Azure's Vera Rubin offer is a technology preview of confidential bare metal, and we expect it to stay in preview through January 2027. Microsoft's spending figures are company wide, rather than specific to Rubin.
At Ignite, Microsoft demonstrated a model with 397 billion parameters on four Rubin GPUs. Nvidia says Vera Rubin NVL72 systems are running inside Azure data centers. A procurement team faces a different proposition: Microsoft's confidential Rubin offering is a technology preview, while Oracle and Google Cloud have announced general availability. We would pass on paying a commercial launch premium for Azure's expansion; investors would be buying a future service launch alongside today's installed hardware.
We call the distance between installed hardware and an orderable cloud service the Deployment Gap. A hotel can have furnished rooms and host invited guests before offering rooms on a public booking page. Running Rubin racks are the furnished rooms; an Azure listing with regions and prices is the booking page. Microsoft dates Rubin validation to March 2026 and production systems to August 2026. We expect its Rubin offering to remain a technology preview through January 2027.
Nvidia's 19 November 2026 announcement describes expanding VM access and Nvidia open models through Microsoft Foundry. Microsoft's 17 November description starts with Confidential Bare Metal Instances in technology preview. Its Ignite demonstration used one node, four Rubin GPUs and vLLM to load a model with 397 billion parameters using four way tensor parallelism. These details describe a showcase for sensitive workloads, rather than the generally available Azure instance an enterprise can compare with another cloud's.
Nvidia attestation lets customers verify the computing environment before releasing confidential models, data or keys, extending protection to bare metal and rack scale. Microsoft's Ignite roundup puts confidential virtual machines, Azure Kubernetes Service containers and Foundry integration in development. Customers building today are directed to existing H100 confidential GPU VMs. The preview has a practical security purpose. Enterprise budgets, meanwhile, depend on which configurations are ready for purchase and at what price.
Oracle declared Vera Rubin NVL72 generally available on 27 October 2026; Google Cloud marked A5X generally available on 6 November 2026. At Nvidia's 19 November announcement, the two rivals had a lead of 23 days and 13 days, respectively. This is an ordering advantage, rather than an inference benchmark. Engineers can evaluate a publicly available product and make architecture commitments around it. Earlier testing can shape where later cloud spending lands. The Deployment Gap therefore creates a commercial opening while Microsoft finishes packaging its service.
Capital spending makes this a less comfortable story than a simple launch race. In the quarter ended 30 September 2026, Microsoft's filing lists $39,367M of additions to property and equipment, versus $19,394M a year earlier; $39,367M over $19,394M is 2.03 times. Against revenue of $92.3B for the same quarter, $39,367M is 42.7% of sales. Microsoft also obtained $5,192M of right of use assets for finance leases. These figures cover the wider Microsoft infrastructure program rather than a single Rubin build.
In the quarter ended 30 September 2026, Microsoft Cloud revenue was $63.9B, up 30%, while Azure revenue grew 48%. Commercial remaining performance obligation reached $677B, up 73%, and Microsoft expects to recognize around 30% of that balance over the next 12 months. Under that recognition scenario, $677B times 30% is $203.1B. These are commitments across Microsoft, covering far more than Vera Rubin. They make investment in capacity plausible while leaving investors to distinguish contracted demand from the launch status of one new cloud service.
Microsoft's scale offers a different interpretation of that spending. On 5 January 2026, Microsoft said it had redesigned its rack architecture for Rubin and projected 3.6 EF NVFP4 per rack, five times GB200 NVL72. By August it had production Rubin systems. With $677B of commercial obligation at the September quarter end, Azure has a reason to prioritize capacity for large contracting customers before publishing a standard product. Under that account, private capacity arrangements could monetize racks first, while general availability would widen distribution. An Azure customer with a large contract might care more about its allocated hardware than a public instance page. A preview therefore need not be the start of commercial use.
Signal65 offers the most compelling argument for Microsoft's execution speed. In its 9 November 2026 assessment, presented in a Microsoft partnership context, Azure's ND GB300 v6 with 72 GPUs delivered 260,098 output tokens per second on DeepSeek-R1 Interactive, 10.3% above the next cloud submission. Signal65 called Azure "the only cloud provider in this round with a multi-rack NVL72 result." The previous cycle went Azure's way: Microsoft's 9 October 2025 announcement described the first GB300 NVL72 production cluster at scale, with more than 4,600 NVIDIA GB300 NVL72. That record makes general availability for Azure Rubin before the end of January 2027 a plausible opposing outcome.
The earlier triumph involved GB300 Blackwell Ultra, while Signal65 used the MLPerf Available category, excluding preview and research systems. On the same public service calendar for Rubin, Oracle was generally available on 27 October 2026, Google on 6 November, and Microsoft announced a confidential bare metal technology preview on 17 November. Microsoft's four Rubin GPUs in the Ignite demonstration provide engineering evidence of a different sort. A public Azure Rubin listing, with region and price, would close the Deployment Gap for enterprise procurement; November's statuses still leave it open.
Microsoft and Nvidia also have ways to make each unit of enterprise computing go further. Nvidia's 19 November 2026 post describes a power efficiency collaboration using MaxLPS and, separately, Nemotron 3.5 models on Foundry Managed Compute. EY adapted Nemotron to classify purchasing records, improving accuracy on its evaluation dataset by 55% while lowering cost by 15x compared with proprietary frontier models. Those are task specific model economics, rather than a Rubin hardware comparison. They matter because a lower inference cost can persuade customers to spend on Azure workloads even as Vera Rubin productization proceeds.
Microsoft's September quarter filing says capacity constraints and changing demand can alter allocations, deployment priorities and pricing. That makes prioritizing strategic users rational, even as Oracle and Google's earlier general availability gives enterprises an alternative for new evaluations. An available cloud service can start shaping architecture choices before a competitor publishes its equivalent. Customers comparing inference bills need a price attached to the hardware, alongside the model savings Nvidia promotes. Investors should price Azure's existing growth separately from a product milestone still ahead.
We would pass on paying today for a completed Azure Rubin commercial rollout, while recognizing that Microsoft has the capital, contracted demand and engineering record to close the Deployment Gap. We would be wrong if an Azure product or blog page lists a Vera Rubin NVL72 instance or bare metal offering as generally available, with regions and pricing, by 31 January 2027.