Taming Azure’s Two Biggest Bills: Logs & AI

Issue #02 · FinOps Uncensored • Production Reality Briefing

• AWS Slashes Amazon EMR on EKS Pricing by 30%
• Cut Microsoft Foundry AI Costs with PTU Reservations
• Actionable Playbook · Discovery Script: KQL query to audit top billable tables in Log Analytics (How to switch and safety checklist)
• 30-Day TCO Benchmark: 100 GB/day across Analytics PAYG, 100 GB Commitment, and Basic Logs

AWS has officially lowered the EMR surcharge for Amazon EMR on EKS by 30%, making containerized big data analytics significantly more cost-competitive against EC2-based EMR deployments and self-managed Apache Spark clusters on Kubernetes.

🔗 AWS What's New Announcement →

Slash hourly AI compute costs in Microsoft Foundry by locking in 1-month or 1-year PTU Reservations (available for Global, Data Zone, or Regional SKUs across OpenAI, DeepSeek, and Llama models).

PAYG vs. PTUs in Microsoft Foundry: How to Scale GenAI and Slash Costs with PTU Reservations

As organizations move generative AI workloads from proof-of-concept into customer-facing production, one of the biggest architectural and FinOps decisions in Microsoft Foundry (including Azure OpenAI, DeepSeek, and Llama models) is choosing the right consumption model: Pay-As-You-Go (Standard) vs. Provisioned Throughput Units (PTUs)—and knowing when to layer Azure Reservations on top. Here is everything FinOps and AI engineering teams need to know.

When deploying models in Microsoft Foundry, you choose between two fundamentally different architectures:

A Provisioned Throughput Unit (PTU) is a standardized, model-independent unit of model processing capacity.

• You allocate a block of PTUs to a deployment, and Microsoft Foundry reserves that compute capacity exclusively for your endpoint.
• Model-Independent Pool: You don't buy PTUs tied to a single model version. The same PTU quota and reservation pool can cover GPT-4o, DeepSeek-R1, or Llama deployments. When you upgrade to a newer model in the future, your existing PTU reservation automatically covers it.

Once you decide a workload needs PTUs, there are two ways to pay for them:

PTU Hourly Billing (On-Demand):

• You pay the list $/PTU/hr rate, prorated to the minute (e.g., a 100-PTU deployment running for 15 minutes bills as 25 PTU-hours).
• When to use it: Short-term benchmarking, load testing, or temporary event spikes (like a 48-hour hackathon).
• Why NOT to use it for production: Leaving production PTUs on hourly billing is the most expensive way to run AI in Azure. Furthermore, scaling PTUs down at night and scaling them back up in the morning is risky: when you scale down, you release that physical GPU capacity back to the regional pool, and there is no guarantee it will still be available when you try to scale back up the next morning.

Microsoft Foundry PTU Reservations (1-Month or 1-Year):

• By committing to a 1-month or 1-year Azure Reservation (in 1-PTU increments, paid upfront or monthly), you unlock a heavily discounted effective $/PTU/hr rate compared to hourly billing.
• Any PTUs deployed up to your reserved quantity are covered by the reservation; if you temporarily scale above your reserved PTUs, only the excess (overage) is billed at the hourly rate.

📘 Microsoft documentation: Provisioned throughput billing & reservations → · What is provisioned throughput? →

Azure Log Analytics tables default to the Analytics tier (~$2.30/GB list price in East US). However, noisy debugging tables rarely require complex KQL joins or multi-row alerting. By switching eligible tables to the Basic Logs plan, you can cut ingestion costs by ~78% ($0.50/GB) while keeping the same 30-day interactive retention window.

Note: Quantity in the Usage table is measured in MB, so dividing by 1000 converts it to billable GB.

Run this in Azure Monitor Logs. If supported tables like ContainerLogV2, AppTraces, or custom _CL tables account for >50% of your volume and are only used for ad-hoc debugging, they are prime candidates for the Basic Plan.

This drops ingestion billing from $2.30/GB down to $0.50/GB (Pay-As-You-Go) for all newly ingested data on that table without interrupting your log pipelines.

• Scheduled Query Alerts & KQL Operators: Tables on the Basic Plan do not support multi-row aggregated Scheduled Query Alerts (they only support single-row Simple Log Alerts), and each query can only read from that one table—joins with other tables are restricted. Verify that no P1/P2 incident rules or Azure Workbooks depend on aggregated alerts or cross-table joins from this table before switching.
• Watch Out for Query Scan Charges ($0.005/GB): Unlike Analytics tables (where queries are free), Basic Logs charge $0.005 per GB of data scanned. Remove Basic tables from auto-refreshing Grafana dashboards or automated polling scripts.
• Retention (30 Days Included + Long-Term Retention): Basic Plan tables include a fixed 30-day interactive retention window (upgraded from the old 8-day limit). For compliance requirements (e.g., 365 days), pair it with Azure Monitor Long-term retention ($0.02/GB/mo).
• Weekly Plan Switch Cooldown: Azure enforces a 7-day waiting period between plan changes on the same table (and you can only change a table's plan once per week). Double-check your dependencies before running the command.

Most FinOps teams try to tame runaway Azure Log Analytics bills by purchasing a Commitment Tier. But locking in a 100 GB/day commitment ($196/day) only trims your bill by ~13.4%. If the bulk of your ingestion is unindexed stdout (like AKS ContainerLogV2) or high-frequency application traces, reclassifying those tables to the Basic Logs plan cuts the same workload from $7,548.50 down to $2,230.00/month—saving $63,822 every year.

Below is the line-by-line 1-year TCO benchmark for a 100 GB/day (3,000 GB/month) baseline stored for 12 months (30 days interactive + 11 months long-term retention), factoring in 20 ad-hoc 7-day diagnostic queries per month:

Attending Microsoft Ignite in San Francisco? Let's grab a coffee and talk Azure pricing, reservation lock-in, and how to protect your cloud budget.

👉 Let's talk →

To cleaner architectures and lower cloud bills,
The CloudHiro Team