I wanted to include a couple comments I got from Lori MacVittie and Joel Moses on the subject of the Fungible acquisition.
I think perhaps the difficulty of development and lack of OPI support did influence the price of Fungible, but it’s not like MSFT cares all that much. It didn’t buy Fungible to turn it into a consumer-consumable service. It’s gonna integrate it into Azure and slap its own interfaces atop that. Consider what MSFT did for NVIDIA with DirectX… It’s not like NVIDIA doesn’t have its own interfaces, it does. But it was DirectX across console and PC that made it explode into the behemoth it is today.
Also, “HARDWARE MATTERS AGAIN!!!” 😊
Microsoft’s interest in Fungible is much like Amazon’s acquisition of Annapurna a few years ago – it’s an attempt to invest in technologies that make the operation of hyperscale clouds more efficient and cost-effective. Amazon turned Annapurna first into an internal infrastructure tier and then into a public-accessible ARM-based EC2 compute offering. Microsoft has been using ARM and DPU/SmartNIC technologies from others for a long time internal to Azure, and this lets them get ownership of those costs they’d source from other vendors.
The quote there is very telling, ("But its DPU architecture was difficult to develop for, reportedly, which might’ve affected its momentum.") and is the reason why we’re so bullish on creating an enterprise marketplace around DPU/IPU technologies – SmartNICs and DPUs are incredibly hard to develop for unless you’ve got a staff of folks dedicated to creating abstractions on top of them that are unique to the way you do business. That works for hyperscalers (like Microsoft) but not for enterprises or to a large degree Service Providers/Telcos. We’re placing a lot of hope in the industry motion around Open Programmable Infrastructure to help here.
In the end analysis, Microsoft’s motion is proof of two things: 1) the rising wave of DPU/IPU technologies inside hyperscalers (service providers are next, followed by large enterprises!), and 2) proof that the forecasts are going to be accurate: ARM workloads within cloud and datacenter environments are on track to comprise at least 15-25% of compute workloads by 2025.