How Dell GB10 Pro Max Signals the Shift Towards Real On-Device AI

For years, the AI community expected progress to come only from bigger clusters and larger cloud deployments.

Instead, a parallel trend has reshaped how developers build models, leading to small and mid-sized language models becoming dramatically more capable.

This shift has reopened an old question with new urgency: If developers can do more with smaller models, why is much of AI development still locked behind remote, expensive and capacity-constrained infrastructure?

Local computing has struggled to keep pace. Even top-end workstations hit memory ceilings well before loading these improved models.

Teams working on 30B or 70B parameter models often find that their hardware forces them to use compression techniques, model sharding or external GPU servers. 

For regulated industries, none of those workarounds are straightforward because moving data off-premise is restricted. For researchers and startups, accessing cloud instances becomes a recurring drain on budgets and iteration velocity. 

The gap between what models can do and what local machines can support has grown wider.

To address challenges in this space, hardware manufacturers like Dell have invested significant effort. The company’s latest Dell Pro Max accelerated by NVIDIA GB10 Grace Blackwell Superchip is a response to developers capable of building more ambitious on-device AI but blocked by hardware limits. 

“Training models with more than 70 billion parameters demands computational resources far beyond what most high-end workstations deliver,” the company said. 

By bringing NVIDIA’s Grace Blackwell architecture—previously limited to data centres—into a deskside form factor, Dell is attempting to realign hardware with this new generation of compact but computationally demanding AI workloads.

The device is designed to give AI developers high performance, simplicity, security, and cost control in a single, reliable setup.

The Dell Pro Max with GB10 ships with 128GB of unified LPDDR5X memory and runs on Ubuntu Linux with NVIDIA DGX OS, preconfigured with CUDA, Docker, JupyterLab and the NVIDIA AI Enterprise stack. 

Dell says the system delivers up to 1,000 trillion operations per second (TOPS) of FP4 AI performance, giving developers the headroom to fine-tune and prototype models up to 200B parameters locally.

“Packing this much power into a compact 1.2-kg device, measuring just 150 mm by 150 mm by 50.5 mm, represents a significant engineering achievement,” the company stated. 

Besides, unified memory avoids many bottlenecks that arise from juggling separate CPU/GPU memory pools and lets developers work with large models in a single address space.

The value is practical rather than theoretical. Academic labs can run Meta’s open source Llama-class models without waiting for shared clusters. 

Startups can experiment with product features locally instead of committing to cloud spend during early R&D. Banks and healthcare organisations can build AI systems while keeping data inside their compliance perimeter. Independent developers can test and refine models that were previously out of reach without renting external GPUs. 

As Dell put it, the Pro Max with GB10 is designed to “streamline development” and eliminate the recurring problem of local devices hitting their limits too early in the workflow.

When working with larger-scale AI workloads, Dell also shared a potential solution. “Teams that need more capacity can bond two GB10 systems to act as a single node, accommodating models of up to 400 billion parameters,” Dell stated. 

It added that teams can get started quickly as DGX OS comes preconfigured. They can launch training jobs within minutes, use additional SDKs and orchestration tools as needed, and pull model checkpoints directly from the NVIDIA Developer portal and the NGC catalogue.

It is an AI development node, and a tool for running and iterating on models directly, not an all-purpose PC. Teams without experience in machine learning stacks or DGX-style workflows will face a learning curve.

Still, the direction is consistent with how the AI ecosystem is shifting. As more capable smaller models emerge, on-device AI is becoming viable for tasks that previously required remote compute. Developers want faster iteration, predictable costs and greater control over data. 

The device exists to meet those conditions. It does not replace large clusters for final training runs, but it changes what can happen locally in the earliest, most experimental stages.

“Personal computers have brought software development to everyone. Cloud computing has made large-scale apps easy to access,” Dell added. “Now, desktop supercomputing will make advanced AI engineering possible for anyone ready to explore.”