AWS recently introduced another impressive list of new and improved capabilities at its annual re:Invent conference. I believe the most important news items from this year’s event are AWS’s entry into the broader hardware market, via vertical integration of its own processors into its cloud infrastructure, and expansion of that cloud infrastructure into private, on-premises enterprise deployments. There were three important announcements in these areas.
AWS Graviton for general purpose, scale-out applications
Graviton, available now, is an Arm-based CPU integrated into a new AWS EC2 “A1” VM type. AWS is the first large public cloud provider to deliver an Arm-based compute capability. AWS positions these new VMs as delivering cost savings for general purpose applications such as web servers, containerized microservices, caching fleets, and distributed data stores with Arm-based applications available.
The more that general purpose applications, especially common open source applications, become available in editions that run on Arm processors, the more AWS may be able to differentiate on compute cost. If successful, competitors will likely to need to respond with their own Arm processor-based offerings. This forms a long-term threat to Intel’s dominant CPU market share amongst cloud providers. This, along with AWS’s recent introduction of AMD CPUs, also gives AWS leverage in negotiation with Intel on CPU pricing of the most popular VM type AWS offers. For more information on Graviton, check out this post from my colleague, Matt Kimball.
AWS Inferentia for machine learning inference
Inferentia is a custom machine learning inference processor designed to improve performance and reduce the cost of inference workloads. With availability projected for late 2019, AWS is a few years behind Google and its custom machine learning Tensor Processing Unit (TPU). That being said, Inferentia’s performance compared to Google’s TPU is unknown. AWS said Inferentia will support Tensorflow, Apache MXNet, and PyTorch deep learning frameworks, whereas Google’s TPU only supports Tensorflow.
Machine learning is a critical area of cloud provider capability given how quickly and broadly machine learning is being adopted by the enterprise. If Inferentia drives differentiation for AWS versus the cloud providers who currently rely on NVIDIA’s GPUs to address inferencing workloads, the competition will likely to need to respond more specifically (e.g. Microsoft, IBM, Oracle, and Alibaba). This is a long-term threat to NVIDIA’s market dominance in machine learning accelerators amongst cloud providers. Similar to the situation with Intel in CPUs, this will also likely provide AWS with leverage to negotiate with NVIDIA on GPU pricing for one of the most popular applications for GPUs in cloud infrastructure. For more information on Inferentia, check out this postfrom my colleague, Karl Freund.
AWS Outposts for private deployments of AWS services
AWS also announced Outposts—infrastructure for deployment in private enterprise on-premises or colocation data centers that will be fully managed, maintained, and supported by AWS. Available starting in the second half of 2019, AWS said Outposts customers will have the option of using AWS’s native infrastructure services or VMware’s Cloud on AWS services. AWS is following Microsoft’s lead with this hybrid cloud capability—Microsoft has been in market for a year with its Azure Stack offering. There are two significant aspects to this AWS introduction.
First, Microsoft’s Azure hybrid cloud offering has emerged as one of its top differentiators versus AWS and VMware. If AWS can cut into Microsoft’s market position as the only hyperscale cloud provider delivering a seamless, native hybrid cloud experience, it could potentially both reduce Microsoft’s growth and yield much more growth to AWS from the larger private infrastructure market. Microsoft Azure is growing the most of any cloud provider other than AWS, yet it is still losing market share as AWS absorbs more of the market’s growth each quarter. This growth gap would only increase if Outposts is successful.
Second, Microsoft partnered with the global OEM channel to deliver its Azure Stack offering for private enterprise deployment. AWS is bypassing the OEM channel and setting up a potentially significant long-term threat to the market share of private infrastructure leaders Cisco, Dell EMC, Hewlett Packard, and Lenovo—especially if its model is successful and Microsoft follows with a similar direct go-to-market (as it did with its Surface client devices).
On the other hand, Outposts is still at least 6-12 months away and we have no specifications on hardware options or available AWS services, pricing, or support. There are plenty of details AWS will need to get right even beyond the specifications (e.g. private network integration and performance) to see strong adoption in private deployments. Still, AWS will be starting with the advantage that its cloud platform has many more customers already on it than Microsoft and other competitors.
There is a whole set of questions related to each of these three new offerings that need answering in order to predict how compelling and competitive they will be in the market. We will start to get some answers soon on the Arm compute offering since it is available now, but for the most part, it will be this time next year before we have a sense of how successful these offerings could be long-term. What is for sure is that if they are successful, they could force a shift in approach by other cloud providers, beyond the existing Google and Microsoft efforts. This could further shift market power from hardware vendors to cloud providers in a way that could significantly impact the shape of the margins in the infrastructure value chain.