DEV Community: Gísli Guðmundsson

FinOps for AI Workloads on Azure: Tokens, PTU and Pay-As-You-Go

Gísli Guðmundsson — Tue, 27 Oct 2026 11:18:34 +0000

Originally published on Substack.

Part 3 of my Azure AI landing zone series is about FinOps for AI workloads. Tokens are the AI bill most teams actually get, so the post covers:

Picking the deployment type first: Global Standard, Data Zone, Batch and Flex (preview)
Why quota isn't billing, and why max_tokens eats your TPM
PTU vs pay-as-you-go, and when hourly PTU actually pays off
Sizing PTU with Microsoft's formula, and how prompt caching helps
Spillover: PTU for the baseline, pay-per-token for the peaks
Showback per team with llm-token-limit and llm-emit-token-metric in API Management
Cost Management: group by meter, tag your deployments, and remember budgets only alert

Read the full article on Substack: https://gisli78.substack.com/p/finops-for-ai-workloads-on-azure

Bicep sample: https://github.com/gislig/azure-ai-landing-zone/tree/main/examples/ai-finops

GPU Quotas, VM Sizes and Cost for AI Workloads on Azure

Gísli Guðmundsson — Mon, 19 Oct 2026 10:09:16 +0000

Part 2 of my Azure AI landing zone series is out on Substack.

Before anyone asks for an H100, this post covers what to check: whether you need your own GPUs at all (Foundry pay-per-token, PTU, managed compute vs. your own VMs or AKS), the current Azure GPU VM families and what's retired, how vCPU family quota works and how to check it with az vm list-usage, az quota and az vm list-skus, what Spot, capacity reservations, reservations and savings plans actually guarantee, and guardrails with Azure Policy, budgets by tag and stopping idle GPUs. It also covers the regional VM price increase coming on 1 February 2027.

Read the full article: https://gisli78.substack.com/p/gpu-quotas-vm-sizes-and-cost-for

Bicep sample (policy + budget): https://github.com/gislig/azure-ai-landing-zone/tree/main/examples/gpu-guardrails

A Secure Azure Landing Zone for AI Workloads

Gísli Guðmundsson — Mon, 12 Oct 2026 16:41:51 +0000

Before the first model call, I lock down the platform: Azure AI Foundry (AIServices) behind a private endpoint with the three required private DNS zones, local auth disabled in favour of managed identity, and Azure API Management as the AI gateway using llm-token-limit and llm-emit-token-metric. Everything is in Bicep with automated tests.

👉 Read the full article on Substack: https://gisli78.substack.com/p/a-secure-azure-landing-zone-for-ai

💻 Code: https://github.com/gislig/azure-ai-landing-zone

Tags: #mvpbuzz #azure #azug #microsoft