LATEST
AI Tools

Azure HorizonDB Goes Live: Postgres for AI Apps at $0.37 an Hour

Updated November 19, 2026
12 min read
Editor's Note: This article was human-reviewed for accuracy
Server racks in a data center aisle with a monitoring workstation

Quick Answer & Key Takeaways

  • • Microsoft made Azure HorizonDB generally available at Ignite on November 17, 2026, in 13 Azure regions.
  • • The PostgreSQL-compatible database starts at $0.37 an hour for 2 vCores in East US, about twice the compute rate of Azure Database for PostgreSQL Flexible Server.
  • • Its pitch is AI inside the database: vector search, graph queries, and model calls from SQL, but several headline features, including Knowledge Graphs and customer-managed keys, are still in preview.

What Azure HorizonDB actually is

Microsoft made Azure HorizonDB generally available at its Ignite conference on November 17, 2026, after a year in preview. It is a fully managed, PostgreSQL-compatible database service built on a different architecture from the PostgreSQL Microsoft already sells on Azure, and it is aimed squarely at the workloads AI applications create: agents that spin databases up and down, retrieval pipelines that need vector search next to transactions, and apps that query relationships as well as rows.

The core idea is a database-as-log design. Compute and storage are separated, transactions commit in a single step to shared write-ahead log storage, and read capacity scales out without touching PostgreSQL replication. Microsoft says every database is resilient across availability zones by default, with sub-millisecond commit latency and failover in under five seconds. Databases can grow to 128 TB, and a single compute node scales to 192 vCores.

That is a different product from Azure Database for PostgreSQL Flexible Server, which runs open-source PostgreSQL in the conventional way. Microsoft keeps both: Flexible Server is the service for running PostgreSQL as it exists today, while HorizonDB is the rebuilt version for mission-critical and AI-heavy workloads. The distinction matters for the price, which is roughly double.

What went generally available, and what did not

The GA label covers the core service: the managed database, zone-resilient storage, read scale-out, Entra ID authentication, and the AI search and model features described below. It is live in 13 Azure regions: Canada Central, Central US, East US, South Central US, West US 2 and West US 3 in the Americas; Germany West Central and Sweden Central in Europe; and Australia East, India South Central, Indonesia Central, Japan East and Korea Central in Asia Pacific. Microsoft says more regions are coming, but it has published no dates.

Several headline features are not GA yet, and the launch materials are easy to misread on this point. Knowledge Graphs, the feature that turns documents into entity maps inside PostgreSQL, is in public preview. AI Model Management, which connects the database to models in Microsoft Foundry, is also in preview. Customer-managed encryption keys are in public preview too, according to Microsoft Learn, even though the launch blog lists them among the security capabilities. Private Link private endpoints are in preview as well. Monitoring through Azure Monitor alerts and the Log Analytics integration did reach general availability alongside the service.

The other half of the announcement is Microsoft Fabric HorizonDB, a separate, serverless PostgreSQL experience inside Microsoft Fabric. It is in public preview, not GA. It provisions databases in seconds, pauses compute after five minutes of inactivity, and resumes on the next query. It also brings translytical queries that run PostgreSQL queries on Fabric Data Warehouse, mirroring of tables into OneLake in Delta format, and synced tables that flow enriched warehouse data back into the database. Do not confuse the two: the serverless auto-pause belongs to Fabric HorizonDB, not to Azure HorizonDB.

What it costs

HorizonDB is pay-as-you-go with no free tier and no published reserved or savings-plan tier. Compute is billed per vCore per hour, and you pay for the compute you provision on the primary instance and on high-availability replicas alike, so turning on high availability roughly doubles the compute line. Storage for data and logs is billed per GiB per month. Backup storage is free up to 100% of the initial cluster size, and extra backup storage is charged per GiB per month. There is no extra HorizonDB charge for model management; model usage is billed by Microsoft Foundry at its standard rates and appears on the HorizonDB invoice. All prices below are the East US rates from Microsoft's own pricing page, read on November 19, 2026. Rates vary by region, and Microsoft states they are estimates, not quotes.

