On the inaugural day of Microsoft Build 2026, held in San Francisco, Microsoft unveiled the public preview of Azure HorizonDB, a fully managed database compatible with PostgreSQL, meticulously re-engineered to cater to the demands of agentic AI workloads. In conjunction with this, the company also announced the limited availability launch of Web IQ, a web-grounding API layer that is already integral to the grounding systems of both Microsoft Copilot and OpenAI’s ChatGPT. For enterprise developers and cloud architects, this dual announcement presents an immediate opportunity: HorizonDB is open for preview signups across five Azure regions as of June 2, while access to Web IQ is available through application for select Azure customers.
At the keynote, CEO Satya Nadella emphasized that the primary bottleneck hindering enterprise AI is no longer the capabilities of models but rather the underlying infrastructure. The unique data call patterns created by agents—characterized by continuous storage, retrieval, reasoning, action, and looping at hundreds of concurrent queries—differ significantly from the predictable read-and-write cycles that traditional databases were designed to handle. HorizonDB and Web IQ aim to bridge this infrastructure gap.
Azure HorizonDB in Public Preview: What Changed Since Ignite
Azure HorizonDB represents a significant architectural shift from Azure Database for PostgreSQL. It employs a “database-as-logs” design, which commits transactions directly to shared write-ahead log (WAL) storage, eliminating the multiple coordination steps typical of traditional PostgreSQL commits. This innovation results in sub-millisecond multi-zone commit latency, a critical advantage for workloads where agents execute numerous read-write cycles each second.
The storage engine is crafted in Rust, a deliberate choice by Microsoft to enhance memory safety. Unlike C-based engines, Rust inherently mitigates various buffer overflow vulnerabilities, which is crucial for databases queried frequently by autonomous processes that lack the error tolerance of human oversight.
In terms of scalability, compute and storage are disaggregated, allowing them to grow independently. Storage can automatically expand to 128 TB, while compute can scale up to 3,072 virtual CPU cores across primary and replica nodes. Stateless compute nodes, which do not store data, facilitate up to 15 read replicas with minimal downtime during maintenance. Microsoft claims that this architecture can deliver up to three times faster transactions and search performance compared to self-managed PostgreSQL, although these figures are based on internal benchmarks and have yet to be independently verified.
Mohsin Shafqat, Director of Software Engineering at NASDAQ, remarked that HorizonDB “brings transactional data, vector search, and AI capabilities into a single platform, which simplifies the architecture without forcing a complete rethink.” Andrew Brust, CEO of Blue Badge Insights and a Microsoft MVP, noted that HorizonDB fills a significant gap in Azure’s offerings, stating, “Now that gap is filled.” However, it is important to note that HorizonDB is not yet serverless; unlike AWS Aurora DSQL, CockroachDB, and YugabyteDB, it requires developers to manually configure compute and manage replicas based on throughput needs, with only storage scaling automatically.
DiskANN and Spherical Quantization: Vector Search Inside the Database
The standout feature of HorizonDB is its native approach to vector search, which sets it apart from Amazon Aurora PostgreSQL and Google AlloyDB. While the latter handle vector workloads through extensions, HorizonDB incorporates vector search directly via DiskANN with spherical quantization. This design, initially developed by Microsoft Research, enables efficient nearest-neighbor vector search over extensive disk-resident datasets without necessitating the loading of entire indexes into memory—an essential capability for enterprise AI workloads that often involve billions of vectors.
Spherical quantization normalizes vector representations and encodes them into compact forms that preserve angular distances, allowing for minimal accuracy loss while significantly reducing memory and storage overhead. This advancement supports vectors with up to 16,000 dimensions, far surpassing the 2,000-dimension limit of HNSW-based indexes in pgvector, making it suitable for high-dimensional models like text-embedding-3-large.
HorizonDB further enhances its capabilities by combining DiskANN with hybrid search, enabling simultaneous retrieval of semantic meaning and keyword relevance through a single query. This integration is particularly beneficial for retrieval-augmented generation (RAG) applications. Additionally, HorizonDB facilitates in-database model invocation via the azure_ai extension, allowing for model inference directly within the PostgreSQL engine through SQL, thus eliminating the need for an external orchestration layer.
Web IQ: Why Microsoft’s AI Grounding Layer Was Already Everywhere
While both announcements are significant, Web IQ stands out as the more strategically important of the two. Jordi Ribas, Microsoft’s president of Search and AI, explained that Web IQ was completely restructured across all layers—indexing, retrieval, ranking, passage selection, and orchestration—to serve inference-time grounding rather than the traditional page-ranking needs of human search.
The architecture comprises five layers, beginning with Bing’s extensive global index, which has been maintained for decades, ensuring freshness and coverage that a specialized index could not match. Above this index lies an embedding model, which Microsoft open-sourced in April 2026, projecting content into a vector space where semantic similarity can be efficiently computed. The retrieval layer utilizes DiskANN-powered distributed partitioning, while content understanding and ranking models tailored for large language model-driven reasoning sit above retrieval. The orchestration layer at the top manages query routing, retrieval fan-out, result merging, and evidence transformation.
A key architectural choice is the unit of information returned by Web IQ. Unlike traditional search APIs that return full documents, Web IQ provides passage-level “structured evidence objects.” This design principle emphasizes that models require information rather than documents, which often serve as inadequate proxies. By focusing on passage-level data, Web IQ enhances the signal-to-noise ratio, resulting in a higher information density per token, thereby reducing costs and improving precision in reasoning under constrained context windows.
