Next up in the AI Revolution: Hitachi iQ

Introduction

Every enterprise AI roadmap eventually runs into the same wall, and it is rarely a lack of ideas or data. It is capacity. The model you want to fine-tune does not fit in memory. Inference slows down as more users arrive. A training run that should take days stretches into weeks. In most of these cases the bottleneck is the GPU: how much memory each accelerator holds, how quickly it can move data, and how many accelerators can work together without waiting on the network.

That is why GPU capacity has become the number that matters most when choosing AI infrastructure. Compute counts, but a GPU that cannot hold your model, or the data and context the model needs while it runs, forces trade-offs that show up as higher costs, slower results, and more complicated architecture.

servers hitachiIQ

This is the problem the new Hitachi iQ servers are built to solve. As a Hitachi Vantara Premier Partner, Sannet Solutions is making these nodes available with AMD Instinct MI325X and MI350X GPUs, combined with AMD Turin CPUs, 200GbE or 400GbE networking, and high-capacity NVMe flash in an 8U chassis. In this article we explain why demand for AI infrastructure keeps accelerating, why GPU memory capacity is the specification to put first, and what the new nodes mean for CIOs, IT directors, and data science teams planning their next deployment.

Why AI infrastructure demand is surging?

The scale of the investment is hard to overstate. Gartner forecasts that global data center spending will surpass $650 billion in 2026, up from nearly $500 billion the year before, with server spending alone growing almost 37%. IDC reports that AI infrastructure spending reached $89.7 billion in the first quarter of 2026 alone, up 33% year over year, and it raised its full-year 2026 forecast to $497 billion. IDC puts the 2025 total at $318 billion, more than double the $153 billion spent in 2024.

Several forces are driving that curve.

  • AI is moving from pilot to production. Companies that spent 2024 and 2025 experimenting with large language models (LLMs) are now deploying them in customer service, software development, document analysis, and operations. Production workloads run continuously and serve many users at once, which is far more demanding than a proof of concept.
  • Models and context windows keep growing. Larger models, longer prompts, and retrieval-augmented pipelines all consume more GPU memory, both for the model weights and for the working memory used during inference.
  • Data has to stay close to the compute. Enterprises increasingly want to run AI on their own data, on their own premises, for reasons of privacy, regulation, latency, and cost predictability. That puts the infrastructure decision in the hands of IT leaders rather than only cloud providers.
  • Supply is being reshaped. As AI servers add more memory and storage, those components have become a larger share of the overall bill, which makes choosing the right configuration the first time more important.

The accelerator market has responded with a race to add capacity. When AMD launched the Instinct MI325X, ITPro reported that it carries 256GB of HBM3E memory with 6TB/s of bandwidth, and that AMD claimed up to 1.3x the inference performance on the Mistral 7B model and 1.2x on Llama 3.1 70B. AMD aimed the chip directly at Nvidia’s market-leading H200 generation. At Computex, DatacenterDynamics reported that AMD’s comparison against the competition cited roughly double the memory capacity, about 1.3x the memory bandwidth, and about 1.3x the AI compute performance. The shipping MI325X carries 256GB of memory, slightly below the 288GB figure quoted at that early announcement, so the capacity advantage is smaller than the earliest headline. These are vendor figures, and your own workloads should be benchmarked, but they show where the industry is heading: more memory per GPU, faster.

The MI350 series pushed this further. AMD introduced the MI350X and MI355X at its Advancing AI event in June 2025, alongside ROCm 7, the latest version of its open-source AI software stack.

Why GPU capacity is the specification to put first

server HitachiIQ

When teams compare AI servers, they often start with peak compute figures. Capacity deserves equal or greater attention, for four reasons.

  1. It sets the largest model you can run. Model weights have to live in GPU memory. If they do not fit on one node, the model must be split across nodes, which adds networking overhead, complexity, and cost.
  2. It determines how many users you can serve. During inference, each active conversation needs working memory for its context. More memory per GPU means more simultaneous users and longer prompts on the same hardware.
  3. It reduces data movement. Higher-bandwidth memory feeds the compute units faster. For many LLM workloads, memory bandwidth, not raw compute, is what limits throughput.
  4. It protects your investment. Models are growing every year. A node with generous memory headroom stays useful longer, while a tight fit can force an early refresh.

This is where the MI325X and MI350X stand out. The MI325X offers 256GB of HBM3E per GPU, and the MI350X offers 288GB of HBM3E on AMD’s newer CDNA 4 architecture, built on a 3nm process with support for FP4 and FP6 data types that make inference more efficient. AMD’s reference platforms group eight accelerators on a single baseboard, which for the MI325X works out to 2TB of GPU memory in one system. [Confirm the number of GPUs per Hitachi iQ node before publishing.]

What’s new in Hitachi iQ

Hitachi iQ launched as a family of industry-specific AI systems that combine Hitachi Vantara storage with GPU servers, and it initially centered on NVIDIA hardware, including HGX systems with H100 and H200 GPUs. The new AMD-based nodes extend the portfolio, giving enterprises a second high-end accelerator option within the same ecosystem. They are built on an 8U chassis with an AMD motherboard.

How to get started with Sannet Solutions

The new Hitachi iQ nodes are ready to be configured and deployed. Our senior engineering team can walk you through the technical specifications, help you size GPU capacity for your workloads, and design a configuration that fits your environment.

Frequently asked questions

What is the difference between the MI325X and the MI350X? The MI325X has 256GB of HBM3E memory and 6TB/s of bandwidth. The MI350X offers 288GB of HBM3E, higher bandwidth, and a newer CDNA 4 architecture with FP4 and FP6 support.

Does Hitachi iQ support AMD GPUs? Yes. The new Hitachi iQ nodes available through Sannet Solutions integrate AMD Instinct MI325X and MI350X GPUs.

Why does GPU memory matter for LLMs? Model weights and the working context for each request must fit in GPU memory. More memory means larger models, longer prompts, and more concurrent users per node.

What networking do the nodes use? Dedicated 200GbE or 400GbE NICs connect the GPUs to your data and storage environment.

Can I see the technical specifications? Yes. Contact Sannet Solutions’ engineering team for the full specifications or to configure a node.

Leave feedback about this

  • Quality
  • Price
  • Service
Choose Image