On the first day of Microsoft Build 2026 in San Francisco, Microsoft announced the public preview of Azure HorizonDB, a fully managed PostgreSQL-compatible database rebuilt from the ground up for agentic AI workloads, alongside the limited availability launch of Web IQ — a web-grounding API layer that, Microsoft revealed, already powers the grounding systems behind both Microsoft Copilot and OpenAI's ChatGPT. For enterprise developers and cloud architects, the double announcement is actionable today: HorizonDB is open for preview signups across five Azure regions as of June 2, and Web IQ access is available by application to select Azure customers.

The strategic frame behind both products is the same. CEO Satya Nadella argued at the keynote that the bottleneck holding back enterprise AI is no longer model capability — it's the infrastructure underneath. Agents create fundamentally different data call patterns: they store, retrieve, reason, act, and loop continuously, often at hundreds of concurrent queries, rather than the relatively predictable read-and-write cycles that traditional databases were designed to serve. HorizonDB and Web IQ are Microsoft's answer to that infrastructure gap.

Azure HorizonDB in Public Preview: What Changed Since Ignite

Azure HorizonDB is not an incremental update to Azure Database for PostgreSQL. It is architecturally distinct, built on what Microsoft calls a "database-as-logs" design that commits transactions directly to shared write-ahead log (WAL) storage rather than requiring the multiple coordination steps that traditional PostgreSQL commits involve. The practical consequence is sub-millisecond multi-zone commit latency — a meaningful difference for workloads where agents are executing dozens of read-write cycles per second.

The storage engine itself is written in Rust — a choice Microsoft made explicitly for memory-safety guarantees. Unlike C-based storage engines, Rust eliminates whole categories of buffer overflow vulnerabilities by construction, which matters when a database is being queried at high frequency by autonomous processes that lack the error tolerance of human-supervised workflows.

Compute and storage are disaggregated: the two scale independently. Storage grows automatically to 128 TB; compute scales to 3,072 virtual CPU cores across primary and replica nodes. Stateless compute nodes — which hold no data themselves — enable up to 15 read replicas with near-zero downtime during maintenance. Microsoft claims this architecture delivers up to 3x faster transactions and search performance compared to self-managed PostgreSQL, though those figures are from Microsoft's own internal benchmarks and have not been independently verified by third parties.

Mohsin Shafqat, Director of Software Engineering at NASDAQ, said in a statement provided by Microsoft 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, described the gap HorizonDB fills when the product first entered private preview at Ignite 2025: Azure lacked "a big, distributed, all-you-can-eat, vanilla Postgres," he said. "Now that gap is filled."

One important constraint: HorizonDB is not yet serverless. Unlike AWS Aurora DSQL, CockroachDB, and YugabyteDB — all of which offer serverless SKUs — HorizonDB requires developers to manually configure compute and add or remove replicas depending on throughput needs. Only storage scales automatically. This is a meaningful operational difference for teams that rely on zero-configuration autoscaling, and Microsoft's own vice president of engineering confirmed the limitation directly when asked.

DiskANN and Spherical Quantization: Vector Search Inside the Database

The feature that most sharply differentiates HorizonDB from Amazon Aurora PostgreSQL and Google AlloyDB is its approach to vector search. Aurora and AlloyDB handle vector workloads primarily through extensions — pgvector on Aurora, ScaNN on AlloyDB — that run alongside the database engine rather than inside it. HorizonDB embeds vector search natively through DiskANN with spherical quantization, a design originally developed by Microsoft Research.

DiskANN enables nearest-neighbor vector search over large, disk-resident datasets without loading entire indexes into memory — a distinction that matters at the scale of enterprise AI workloads where embedding datasets routinely reach billions of vectors. Spherical quantization works by normalizing vector representations and encoding them into compact angular-distance-preserving forms, which allows the system to compare embeddings with minimal accuracy loss while dramatically reducing memory and storage overhead. The result supports vectors up to 16,000 dimensions — far beyond the 2,000-dimension ceiling of HNSW-based indexes in pgvector — making it viable for high-dimensional models such as text-embedding-3-large.

HorizonDB combines DiskANN with hybrid search — pairing vector similarity via pgvector with full-text search through the pg_textsearch extension in a single query. The result is retrieval that matches both semantic meaning and keyword relevance simultaneously, which is the pattern most retrieval-augmented generation (RAG) applications require. HorizonDB also adds in-database model invocation through the azure_ai extension, which brings model inference directly into the PostgreSQL engine via SQL, eliminating the external orchestration layer that most current AI-database stacks require.

Web IQ: Why Microsoft's AI Grounding Layer Was Already Everywhere

Web IQ is the more strategically significant of the two Build 2026 announcements, even though it received less keynote time. According to Jordi Ribas, Microsoft's president of Search and AI, the system was rebuilt from the ground up across every layer — indexing, retrieval, ranking, passage selection, and orchestration — so that each component serves inference-time grounding rather than the page-ranking requirements of traditional human search.

The architecture has five layers. At the base sits Bing's global index, maintained for decades, providing freshness and coverage that a purpose-built index could not replicate. On top of that sits an embedding model — Microsoft open-sourced the underlying model in April 2026 — which projects content into a vector space where semantic similarity becomes computationally tractable. The retrieval layer uses DiskANN-powered distributed partitioning routed globally. Content understanding and ranking models trained specifically for large language model-driven reasoning sit above retrieval. The orchestration layer at the top routes queries, fans out retrieval, merges results, and transforms them into evidence.

