Graphics cards and a processor are displayed before server racks, with a glowing cloud network above.
Microsoft's Ignite message this week was that Azure will offer more kinds of AI silicon. It did not say Nvidia is on its way out. Headlines about "reducing reliance on Nvidia" are accurate in direction, but they overstate how far the shift has gone so far.

Graphics cards and a processor are displayed before server racks, with a glowing cloud network above. What Microsoft actually said​

According to Korea's JKN, citing Yonhap, Jeremy Winter, Corporate Vice President of Azure Core, told Ignite in San Francisco on November 18 that Microsoft has long believed a "heterogeneous approach" to silicon is the right one. The same report says Winter stated that customer demand for Nvidia chips remains very high and that the Nvidia partnership will continue.

That is portfolio management, not a breakup. The JKN piece itself describes the strategy as picking chips on supply stability, performance and cost, rather than excluding Nvidia.

Azure's AMD options​

Microsoft's Azure Infrastructure Blog, dated November 12 and updated November 17, sets out what AMD hardware is available now and what is still coming:

  • AMD Instinct MI300X: Microsoft says it is generally available on Azure today. Through Microsoft Foundry Managed Compute, teams can deploy open models from the Foundry catalog onto dedicated MI300X capacity behind a private endpoint. Microsoft manages the software stack for each model, and billing is per GPU-hour. Billing stops when instances are spun down or the deployment is deleted.
  • Azure ND MI300X v5 VMs: For teams that want full control of the software stack, or that fine-tune and train their own models, these VMs carry eight MI300X accelerators and 1.5 TB of HBM each. Each accelerator has 192 GB of HBM3.
  • AMD Instinct MI455X: Azure and AMD are working on infrastructure built around MI455X accelerators and AMD's Helios reference design. Microsoft describes this as "coming to Azure". It is not generally available yet.
  • Azure HDv2 (CPU): Preview registration is open for VMs built on AMD's sixth-generation EPYC "Venice" processors. These are for AI data preparation and HPC, not AI acceleration, so they are a separate category from the MI-series GPUs.

The JKN report says MI300X and MI455X virtual machines will serve Azure's AI services. The Microsoft post is more careful: MI300X is available today, and MI455X is described as upcoming.

How to read Microsoft's benchmark claims​

Microsoft's post cites a Signal65 study of MI300X on Foundry Managed Compute. In that study, endpoints for models including gpt-oss, Gemma and GLM went live in 14 to 24 minutes with no serving-stack tuning.

Microsoft's own footnotes limit what that shows:

  • The study was commissioned by Microsoft.
  • The economics are GPU-only, at a tested sustained load and $7.91 per GPU-hour.
  • The workloads differed, and there is no model-quality ranking.

It is useful evidence about deployment speed and cost on Microsoft's platform. It is not a general head-to-head against Nvidia.

The HDv2 figure has similar limits. A 1.3 TB dataset was prepared in about 37 minutes on a preproduction system. Microsoft projects that is 49% faster than the next-best tested configuration, at about 73 minutes. Microsoft says results may vary with workload, configuration and the final Azure environment.

Maia 300: Microsoft's own silicon​

Microsoft's official blog from October 12 supplies the Maia detail. Ignite built on it rather than introducing it.

  • Timing: Maia 300 was unveiled at the OCP Global Summit in San Jose. Per Microsoft's blog, production deployments begin in 2027. It was not running in production at Ignite.
  • Role: Microsoft says Maia will become the primary inference system for its own products and services. Maia 200 already serves production workloads, including Copilot. GitHub Copilot and Microsoft Foundry are due by the end of the year.
  • Specifications (Microsoft's figures):
    • Eight HBM4 stacks with 18.0 TB/s of peak memory bandwidth and a 950 W TDP.
    • Compute chiplets on TSMC's N2P node.
    • A 96-accelerator rack at 167 kW, with 27.6 TB of HBM and 1.7 PB/s of aggregate bandwidth.
  • Efficiency claims: Microsoft expects 40% more tokens per megawatt than leading GPU-based systems. It also claims up to 7x the rack-scale throughput of Maia 200 in a full 96-accelerator configuration.

