Microsoft Build 2026: Enterprise AI Wins on Data Context

TL;DR

Microsoft announced HorizonDB, a PostgreSQL-compatible database scaling to 128 TB storage and 3,072 vCores, with built-in vector search and AI model management, now in public preview
GPU acceleration is coming to Fabric Data Warehouse (early access July 2026), delivering up to 7x faster performance than unnamed competitors at 64-user concurrency
Fabric IQ is now generally available, extending Power BI’s semantic models with operational context, business rules, and real-time signals—positioning it as the “context layer” agents need
Microsoft’s core thesis: Enterprise AI success depends on organizational memory and data context, not just model capabilities

What Happened

Microsoft used Build 2026 to make a direct argument: the bottleneck in enterprise AI isn’t model quality anymore. It’s organizational context. The company announced three major pieces of infrastructure—HorizonDB database, GPU-accelerated Fabric Data Warehouse, and general availability of Fabric IQ—all designed to solve the same problem.

That problem, according to Amir Netz, CTO of Microsoft Fabric, is that agents start from zero every time. “You want the AI to be like an employee of a company—an insider who knows how the machinery operates, what the goals are—rather than a stranger on the outside,” Netz told The New Stack.

The announcements position Microsoft Fabric as both the data foundation and deployment target for enterprise agents. Instead of forcing companies to stitch together separate systems for transactions, analytics, search, and AI, Microsoft is betting that a unified platform with a shared “context layer” will win.

Why It Matters

Agents without context are expensive guesswork machines. Every enterprise experimenting with AI agents hits the same wall: the agent doesn’t understand how the business works. It doesn’t know that “Q4 revenue” means something specific in their data warehouse, that certain actions require approval workflows, or that real-time inventory signals should override historical forecasts.

Microsoft’s play is to turn organizational knowledge—semantic models, business rules, entity relationships, operational signals—into a structured layer that any agent can query. Fabric IQ sits on top of Power BI’s semantic models, which are already used by roughly 500,000 organizations. That’s not a greenfield integration challenge. It’s an installed base.

The infrastructure choices matter for developers. HorizonDB isn’t just another PostgreSQL fork. Scaling to 128 TB with sub-millisecond multi-zone commits means transactional workloads that would previously require sharding or external caching can now run in a single system. The built-in vector search and AI model management eliminate the need to bolt on Pinecone or manage separate embedding pipelines.

For data teams, the GPU acceleration in Fabric Data Warehouse changes the performance ceiling. Netz noted that 10% annual performance gains are typical in warehousing—GPU acceleration is delivering 5x to 100x. UNC Health reported 5x query speed improvements in preview, which translates directly to faster agent responses and lower infrastructure costs at scale.

Key Details

Azure HorizonDB (Public Preview)

PostgreSQL-compatible, fully managed
Elastic storage: up to 128 TB
Compute scaling: up to 3,072 vCores
Sub-millisecond multi-zone commit latency
Built-in vector search and AI model management
Direct integration with Microsoft Foundry and Fabric

Fabric Data Warehouse GPU Acceleration (Early Access July 2026)

NVIDIA accelerated computing integrated at warehouse layer
No query rewrites required
Up to 7x faster than three unnamed cloud competitors at 64-user concurrency (internal benchmarks, May 2026)
Underlying research won Best Industry Paper at ACM SIGMOD 2026

Fabric IQ (Generally Available)

Semantic and ontology layer built on Power BI models
Includes business entities, relationships, rules, and real-time signals
Integrated into Microsoft Foundry, Agent 365, Copilot Chat, and GitHub Copilot CLI
Operations agents (GA) continuously monitor and act on business logic
Ontologies expected to reach GA in coming months

Additional Releases

Graph in Fabric (GA): Models relationships between entities and systems
Planning in Fabric (GA later this month): Adds forward-looking forecasts with write-back capability
Database Hub (Private Preview): Centralized management for HorizonDB, Azure PostgreSQL, Cosmos DB
Cosmos DB Linux Emulator (GA), semantic reranking and agent memory toolkit (Preview)
Rayfin SDK (Open Source): Build and deploy app backends directly to Fabric

Implications

This is Microsoft’s answer to Snowflake and Databricks. Netz positioned Fabric explicitly against competitors he characterized as “primarily analytical.” The argument is that agents need operational and analytical workloads in one platform—transactions, real-time signals, historical data, and forecasts—with shared context layered across all of it.

If the strategy works, Microsoft captures the entire enterprise AI stack: the database (HorizonDB), the warehouse (Fabric with GPU acceleration), the context layer (Fabric IQ), and the agent runtime (Foundry, Agent 365, Copilot). Companies building agents wouldn’t just use Microsoft tools—they’d run on Microsoft infrastructure end-to-end.

The temporal coverage argument is compelling. Netz described Fabric’s evolution in terms of time: “We had the past. We had the present. The missing piece was the future—what is supposed to happen.” With planning capabilities writing back into Fabric, agents gain a closed-loop view spanning historical data in OneLake, real-time signals from Fabric Real-Time Intelligence, and forward-looking forecasts. That’s the kind of unified data model that’s painful to replicate across multiple platforms.

OpenAI choosing Cosmos DB as its primary operational database is the kind of reference customer that matters. It signals that Microsoft’s infrastructure can handle the scale and performance requirements of the most demanding AI workloads.

Our Take

Microsoft is right about the problem, but execution is everything. The context layer thesis is sound—agents that don’t understand organizational structure are limited to generic tasks. But “organizational memory” is a hard technical and political problem. It requires semantic models that stay current, governance policies that don’t break agents, and enough performance that latency doesn’t kill the user experience.

The HorizonDB and GPU acceleration announcements suggest Microsoft is serious about removing infrastructure bottlenecks. Sub-millisecond commits and 7x query performance aren’t incremental gains—they’re the kind of improvements that unlock new use cases.

Watch how Fabric IQ adoption plays out across the agent ecosystem. Integration with Foundry, Agent 365, Copilot, and GitHub Copilot CLI is strategic, but the real test is whether developers building agents outside Microsoft’s ecosystem can access Fabric IQ easily. If it becomes a proprietary moat, competitors will build their own context layers. If Microsoft opens it up more broadly—potentially through expanded MCP tooling—it could become a standard.

The Rayfin partnership with Replit is notable because it lowers the barrier for developers and coding agents to target Fabric directly. That kind of distribution matters when you’re trying to establish platform lock-in early in a new market.

Bottom line: Microsoft is making a structural bet that the enterprise AI platform that wins will be the one that solves data context, not just model access. The infrastructure looks capable. The question is whether enterprises adopt it faster than they adopt competing approaches from Snowflake, Databricks, or open-source stacks.