Microsoft plugs AMD's newest chips into Azure to handle the AI compute crunch

Key points

• Microsoft announced three new Azure virtual machine types in 2025, all built around AMD's newest processor technology.
• HDv2 virtual machines pack nearly 500 CPU cores and 4 terabytes of RAM, aimed at feeding data to large AI training and agent systems.
• HXv2 virtual machines reach clock speeds above 5 GHz and carry 50 percent more cache per core than the previous generation, targeting chip designers.
• The ND MI455X v7 virtual machine uses AMD's Helios rack-scale system to run AI inference, the step where a finished AI model answers real user queries.
• AMD's CTO Mark Papermaster confirmed AMD uses Azure HX machines itself to design future EPYC and Instinct chips.

More models, more software agents acting on users' behalf, more raw compute eaten every quarter. Microsoft's answer, announced on the official Microsoft AI blog, is three new flavours of Azure virtual machines, each built around AMD's latest silicon. Think of a virtual machine as a rented computer inside a massive data centre: you pay for what you use, Microsoft handles the physical hardware.

The data feeder: Azure HDv2

AI models learn from data, and agents need a constant supply of it. HDv2 keeps that pipeline full, using AMD's 6th-generation EPYC processors and squeezing in nearly 500 cores, 4 terabytes of RAM and 32 terabytes of fast local NVMe storage. That's a lot of horsepower dedicated purely to moving and organising information.

The chip designer's workhorse: Azure HXv2

Designing a new processor means simulating billions of transistors in software before any silicon gets made. Brutal work. HXv2 targets those customers with AMD's 3D V-Cache technology, a trick that stacks extra memory directly on top of the processor to cut data-travel time, pushing clock speeds past 5 GHz. It's a meaningful step up from the HX series that launched in 2023: 50 percent more addressable cache per core, up to 4 terabytes of RAM, and 800 Gb InfiniBand networking for large-scale simulations. AMD's own engineering teams use the HX series to design next-generation EPYC CPUs and Instinct GPUs, and Mark Papermaster, AMD's Executive Vice President and CTO, said the company is "excited about Azure HXv2, which is designed to deliver even greater performance and scalability."

As we've tracked in nine AMD stories since July and reported separately on the cost pressures reshaping AI infrastructure, specialisation like this is exactly what enterprises are starting to demand when they scrutinise their cloud bills.

The inference engine: ND MI455X v7

Inference is when a finished AI model does its actual job: answering a question, generating an image, running a search. ND MI455X v7 handles that at production scale via AMD's Helios rack-scale platform, a system linking multiple processors across an entire server rack.

Should you care if you don't run a data centre?

Directly, nothing changes in your day-to-day apps. But the companies building those apps now have more specialised options rather than forcing every task onto the same general-purpose hardware. That tends to mean faster responses and lower costs over time. The bigger story is that power efficiency, not raw speed, is increasingly the constraint: more specialised silicon helps squeeze more useful work from each watt, which matters to everyone paying an electricity bill.