Compute sizePer hour (USD)Per month, 730 hours (USD)Per month (GBP, approx.)Per month (CAD, approx.)
2 vCores$0.37$266.68£202.52CA$382.26
4 vCores$0.73$533.37£405.04CA$764.53
8 vCores$1.46$1,066.73£810.08CA$1,529.05
16 vCores$2.92$2,133.47£1,620.16CA$3,058.11
32 vCores$5.85$4,266.94£3,240.32CA$6,116.23
64 vCores$11.69$8,533.88£6,480.63CA$12,232.46
192 vCores$35.07$25,601.63£19,441.88CA$36,697.38

Storage adds $0.25 per GiB per month in East US, and backup storage beyond the free allowance costs $0.10 per GiB per month. A small example makes the shape of the bill clear: a 2-vCore cluster with 100 GiB of storage costs about $266.68 for compute plus $25.00 for storage, or roughly $291.68 a month before any replica, backup overage, or model usage. Add a high-availability replica and the compute line doubles to about $533.36, before storage.

Regional pricing is not flat. The same 2-vCore node costs $0.45665 an hour in UK South, about 25% more than East US, and $0.438384 in Canada Central. Anyone budgeting outside the US should read their own region's column on the pricing page rather than converting the US figure. Exchange rates used above are November 19, 2026 market rates: 1 USD = 0.7594 GBP and 1 USD = 1.4334 CAD. They are approximations, and Azure bills in USD before conversion.

One cost detail catches people out. Azure HorizonDB itself has no scale-to-zero mode: provisioned compute is billed while it is provisioned, even when the database sits idle. The auto-pause after five minutes of inactivity is a Fabric HorizonDB feature. Short-lived agent workloads that spin up databases for a prototype and drop them will pay for the hours they hold, not just the queries they run.

HorizonDB vs Azure Database for PostgreSQL Flexible Server

The natural comparison is with the PostgreSQL service Microsoft already sells. The table below uses East US pay-as-you-go compute rates from Microsoft's own pricing pages for both services, read on November 19, 2026. Flexible Server storage costs $0.115 per GiB per month in East US, against HorizonDB's $0.25.

Azure HorizonDBFlexible Server, General Purpose (Ddsv5)
2 vCores, per hour$0.37$0.18
8 vCores, per hour$1.46$0.71
16 vCores, per hour$2.92$1.42
32 vCores, per hour$5.85$2.85
Storage, per GiB/month$0.25$0.115
Max vCores per node192128 (Ddsv6 series)
ArchitectureDatabase-as-log, decoupled compute and storageConventional managed PostgreSQL
Zone resilienceBy defaultOptional zone-redundant HA
Built-in AI search and graphYes: vector, BM25, hybrid search, Apache AGEExtensions available; no built-in AI platform
Burstable low-cost tierNoYes, from $0.068 an hour (B2s, 2 vCores)
Stop/start to pause billingNoYes
StatusGenerally available November 17, 2026Generally available for years

The pattern is consistent: HorizonDB costs about 2.05 times Flexible Server's General Purpose compute at every size, and about 2.17 times for storage. What that premium buys is the rebuilt architecture, the higher ceilings, zone resilience by default, and the AI features built into the database rather than bolted on. For a conventional application that does not need those, Flexible Server remains the cheaper PostgreSQL on Azure, and its burstable tier starts far below HorizonDB's smallest node. For an AI workload that would otherwise pay for a separate vector database, a graph database, and an orchestration layer, the comparison is less one-sided, but Microsoft has published no worked example of that saving, so it should be tested rather than assumed.

The AI features, and how much to trust the numbers

The launch blog frames HorizonDB as a database for agents, and the feature list follows that pitch. Vector search runs inside PostgreSQL with pgvector support and DiskANN indexing, which Microsoft says delivers up to 1.5 times the throughput of HNSW at comparable recall with indexes up to five times smaller. BM25 keyword ranking comes through pg_textsearch, and Reciprocal Rank Fusion functions combine vector and keyword results into hybrid search. Graph modeling runs through Apache AGE, with Microsoft reporting simple reads up to 3,900 times faster and multi-hop queries up to 512 times faster than community Apache AGE on the LDBC SNB Interactive workload, and MERGE and batch upserts up to 190 times faster with about six times lower peak memory in its own internal testing.