Treat those comparisons with care. Analyst coverage from Jon Peddie Research notes that Microsoft names only Maia 200 as a comparator in its deep dive and gives no test conditions for the GPU claim. The "tokens per megawatt" claim (October blog) and the "tokens per watt" phrasing attributed to the Ignite keynote are different units, so don't treat them as one metric.

OpenAI and AMD: a separate, earlier deal​

The OpenAI half of the headline is not new. OpenAI announced its AMD agreement on October 6, 2025:

  • It covers 6 gigawatts of AMD Instinct GPUs across multiple generations.
  • The first 1 GW deployment is of MI450-series GPUs, set to begin in the second half of 2026.
  • AMD issued OpenAI a warrant for up to 160 million AMD shares. It vests in tranches tied to deployment milestones, scaling toward 6 GW, AMD share-price targets, and OpenAI technical and commercial milestones. That is a vesting structure, not a completed issuance.

JKN adds that cloud specialist TensorWave announced a six-year contract to provide OpenAI with 457 MW of AMD-based infrastructure. It would use MI455X and Helios racks. I could not independently verify that contract in the available materials, so treat it as JKN's report of TensorWave's announcement.

OpenAI's announcement does not say the AMD systems replace Nvidia hardware. It also does not show the full capacity is online yet.

Is this really "reducing reliance on Nvidia"?​

Directionally, yes, but with caveats:

  • Analysts quoted in market coverage suggest Maia may be only a modest share of Microsoft's AI chip spending in the near term. One estimate, attributed to Evercore ISI in an EdgeX summary, is about 10% by early 2027. Treat that as an analyst estimate, not a Microsoft figure.
  • Maia is inference-focused. Commentary from IO Fund argues that training stays on GPUs, where Nvidia's software and networking still lead. Microsoft has also said it partners with Nvidia and AMD for training.
  • JKN reports that Nvidia is supplying its next-generation "Vera Rubin" systems to major cloud operators, including Azure.
  • No source quantifies Nvidia's share of Microsoft's or OpenAI's compute, or sets a target for cutting it.

The more defensible reading is that hyperscalers are building negotiating leverage and supply resilience. Choosing among Nvidia, AMD and in-house silicon by workload, cost and availability is easier than depending on one vendor.

What this means for Azure customers​

  • Inference on open models: MI300X Managed Compute in Foundry is a real option today. Try it with your own models and measure cost per session on your workload. Vendor studies, including Microsoft's, use their own test conditions.
  • Control over the stack: ND MI300X v5 VMs suit teams that want to manage software themselves or fine-tune. Check that your frameworks and kernels support AMD's software ecosystem before committing.
  • Planning ahead: MI455X and Maia 300 are future capacity. Don't plan around them for production until Microsoft gives regional availability and pricing.
  • Maia access: Maia is positioned for Microsoft's own services, such as Copilot. Whether it will reach Azure customers directly has not been established in the evidence reviewed.
  • Vendor neutrality: Jon Peddie Research advises CIOs to keep Azure inference planning flexible and to demand independent, matched benchmarks. That seems sound.

Bottom line​

Microsoft is widening its silicon menu: Nvidia, AMD in two generations, and its own Maia line. OpenAI is pursuing a large AMD supply deal in parallel. Neither company has announced an Nvidia exit, and the evidence shows Nvidia remains central. The story is more choice for customers, and more leverage for Microsoft and OpenAI when negotiating with suppliers.

 

References

  1. Microsoft and OpenAI to Reduce Reliance on Nvidia by Expanding Use of AMD and In-House Chips - 재경일보 재경일보 2026-11-19T04:25:12+00:00
  2. AMD and OpenAI announce strategic partnership to deploy 6 gigawatts of AMD GPUs openai.com
  3. Microsoft to expand Azure Maia system to multi-gigawatt deployments - The Official Microsoft Blog blogs.microsoft.com