Microsoft used Build 2026 to put GPU acceleration inside its Fabric Data Warehouse, calling it the first fully managed cloud data warehouse to run analytics queries on NVIDIA accelerated computing. In its own testing, eligible queries ran up to 7x faster than three rival cloud warehouses at 64-user concurrency, with no query rewrites and no clusters to size.
The speed is real. The harder problem sits upstream. Most enterprises still can’t feed an engine this fast with data clean and governed enough to trust, and that gap is what now separates companies scaling AI from companies stuck in pilots.
What Microsoft Put Inside the Warehouse
Turning it on is a workspace toggle. Teams flip one setting in Fabric workspace options and it applies to every SQL Analytics Endpoint and Data Warehouse in that workspace, no query rewrites involved. When a user hits run, the query optimizer routes eligible work to the GPU; anything it can’t accelerate falls back to the CPU engine and returns the same result. That fallback is the design choice that matters. It makes GPU acceleration additive for shops that spent years stabilizing their data estates, instead of a rip-and-replace migration.
The headline numbers come from Microsoft’s own TPC-H 300GB run (a standard analytics benchmark), conducted internally in May and set out in the Fabric and Databases Build 2026 update.
| Concurrency level | Speed vs three rival cloud warehouses |
|---|---|
| Single user | Up to 3x faster |
| 16 users | Up to 6x faster |
| 64 users | Up to 7x faster |
Microsoft also put the engine at up to 6x faster than its own CPU baseline, and paired the benchmark with production reports: a healthcare customer in early access cited up to 5x faster query speeds, a professional-services firm 3.4x at single concurrency on complex workloads, and a manufacturer gains on reporting-heavy analytics. The detail worth reading in those figures is the curve, because most warehouses slow down as more users pile on, and Microsoft’s held its lead as the load climbed.
GPU-accelerated Fabric Data Warehouse is the next step: same queries, same scale, but answers fast enough to keep up with how people think.
That line is from Bogdan Crivat, corporate vice president for Azure Data Analytics at Microsoft, in the Fabric announcement. His framing puts the value in what gets done with the speed, not the speed itself.
The Engine That Won a SIGMOD Award
The speed has a paper trail. The architecture underneath is described in CoddSpeed, a Microsoft Research project that took the Industrial Track Best Paper award at the ACM SIGMOD 2026 database conference in Bengaluru. The work, led by Microsoft researcher Matteo Interlandi and colleagues, is detailed in the CoddSpeed research publication.
It is built on the Tensor Query Processor (TQP, a GPU-based execution engine), with a data-movement layer that uses NVLink and InfiniBand interconnects to keep the accelerators fed. On the TPC-H 1TB benchmark the paper reported up to 30x faster execution than CPU, and more than an order of magnitude across production scenarios.
The principle the authors keep returning to is hardware independence. CoddSpeed is built around an abstraction layer meant to run across more than one kind of silicon:
- Compute – GPUs today, with FPGAs and custom ASICs supported by the same layer
- Interconnect – NVLink and InfiniBand for moving data between accelerators at speed
- Fallback – the CPU engine, for any query the accelerators can’t take
That abstraction is a hedge. Nobody knows which accelerator wins the next five years, so building the engine to swap GPUs for FPGAs or in-house chips without rewriting the query layer protects the investment as the silicon market churns.
Concurrency Is the Number That Matters
For years enterprise software was built around CPU transactional systems tuned for predictable workloads and human users. Agentic AI breaks that assumption. An agent doesn’t run one query and read a dashboard; it fires many requests, checks conditions and acts, often dozens of times to finish a single task. Multiply that across a workforce of agents and the request volume hitting operational data climbs by orders of magnitude.
So the 64-user figure is a stand-in for something bigger. The job is no longer one fast query; it is thousands of simultaneous interactions across applications, analytics, workflows and agents without the system buckling. That is why the concurrency curve, more than the top-line multiple, is the part Microsoft kept pointing at.
Multiply that across a workforce of agents and the request volume hitting operational data climbs by orders of magnitude.
Speed also changes behavior. When data comes back fast, people ask more questions and test more scenarios before deciding. Agents do the same at machine pace, interacting with operational systems more often, evaluating more conditions and acting closer to real time. It is the same dynamic Microsoft is leaning on across NVIDIA’s accelerated-computing stack for agentic AI, which now reaches from Azure down to Arm-based Windows PCs. Faster access turns insight into action with less friction, and that is the behavior these systems are built to provoke.
From Single Platform to Ecosystem
Microsoft is not the only vendor chasing this. The whole industry is racing to own the place where data, AI, applications and business processes meet, and a benchmark only buys a seat at that table.
The Vendors Fighting for the Layer
AWS, Google Cloud, Oracle, Snowflake and Databricks are all pitching versions of the same layer. Benchmarks win the headline; production behavior wins the renewal. Over the next few years, enterprise buyers will judge these platforms less on a TPC-H multiple and more on how they hold up under real, mixed workloads at scale. Interoperability, ecosystem reach and operational simplicity carry more weight in that decision than peak query speed.
Open Formats and Old Silicon
Here Microsoft has a defensible position. Fabric, OneLake and Azure AI are built to span platforms, support open table formats like Delta and Iceberg, and reach data that lives outside Microsoft’s own walls. Very few enterprises run a single vendor; most already stitch together several clouds, ERP systems and data platforms, so the company that reduces architectural friction has an edge over one selling a locked stack.
There is a money angle too. Microsoft says the engine can run on older-generation GPUs already racked in its data centers while keeping much of the performance gain. If that holds in practice, the conversation shifts from raw speed to cost, utilization and return on investment, the vocabulary of CFOs as much as CIOs. It also folds neatly into Azure’s recent cloud growth, where squeezing more work out of installed silicon protects margins as AI demand scales.
AI Readiness Is the Binding Constraint
Faster infrastructure does not fix the thing most enterprises actually struggle with. That is the catch buried under the benchmark.
Why AI Amplifies Bad Data
Most organizations do not have an access problem. They have a consistency problem: keeping data governed, owned and in business context as it moves across applications and platforms. AI does not solve that. It scales it. A fast engine pointed at fragmented or poorly governed data just produces wrong answers more efficiently. So AI readiness, the state of an organization’s data, governance and processes, is becoming the constraint that decides who scales past isolated pilots and who stalls.
ERP Becomes a System of Action
ERP (enterprise resource planning, the software running finance, supply chain and operations) is where this lands. For years ERP was a system of record. Now it is turning into a system of action driven by APIs, real-time data and AI-enabled workflows. Agents do not wait for overnight batch jobs; they need continuous access to operational data and the ability to execute decisions across finance, procurement, manufacturing and customer operations. Query performance and low-latency concurrency become operational requirements, not reporting niceties. The trouble is that ERP, data and cloud vendors are shipping faster than customers can modernize. The blocker is rarely the technology. It is the data cleanup, governance and organizational alignment needed to use it, and that work is slow, political and unglamorous.
What Microsoft Still Has to Prove
The next phase of enterprise AI gets decided by which vendors can operationalize trusted data and scalable infrastructure across messy real environments, not by who demos the slickest chatbot. GPU-accelerated Fabric Data Warehouse is an early marker of where Microsoft’s data platform is heading, and the engineering is credible. The validation is thinner. Microsoft backed the launch with a handful of customer reports and its own benchmarks, and enterprises will want more production deployments, more workload variety and clearer business outcomes before they commit budget.
Early access opens in the coming weeks. The proof that counts comes after, when customers try to turn 7x into measurable value against their own imperfect data.
