eligibility for the target infrastructure
Which updates require action now?
Not every Cloud Next announcement deserves the same level of FinOps attention.| Update | Recommended response | Why |
|---|---|---|
| Agent-related billing | Review now | Additional agent services have been billable since January 2026 and may already appear in cloud bills. |
| BigQuery fluid scaling | Review now | Billing behavior can change the baseline used for commitment sizing. |
| C4N and M4N | Review before migration | Machine-family changes can affect resource-based CUD coverage. |
| TPU 8i and TPU 8t | Pilot when relevant | Provider performance claims should be validated against real workloads. |
| GKE infrastructure improvements | Pilot and measure | Cost effects depend on node density, utilization, and workload architecture. |
| T4 and P4 lifecycle | Plan migration | August 1, 2027 creates a future infrastructure deadline. |
| Other Next ’26 announcements | Monitor | Not every product release changes current cloud economics. |
What FinOps teams should do now
Use this checklist to turn the announcements and lifecycle changes into an operating review:Pull agent-related billing by SKU and compare current costs with the pre-January 28 baseline.
Separate Agent Runtime, Sessions, Memory Bank, Code Execution, model inference, storage, and network costs.
Identify BigQuery reservations currently using autoscaling.
Confirm whether fluid scaling is enabled for relevant reservations.
Recalculate the BigQuery baseline before renewing long-term commitments.
Map existing resource-based CUDs by machine series and region before moving workloads to C4N or M4N.
Benchmark TPU 8i or other accelerator changes using workload-level unit economics.
Add the August 1, 2027 T4 and P4 support deadline to infrastructure planning where those GPUs remain in use.
How Usage.ai helps manage GCP commitments
We analyze connected GCP billing data and existing commitments to identify additional eligible commitment opportunities.We help teams identify and manage eligible GCP commitment opportunities based on actual usage patterns.With Flex Insured Commitments, teams can capture commitment-based savings while reducing the risk of unused capacity. Our Flex Commitments reduce cloud spend by 30–50% on average across covered workloads, while qualifying commitments include cashback protection for eligible underutilization.
For GCP teams, we analyze billing data to identify stable usage that may be suitable for commitments, help determine an appropriate commitment level, and support eligible commitment actions.
This helps teams capture more of the savings available through GCP commitments without taking on the traditional risk of unused capacity.
Frequently asked questions
Does BigQuery fluid scaling apply automatically?
No. Eligible autoscaling reservations need the appropriate fluid-scaling configuration. Google’s current BigQuery documentation explains the relevant autoscaling and fluid-scaling behavior.
Will my existing resource-based CUD automatically cover C4N or M4N?
Not necessarily. Resource-based CUDs apply to specific eligible machine series and regions. Review the exact commitment and target machine family before migrating rather than assuming the commitment will move with the workload.
When did Sessions, Memory Bank, and Code Execution start charging?
Google Cloud’s release notes state that these additional services began charging on January 28, 2026.
Does 80% better TPU 8i price-performance mean an 80% lower inference bill?
No. The figure is Google’s performance-per-dollar comparison with the previous TPU generation. Realized cost depends on workload performance, utilization, latency requirements, architecture, scaling behavior, and supporting infrastructure.
Is NVIDIA T4 already retired on Google Cloud?
No. T4 and P4 are deprecated, with end of support scheduled for August 1, 2027. NVIDIA P100 reached end of support on September 15, 2026.