The key architectural decision is the unit of information returned. Traditional search APIs return full documents. Web IQ returns passages and what Microsoft calls "structured evidence objects." The design principle is explicit in the announcement: models do not need documents — they need information, and documents are often a poor proxy for that. By operating at the passage level, Web IQ concentrates useful signal while eliminating irrelevant context, producing a higher information-to-token ratio that lowers cost per API call and supports more precise reasoning under constrained context windows.

Microsoft claims sub-165ms P95 latency across five data centers, and reports that Web IQ is roughly 2.5 times faster at retrieval than the nearest alternative in its benchmark set. Both figures are vendor-published. The competing systems in the latency benchmark are labeled Competitor A through G, and specific configurations may not reflect the full range of production conditions developers would consider. No independent third-party evaluation of these claims had been published as of June 3, 2026.

Web IQ is model-agnostic and delivered via JSON-RPC 2.0 transport, making it compatible with any foundation model without requiring inference lock-in. Developers who want to evaluate access can express interest at Microsoft's Web IQ product page; no public pricing or general availability date had been announced as of June 3.

Microsoft Fabric as the Unified Agent Platform: GPU Analytics and the Database Hub

HorizonDB and Web IQ are designed to slot into a larger architecture Microsoft is building around Microsoft Fabric as a unified data and AI deployment platform. The stated problem is that every new agent currently starts from zero — relearning organizational context, data locations, and business rules without any shared foundation. Fabric IQ, which reached general availability at Build 2026, is Microsoft's proposed solution: a context layer that feeds agents with enterprise data via Work IQ (Microsoft 365 data), Fabric IQ (structured business data), and Web IQ (live web intelligence).

A new Database Hub in Fabric, currently in private preview, will centralize management of Microsoft's database portfolio — HorizonDB, Azure Database for PostgreSQL, and Azure Cosmos DB — with data mirrored to OneLake for cross-database analytics.

Microsoft also announced GPU-accelerated analytics for Fabric Data Warehouse, which entered early access preview and is scheduled to roll out in July 2026. Eligible queries can run directly on NVIDIA accelerated computing inside the execution engine without query rewrites, with Microsoft claiming up to 7x performance improvements in internal benchmarks. The underlying research project, CoddSpeed, won Best Industry Paper at SIGMOD 2026, the Association for Computing Machinery's flagship database conference.

What Developers and Architects Need to Evaluate

For teams already running PostgreSQL workloads on Azure, the question HorizonDB raises is not whether to migrate immediately — the service is in preview with all the instability that implies — but whether the architectural design aligns with planned agentic workloads. The disaggregated "database-as-logs" model, DiskANN-native vector search, and in-engine model invocation represent a meaningfully different operational profile from Azure Database for PostgreSQL Flexible Server, which Microsoft says it will continue to support for standard PostgreSQL deployments.

IDC research director Devin Pratt noted that the big cloud providers are converging on the same pattern of PostgreSQL-compatible services with cloud-native storage, compute, and AI features — a framing that positions HorizonDB as a catch-up move as much as an innovation, given that Amazon Aurora pioneered disaggregated database architecture in 2014. HorizonDB's differentiator is AI-native feature depth: native DiskANN embedding, in-database model management, and tighter Foundry and Fabric integration than either Aurora or AlloyDB currently offers.

Developers who want to evaluate Web IQ separately from HorizonDB can do so — it is model-agnostic and not an Azure-exclusive product, at least in principle. The Federal Trade Commission's active antitrust investigation into Microsoft's Azure cloud licensing practices, which escalated in February 2026 when the commission issued civil investigative demands to at least six competitors, is worth tracking for enterprise procurement teams. The probe centers on whether Microsoft's licensing terms make it harder to run Microsoft software outside Azure — context that is relevant to any new Azure infrastructure dependency.

A dedicated Build 2026 session 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 built specifically for agentic AI workloads. It uses a disaggregated "database-as-logs" architecture with a Rust-based storage engine that enables sub-millisecond multi-zone commit latency and scales compute and storage independently — capabilities the standard Azure Database for PostgreSQL Flexible Server does not offer. HorizonDB also embeds DiskANN vector search, in-database AI model invocation, and semantic hybrid search natively, 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 on Bing's global index and redesigned for agentic workflows. Rather than returning full web pages, it returns passage-level "structured evidence objects" that give AI models higher information density per token, reducing both cost and latency per grounding call. It already powers web-grounding in Microsoft Copilot and ChatGPT, and is now available in limited access to select Azure customers.

Is Azure HorizonDB serverless?

Not yet. Unlike AWS Aurora DSQL, CockroachDB, and YugabyteDB, HorizonDB does not currently offer a serverless compute option. Storage scales automatically, but compute must be manually configured and replicas added or removed by the developer. Microsoft has confirmed this limitation directly; it may change in a future release.

What benchmarks support Microsoft's performance claims for HorizonDB and Web IQ?

All performance figures for both products are vendor-published from Microsoft's own internal benchmarks. HorizonDB's claimed 3x performance gain is measured against self-managed PostgreSQL, not against Aurora or AlloyDB. Web IQ's claimed 2.5x latency advantage is measured against competitors labeled only as Competitor A through G, and the configurations used may not reflect all real-world deployment scenarios. No independent third-party benchmark results for either product had been published as of June 3, 2026.