Analysts predict fierce competition in the AI infrastructure market as enterprises significantly increase spending over the next several years. This projection has prompted suppliers to respond to the anticipated demand swiftly.
This week, HPE launched its Nvidia AI Computing by HPE portfolio. The offering came roughly two weeks after Cisco and Arista introduced AI application platforms Nexus HyperFabric and Etherlink AI, respectively.
Cisco’s product also stems from a partnership with chipmaker Nvidia, whose AI-powering GPUs are the silicon of choice for cloud providers and enterprises running LLMs. Arista’s technology uses Broadcom chips.
The companies are racing to seize a share of the promising GenAI market, which IDC projects will skyrocket by 2027. According to the analyst firm, global enterprise spending on GenAI technology — encompassing software, infrastructure hardware and IT services — is set to surge from $19.4 billion in 2023 to $151.1 billion in 2027, reflecting a compound annual growth rate of 86.1%.
“Everyone wants to get in on the gold rush,” said Jack Gold, principal analyst at J.Gold Associates. “And why wouldn’t they?”
The foundation of Nvidia AI Computing by HPE, launched at the company’s Discover conference in Las Vegas, is HPE Private Cloud AI co-developed with Nvidia. At the heart of the software stack is Nvidia’s AI Enterprise software platform, which includes the company’s NIM inference microservices.
Private Cloud AI allows enterprises to fine-tune AI models using proprietary data that’s kept private through a development technique called retrieval-augmented generation. The Nvidia NIM provides a set of microservices for model tuning, a process called inference.
Models built with Private Cloud AI will run on Nvidia GPUs in the data center and use HPE File Storage for managing model-training data. Other components of the Nvidia AI Computing by HPE portfolio include Nvidia Spectrum-X Ethernet networking and HPE ProLiant servers with support for various Nvidia GPUs, including the L40S, H100 Tensor Core and the GH200, also known as the Grace Hopper Superchip. HPE will release the servers in the summer and the fall.
HPE will sell the systems from the Nvidia partnership as a GreenLake product. GreenLake is a suite of infrastructure tools that HPE sells as a service for private data centers. Enterprises use the systems on a pay-per-use basis.
Arista, Cisco offerings
Arista and Cisco are leaning on their expertise as networking companies for their recently introduced AI infrastructure. High-speed networking is critical to the performance of AI systems. HPE plans to become a third major networking supplier after it completes the $14 billion acquisition of Juniper Networks by early next year.
Everyone wants to get in on the gold rush. Jack GoldPrincipal analyst, J.Gold Associates
The three companies are members of the Ultra Ethernet Consortium, which is developing Ethernet as an open standard for AI networking. Today, large enterprises use the proprietary InfiniBand standard for training LLMs. Only a few vendors offer InfiniBand hardware, which is significantly more expensive than Ethernet products.
Cisco co-developed its Nexus HyperFabric with Nvidia. The prebuilt product includes Nvidia GPUs and AI Enterprise software, a Cisco 6000 series switch and a Vast data store. The switch delivers Ethernet speeds of 400 Gbps and 800 Gbps.
HyperFabric uses Nvidia’s H200 Tensor Core GPUs and the Nvidia BlueField-3 SuperNIC and data processing unit to accelerate security, networking and data access. Cisco plans to make the system available for testing by select customers in the fourth quarter.
Arista’s Etherlink AI switching platforms leverage AI-optimized features in the company’s network operating system, EOS. The platforms include the following:
7060X6 AI Leaf switch. It has Broadcom Tomahawk 5 silicon and a capacity of 51.2 terabits per second (Tbps). It supports 64 800 Gbps or 128 400 Gbps Ethernet ports.
7800R4 AI Spine switch. It has Broadcom Jericho3-AI processors and a capacity of 460 Tbps in a single chassis. It supports 576 800 Gbps or 1152 400 Gbps Ethernet ports.
7700R4 AI Distributed Etherlink Switch. Arista designed the hardware, which also uses the Jericho3-AI architecture, for the largest LLM deployments.
“All three of [the companies] are making meaningful efforts, but each one of them has different nuances,” IDC analyst Vijay Bhagavath said. “Cisco and HPE have computing, networking and all the software. Arista is primarily switches.”
The three AI infrastructure providers are targeting three types of buyers, Bhagavath said. First are the largest cloud providers looking for an alternative to expensive InfiniBand products. Second are companies that lease AI data centers, such as Applied Digital, CoreWeave and Equinix. And third are enterprises.
Large enterprises, such as pharmaceuticals, banks and government-funded research institutions, are using LLMs in their data centers. On the other hand, small to midsize organizations are exploring the potential of AI through small models designed to enhance various aspects of their operations, including office productivity, customer service and sales.
Antone Gonsalves is an editor at large for TechTarget Editorial, reporting on industry trends critical to enterprise tech buyers. He has worked in tech journalism for 25 years and is based in San Francisco.
