Software King of the World, Microsoft will deploy rack-scale AI systems from both AMD and Nvidia, ensuring Azure is not left dependent on Jensen Huang’s increasingly expensive hardware empire.
Chief executive Satya Nadella said Azure would be among the first cloud providers to install infrastructure based on AMD Helios and Nvidia Vera Rubin.
The decision gives AMD an important customer while allowing Microsoft to play the two chipmakers against each other on price, supply and performance.
For those who came in late, AMD’s Helios design combines 72 Instinct MI455X accelerators with sixth-generation EPYC Venice processors and Pensando networking hardware.
The system uses AMD’s Infinity Fabric and open ROCm software stack, delivering up to 2.9 exaflops of low-precision AI performance in a single rack.
Nvidia’s Vera Rubin NVL72 connects 72 Rubin GPUs with 36 Vera CPUs, ConnectX-9 networking, BlueField-4 data-processing units and NVLink 6 switches.
Nvidia claims the system can train AI models using one-quarter as many GPUs as Blackwell while cutting inference costs by up to 90 per cent.
Futurum Group estimates that Helios will cost between $5 million and $5.5 million per rack, compared with $3.5 million to $4 million for Vera Rubin. That would make AMD’s system roughly 40 per cent more expensive.
Neither chipmaker has confirmed those prices, and other analysts have estimated that some Vera Rubin configurations could cost considerably more. The comparison should be stored beside other theoretical AI benchmarks and unusually optimistic property valuations.
AMD argues that the purchase price is less important than total operating costs, including energy consumption, memory capacity and the number of tokens produced from each expensive gulp of electricity.
Helios offers more high-bandwidth memory per accelerator and uses an open rack design, allowing hyperscalers to customise components without seeking permission from Nvidia’s ecosystem.
Nvidia still has the stronger software platform, a mature NVLink architecture and a market position that has made CUDA practically compulsory for many AI developers.
Microsoft’s hardware strategy mirrors its increasingly model-agnostic approach. Azure already offers models from OpenAI, Anthropic, Meta, DeepSeek, xAI, Mistral and several others instead of treating one supplier as sacred scripture.
Deploying both platforms should improve Microsoft’s access to scarce accelerators while giving it some leverage over AMD and Nvidia.
Helios does not need to defeat Vera Rubin immediately to succeed. Winning a place beside Nvidia inside Azure gives AMD credibility, Microsoft bargaining power and Nvidia the unfamiliar experience of having a customer with another number to call.
