August 2026 in review: AI spend gets itemized, CloudFlow goes truly multi-cloud, Kubernetes gets guardrails

Know what your AI actually costs

The pattern is the same across every new integration: connect the account, and cost shows up in Cloud Analytics with the dimensions that matter for that provider, ready for budgets, alerts, allocations, and anomaly detection.

• OpenRouter. Billed cost per model and per hosting provider. If you route one workload across three providers for latency or price, you can now see which route the money took.
• xAI. Grok API cost by model, usage type, and API key. The API key dimension is the one teams ask for, because that is how you tie spend back to a service.
• Fireworks AI. Cost and usage by model, deployment type, and GPU, including budgets and attribution. Dedicated deployments and serverless are visible side by side.
• Snowflake and Databricks, attributed deeper. GenAI Intelligence now shows who spends what on which AI product, from individual Snowflake queries to Databricks workloads. Shared warehouse, per-team answer.
• Anthropic spend reports, no manual conversion. DataHub schema templates ingest provider-native CSV files directly. Export from the Anthropic console, upload, done.
• The self-hosting question, answered with data. compute.doit.com now simulates the "should we self-host an open model or keep paying per token" decision three ways: size a deployment, value hardware you already own, or paste your Anthropic or OpenAI bill and get a split verdict on what maps to an open model at parity. Everything behind it, GPU pricing gathered daily, vLLM throughput benchmarks, the closed-to-open equivalence map, is also a free, keyless REST API with an llms.txt entry point, so your agents can run the same analysis in four GET requests.

CloudFlow: more clouds, more control flow

CloudFlow's job is to turn FinOps runbooks into automation. August widened both what it can reach and how it can reason.

• Oracle Cloud support. The same flows you run on AWS, Google Cloud, and Azure now work against Oracle Cloud.
• Azure support (early preview). Azure actions and triggers alongside the other clouds.
• OpenAI and Anthropic Admin APIs. Spend guardrails, API key hygiene, and usage reporting for your LLM providers, as flow nodes. Rotate stale keys on a schedule, or page someone when daily spend crosses a line.
• Switch node. Route each result down the right path based on a field's value. Multi-way decisions no longer mean stacking Branch nodes.
• Parallel paths node. Split a flow into independent paths so unrelated tasks stop sharing one sequence, and a failure in one path leaves the others running.
• Datastore runs SQL. Query, reshape, and maintain Datastore tables with PostgreSQL inside a flow. Joins, aggregations, and cleanup jobs without exporting anywhere.
• Email triggers. Every published flow can have its own email address. Send an email to it, and the flow runs. Useful for wiring in anything that can send mail, which is everything.
• Export and import flows (preview). Two API operations and matching CLI commands serialize a flow, including subflows, into one JSON bundle with no credentials or tenant state, and import it into another tenant as draft flows. Build once, promote from staging to production, or hand a working flow to a sister team.

See it, slice it, catch it: Cloud Analytics

• Trend detection, rebuilt. The "what's trending" analysis now surfaces changes that actually matter, designed for people and agents alike. In our own testing the new detector cut noise dramatically while keeping the real movers.
• Split shared costs by usage metrics. Allocation drivers can now come from observability platforms like Datadog or Prometheus, so shared costs follow actual consumption instead of a fixed percentage that was accurate last quarter.
• Anomaly triage workflow. Review statuses, comments, and activity history on every cost anomaly. Detection was never the hard part; knowing who looked at it and what they decided was.
• Dashboard filters by service, project, region, and more. Dashboards previously filtered by Allocations. Now they filter by the dimensions many teams actually think in.
• Duplicate any widget in one step. Building dashboards usually means variations on a theme: same report, different time range or group-by. Duplicate and adjust instead of rebuilding.
• Cross-period trend charts for daily and weekly budgets. Track per-period spend against budget across the last 30 days or 16 weeks, so a budget is a trend line rather than a single month's verdict.
• Cloud Intelligence™ in Grafana. Live cloud cost data in the dashboards you already run, available now for all Cloud Intelligence™ customers. Correlate spend with the operational metrics next to it, and alert on cost the same way you alert on latency.

PerfectScale: Kubernetes optimization with better judgment

• Suspected CPU runaway detection. When repeated CPU increases fail to resolve resource pressure, that is a workload problem, not demand. PerfectScale now flags those containers and pauses further increases at configurable thresholds you set in automation YAML.
• Full-mode ResourceQuota support. Workloads in quota-governed namespaces can now be right-sized up to the quota caps, so governance stops silently capping optimization.
• Unevictable pods, now fixable via API. Unevictable pods block node scale-down and quietly limit cluster elasticity. They are now visible and fixable programmatically.
• InfraFit public API. Node utilization, efficiency data, and infrastructure optimization insights, available to your own tooling.
• Filter workloads by JVM runtime. Find every Java workload in a cluster in one click, no labels required. Java memory behavior is its own discipline, and now it has its own filter.
• Google Cloud flexible CUDs in PerfectScale for Commitments. Continuous usage analysis and a laddered purchase plan for Compute Engine, GKE, and Cloud Run flexible commitments, executed with approvals or autonomously within guardrails you define. Cloud SQL support is coming next.
• Numbers that read like your locale writes them. Small, but if you have ever misread 1.000 as one, you know why it shipped.