AI Model Management connects applications to models in Microsoft Foundry from the database, AI Functions invoke models from SQL for embedding generation, extraction, and content generation, and AI Pipelines compose those functions into durable asynchronous workflows, for example enriching new rows as they arrive. A PostgreSQL MCP Server and PostgreSQL Skills give agents a standard way to discover and use the database, and the PostgreSQL extension for Visual Studio Code adds cluster topology views, AI-assisted query optimization through GitHub Copilot, and knowledge-graph visualization. Google is pushing the same idea from its side with its Data Agent Kit, free MCP tools for its data cloud services.

Every performance figure in the paragraph above is Microsoft's own measurement, and the launch materials say so in the footnotes. The 8-times-faster OLTP claim is the headline example: it comes from Microsoft's internal HammerDB TPROC-C test with 95,000 warehouses and 512 users on 64-vCore Edsv6 virtual machines running PostgreSQL 17.11 in Korea Central. That is a legitimate benchmark configuration, but it is one workload, on one hardware shape, in one region, run by the vendor. The graph and DiskANN figures carry the same caveat. Treat them as a reason to run your own workload against both services, not as a promise about your queries.

Two of the most useful capabilities are the least finished. Knowledge Graphs are in public preview, which Microsoft defines as available for testing but not for production use, and AI Model Management is in preview too. The RAG pipeline story, where chunking through reranking runs in SQL without an application tier, is real for the customers Microsoft quotes, but it depends on preview features for the graph half of it. The parts that are GA today are the database itself, vector and hybrid search, and the model functions and pipelines.

Migration and the Oracle angle

For existing PostgreSQL workloads, Microsoft's path is the new Azure Database Migration Service, which moves databases from AWS, Google Cloud, on-premises servers, and Azure virtual machines to HorizonDB with online migration of schema and data plus validation and monitoring. Oracle migrations get a separate pitch: schema and application conversion tooling built into the PostgreSQL extension for Visual Studio Code, using models in Microsoft Foundry to generate PostgreSQL DDL, flag objects that need manual changes, and convert Oracle-specific dependencies in SQL scripts, stored procedures, and Java code. Offline Oracle data movement is in preview and online migration is still to come, so Oracle shops should treat that tooling as early rather than settled.

PostgreSQL's own role in this story is worth noting. Microsoft says about 19 of its employees actively contribute to the PostgreSQL open-source project, including committers and core-team roles, and it describes itself as one of the top upstream contributors. That does not change the price, but it is relevant context for a service that asks teams to trust a vendor-specific rebuild of an open-source database.

Common mistakes to avoid

First, check the region before planning anything. Thirteen regions is a start, and the UK is not among them: the regions are Canada Central, Central US, East US, South Central US, West US 2, West US 3, Germany West Central, Sweden Central, Australia East, India South Central, Indonesia Central, Japan East, and Korea Central. If your data must stay in the UK, HorizonDB is not an option yet.

Second, budget for the replica. High availability is the default posture, and compute is billed on the primary and on replicas, so the headline hourly rate is half the real compute cost of a resilient deployment.

Third, do not expect the Azure service to pause. The five-minute auto-pause belongs to Fabric HorizonDB. An Azure HorizonDB cluster left running over a weekend bills for the weekend.

Fourth, do not build a production plan on Knowledge Graphs, AI Model Management, customer-managed keys, or Private Link until they leave preview. The core database is GA; the surrounding capabilities are on different timelines, and Microsoft Learn still labels customer-managed keys as preview.

Finally, verify the 8x claim against your own workload. HammerDB TPROC-C is a transaction-processing benchmark, and a vector-heavy or analytics-heavy workload may show a very different gap between HorizonDB and self-managed PostgreSQL.

