Will Cloud Hyperscalers Replace Traditional Chip Manufacturers as the Provider of Intelligent Accelerators for Enterprises?

The emergence of technologies like Artificial Intelligence (AI), Machine Learning (ML), 5G, quantum computing, etc. has resulted in a significant strain on Central Processing Units (CPUs). ABI Research is seeing cloud hyperscalers designing and building intelligent accelerators or acquiring intelligent accelerator producers, including Amazon Web Services (AWS) with its Nitro System, Alibaba Cloud with its Cloud Infrastructure Processing Unit (CIPU), Baidu Cloud’s Taihang DPU that leverages Intel’s IPU reference architecture, and Microsoft Azure’s acquisition of Fungible, a Data Processing Unit (DPU) hardware provider, to bolster its data center infrastructure offerings.

Currently, intelligent accelerators are almost exclusively deployed by cloud hyperscalers, with Intel and Google Cloud co-designing the E2000 IPU that powers Google Cloud’s C3 virtual machine. AMD’s Pensando Distributed Services Card is one of the first DPUs to support VMware’s vSphere 8, while Oracle Cloud deploys Marvell Technologies’ OCTEON LiquidIO III SmartNIC solution in its data centers to offer customizable cloud network and firewall security features. Also, NVIDIA recently announced its latest BlueField-3 DPU solution, with Oracle Cloud being one of the early adopters.

ABI Research believes that this will create two distinct segments of intelligent accelerators. The first segment includes mega cloud hyperscalers that design and develop in-house chips to be deployed as part of their cloud solutions platform, with flexibility, customization, and cost as major factors in producing proprietary intelligent accelerators.

The other segment will be large enterprises and Communication Service Providers (CSPs) that will deploy intelligent accelerators in their own private data centers to support growing data and business needs. While this segment is quite small compared to the first segment, as more and more businesses adopt solutions, such as AI, ML, the Internet of Things (IoT), etc., there will be growing demand for intelligent accelerators to be used to optimize workload processes within the environment of a local private data center.

As cloud hyperscalers build up workload acceleration capabilities, companies like Intel, Marvell, AMD, etc. will have to consider the impact of this evolution. The possibility of cloud hyperscalers scaling down the purchase of intelligent accelerators might happen in the long run as the technology and solutions mature. This will represent a big shift, as cloud hyperscalers are the largest customer base of intelligent accelerator hardware for chip manufacturers.

Moving forward, cloud hyperscalers will play a role in commoditizing the use of intelligent accelerators among public cloud customers. Increasingly, ABI Research is seeing cloud hyperscalers integrating intelligent accelerators into existing cloud solutions, such as AWS Nitro with the EC2 platform, as well as Google Cloud’s C3 machine series. However, cloud hyperscalers will need to realize that in the current environment, it is imperative to focus on specific use cases to increase the adoption rate of intelligent accelerators. For example, intelligent accelerators can be used for generative AI solutions, such as answering questions on internal Human Resources (HR) policies, enabling a better user experience using real-time inferencing to display answers instantly.

Chip manufacturers will also need to pivot. It will no longer be viable to focus only on selling to cloud hyperscalers, especially Tier One cloud providers. Chip manufacturers should explore working closely with Tier Two cloud providers, such as OVHcloud, IBM Cloud, Rackspace, etc. This could prove to be a better fit in the long term, as these cloud solution providers typically have quite a diverse customer base.

Chip manufacturers should also start targeting large enterprises. ABI Research is seeing increasing traction and demand for intelligent workload acceleration solutions for industries like automotive, telecommunications, and industrial manufacturing. Specifically for automotive, producers of autonomous vehicles are looking to intelligent accelerators to help process large amounts of data from cameras and sensors instantaneously in real time. CSPs are deploying intelligent accelerators to offload data processing tasks, leaving the CPU to focus on core network functions, while industrial smart factories are using intelligent accelerators to help detect supply chain anomalies and product defects.

While cloud hyperscalers may be on the cusp of disrupting the intelligent accelerator ecosystem by building in-house chips, traditional chip manufacturers still play a major role in this market. The legacy of chip-building expertise, large Research and Development (R&D) investments, and expert knowledge accumulated over the years will keep chip manufacturers ahead of the curve. However, chip manufacturers should look into diversifying their current customer base and quickly embrace the enterprise space as the market continues to evolve and mature.