Microsoft claims that Web IQ achieves sub-165ms P95 latency across five data centers and is approximately 2.5 times faster at retrieval than its nearest competitor, although these figures are vendor-published. The competing systems in the latency benchmark are categorized as Competitor A through G, and specific configurations may not reflect the full range of production conditions developers would encounter. As of June 3, 2026, no independent third-party evaluations of these claims have been published.
Web IQ is model-agnostic and delivered via JSON-RPC 2.0 transport, making it compatible with any foundation model without locking developers into a specific inference framework. Developers interested in evaluating access can express their interest on Microsoft’s Web IQ product page; however, no public pricing or general availability date has been announced as of June 3.
Microsoft Fabric as the Unified Agent Platform: GPU Analytics and the Database Hub
Both HorizonDB and Web IQ are designed to integrate seamlessly into a broader architecture that Microsoft is developing around Microsoft Fabric, envisioned as a unified data and AI deployment platform. The challenge identified is that each new agent typically starts from scratch, needing to relearn organizational context, data locations, and business rules without a shared foundation. Microsoft proposes Fabric IQ as a solution, providing a context layer that supplies agents with enterprise data through Work IQ (Microsoft 365 data), Fabric IQ (structured business data), and Web IQ (live web intelligence).
A new Database Hub within Fabric, currently in private preview, will centralize the management of Microsoft’s database portfolio—including HorizonDB, Azure Database for PostgreSQL, and Azure Cosmos DB—with data mirrored to OneLake for cross-database analytics. Additionally, Microsoft announced GPU-accelerated analytics for Fabric Data Warehouse, which is set to enter early access preview and roll out in July 2026. Eligible queries will run directly on NVIDIA accelerated computing within the execution engine without requiring query rewrites, with Microsoft claiming performance improvements of up to seven times in internal benchmarks. The underlying research project, CoddSpeed, was honored as Best Industry Paper at SIGMOD 2026, the flagship database conference of the Association for Computing Machinery.
What Developers and Architects Need to Evaluate
For teams currently operating PostgreSQL workloads on Azure, the introduction of HorizonDB prompts a critical evaluation—not necessarily of immediate migration, given that the service is still in preview and carries inherent instability, but rather of whether its architectural design aligns with anticipated agentic workloads. The disaggregated “database-as-logs” model, native DiskANN vector search, and in-engine model invocation present a distinctly different operational profile compared to Azure Database for PostgreSQL Flexible Server, which Microsoft has committed to continuing support for standard PostgreSQL deployments.
IDC research director Devin Pratt observed that major cloud providers are converging on similar PostgreSQL-compatible services that incorporate cloud-native storage, compute, and AI features. This context positions HorizonDB as both a catch-up move and an innovation, particularly since Amazon Aurora pioneered disaggregated database architecture back in 2014. HorizonDB’s unique differentiator lies in its AI-native feature depth, including native DiskANN embedding, in-database model management, and tighter integration with Foundry and Fabric than currently offered by either Aurora or AlloyDB.
Developers interested in assessing Web IQ independently from HorizonDB can do so, as it is model-agnostic and not exclusively tied to Azure. However, the ongoing Federal Trade Commission investigation into Microsoft’s Azure cloud licensing practices, which intensified in February 2026 with civil investigative demands issued to several competitors, is an important consideration for enterprise procurement teams. This investigation focuses on whether Microsoft’s licensing terms create barriers to running Microsoft software outside of Azure, a context that could impact any new dependencies on Azure infrastructure.
A dedicated session at Build 2026 on HorizonDB integration with Rayfin, Microsoft’s new managed backend-as-a-service for agentic applications, is scheduled for June 3 at 4:30 p.m. ET (1:30 p.m. PT).
Frequently Asked Questions
What is Azure HorizonDB and how does it differ from Azure Database for PostgreSQL?
Azure HorizonDB is a fully managed PostgreSQL-compatible database specifically designed for agentic AI workloads. It utilizes a disaggregated “database-as-logs” architecture with a Rust-based storage engine that enables sub-millisecond multi-zone commit latency and independent scaling of compute and storage—capabilities not available in the standard Azure Database for PostgreSQL Flexible Server. HorizonDB also natively incorporates DiskANN vector search, in-database AI model invocation, and semantic hybrid search, rather than relying on external extensions.
How does Microsoft Web IQ work, and what is it used for?
Web IQ is a suite of AI-native grounding APIs built upon Bing’s global index and redesigned for agentic workflows. Instead of returning complete web pages, it provides passage-level “structured evidence objects,” enhancing the information density per token and reducing both cost and latency per grounding call. Web IQ is already integrated into Microsoft Copilot and ChatGPT and is currently available in limited access for select Azure customers.
Is Azure HorizonDB serverless?
No, HorizonDB is not currently serverless. Unlike AWS Aurora DSQL, CockroachDB, and YugabyteDB, it does not offer a serverless compute option at this time. While storage scales automatically, compute must be manually configured, and developers are responsible for adding or removing replicas based on their throughput needs. Microsoft has confirmed this limitation, which may change in future releases.
What benchmarks support Microsoft’s performance claims for HorizonDB and Web IQ?
All performance metrics for both products are derived from Microsoft’s internal benchmarks. HorizonDB’s claimed threefold performance improvement is measured against self-managed PostgreSQL, not against Aurora or AlloyDB. Similarly, Web IQ’s reported 2.5 times latency advantage is compared to competitors labeled only as Competitor A through G, and the configurations used may not encompass all real-world deployment scenarios. No independent third-party benchmark results for either product had been published as of June 3, 2026.