Who should look at it

HorizonDB makes the most sense for teams already on Azure that run PostgreSQL for AI applications: retrieval pipelines that want vector search beside transactions, agents that need graph traversal over operational data, and workloads where zone resilience and fast failover are requirements rather than extras. The same teams should compare it honestly against Flexible Server, which costs about half as much for plain PostgreSQL, and against self-managed PostgreSQL on virtual machines, which costs less still but leaves the operations to you. The decision turns on whether the built-in AI features replace services you would otherwise pay for separately. If they do not, the premium is hard to justify. If they do, the bill needs to be modeled with the replica, the storage, and the Foundry model usage included, because the hourly compute rate is only the first line.

Microsoft announced HorizonDB alongside a wider Ignite push into agent infrastructure, including Microsoft Perception, its agent risk-scoring system, and the agent runtime work from AWS shows the same direction from the other cloud vendors. The database layer is becoming part of the agent stack rather than a place agents merely read from. HorizonDB is Microsoft's version of that bet, and its price is the first thing to check before buying into it.

Is Azure HorizonDB a replacement for Azure Database for PostgreSQL?

No, not yet. Microsoft sells both. Flexible Server remains the service for conventional PostgreSQL workloads, with burstable tiers, stop/start, and about half the compute price. HorizonDB is the rebuilt service for mission-critical and AI-heavy workloads, with higher ceilings and built-in search, graph, and model features.

How much does Azure HorizonDB cost?

In East US it starts at $0.37 an hour, about $266.68 a month, for 2 vCores, plus $0.25 per GiB of storage per month. The largest node, 192 vCores, costs $35.07 an hour, about $25,601.63 a month. High-availability replicas double the compute line, and model usage is billed separately through Microsoft Foundry.

Which regions is Azure HorizonDB available in?

Thirteen regions at launch: Canada Central, Central US, East US, South Central US, West US 2 and West US 3 in the Americas; Germany West Central and Sweden Central in Europe; and Australia East, India South Central, Indonesia Central, Japan East and Korea Central in Asia Pacific. Microsoft says more are coming but has given no dates.

Is Azure HorizonDB serverless?

No. Azure HorizonDB uses provisioned compute billed per vCore per hour, and it keeps billing while provisioned, even idle. The serverless option with auto-pause after five minutes of inactivity is Microsoft Fabric HorizonDB, a separate service in public preview.

Does Azure HorizonDB support pgvector?

Yes. Existing vector applications work as they are, according to Microsoft, and DiskANN indexing is available for embedding search alongside BM25 keyword ranking and hybrid search that combines both.

What is the 8x performance claim based on?

Microsoft's own internal HammerDB TPROC-C benchmark: 95,000 warehouses, 512 users, 64-vCore Edsv6 virtual machines running PostgreSQL 17.11 in Korea Central, comparing HorizonDB with self-managed PostgreSQL on comparable hardware. It is a vendor-run test, not an independent result.

Microsoft made its PostgreSQL bet at Ignite with a clear price attached: about twice what its existing PostgreSQL service costs, in exchange for a rebuilt architecture and AI features inside the database. The service is live and the rates are published, so the next step is a workload test, not another announcement. Check your region, model the replica, and run your own queries before committing.

Sources

12 min read
39 views
edit_note Written by

Ratul Ahamed

Founder & Editor-in-Chief

Ratul Ahamed is the founder and editor of AI Core Updates, an independent site covering AI tools, models and pricing. He is a web developer based in Bangladesh and also runs All In One Tools, a free collection of browser-based utilities. Contact: [email protected]

Read all stories by Ratul Ahamed arrow_forward
verified_user Human-Reviewed Editorial

forum Comments (0)

Leave a Reply

chat_bubble_outline

No comments yet. Be the first to join the conversation!

Weekly AI Digest

Get our weekly AI tools digest

One email a week analyzing new AI tools, model updates, and practical workflows. Tested and summarized with zero fluff. Unsubscribe any time.

bolt
Hands-on AI Tool Reviews
Pricing math, limits, and alternatives.
verified_user
Strict Anti-Spam Policy
1-click unsubscribe at any time.
mark_email_unread

Subscribe to Digest

Sent weekly • No spam

person_outline
mail_outline

No spam ever. You can unsubscribe with a single click at any time.

cookie

Cookie Settings

We use cookies to enhance your browsing experience, serve personalized content, and analyze our traffic. By clicking "Accept All", you consent to our use of cookies. Learn more