DGX Spark vs Dell Pro Max GB10 vs ASUS GX10 vs HP ZGX Nano vs Lenovo PGX

By AiCybr Editorial · Published August 10, 2026

The NVIDIA DGX Spark, Dell Pro Max with GB10, ASUS Ascent GX10, HP ZGX Nano and Lenovo ThinkStation PGX are much closer internally than their branding suggests. All are built around NVIDIA’s GB10 Grace Blackwell platform, with a 20-core Arm CPU, Blackwell GPU and 128 GB of coherent unified memory.

That means buyers should not expect one vendor’s GB10 box to deliver a fundamentally different class of raw model capacity from another. The useful differences are elsewhere: storage options, I/O implementation, chassis/serviceability, vendor software, support, regional availability and price.

Quick verdict

For most local-LLM buyers, the first question should therefore be whether GB10 itself fits the workload. Only then should the vendor comparison decide the purchase.

What all five systems share

NVIDIA’s current DGX Spark documentation describes the GB10 platform with:

• a 20-core Arm CPU using 10 Cortex-X925 and 10 Cortex-A725 cores;
• a Blackwell-generation GPU with fifth-generation Tensor Cores;
• 128 GB LPDDR5X coherent unified memory;
• a 256-bit memory interface;
• 273 GB/s memory bandwidth;
• up to 1 PFLOP FP4 tensor performance with sparsity;
• 10 GbE plus NVIDIA ConnectX-7 high-speed networking;
• NVIDIA DGX OS and the NVIDIA AI software stack.

NVIDIA positions one GB10 system for models up to roughly 200 billion parameters, and a paired two-system configuration for models up to 405 billion parameters. Those numbers are platform-support claims, not promises that every 200B or 405B model will run at a useful speed or with any arbitrary context length.

Model practicality still depends on weight precision, quantization, KV-cache size, runtime overhead, framework support and whether the model architecture is supported efficiently on GB10.

The comparison table

The table deliberately avoids pretending that shared GB10 hardware becomes a different GPU because it is placed in a different enclosure. Where a vendor does not publish a separate memory-bandwidth figure, the system still uses the same GB10 memory architecture; however, buyers should use the vendor’s own specification sheet for warranty-critical details rather than assuming every implementation is identical in every subsystem.

NVIDIA DGX Spark: the reference system

DGX Spark is the easiest baseline because NVIDIA documents the platform directly.

The current hardware guide lists:

• 128 GB LPDDR5X unified memory;
• 273 GB/s bandwidth;
• 1 TB or 4 TB NVMe M.2 storage;
• 10 GbE;
• ConnectX-7 networking;
• Wi-Fi 7;
• four USB-C ports;
• HDMI 2.1a;
• a 240 W external power supply;
• 140 W GB10 SoC TDP;
• 150 × 150 × 50.5 mm chassis dimensions.

NVIDIA also states that partner GB10 systems may not receive firmware/software updates at exactly the same time as the Founders Edition. That is a small but real reason to prefer the NVIDIA box if you want the reference update path.

In February 2026 NVIDIA raised DGX Spark’s Founders Edition MSRP from $3,999 to $4,699, citing worldwide memory-supply constraints. NVIDIA explicitly said the price change did not include a hardware/configuration change and did not dictate OEM GB10 pricing.

Dell Pro Max with GB10: strongest business-PC style support path

Dell’s current US product page lists the Pro Max with GB10 with the same 20-core Grace CPU, Blackwell GPU and 128 GB LPDDR5X memory.

The practical differentiators are more conventional workstation details:

• 1 TB, 2 TB and 4 TB SSD configuration options on Dell’s configurator;
• 10 GbE;
• ConnectX-7 with two 200G QSFP ports;
• three USB-C Gen 2x2 ports with DisplayPort Alt Mode plus a USB-C power input;
• HDMI 2.1b;
• Wi-Fi 7 and Bluetooth;
• Dell support/warranty purchasing options.

Dell’s owner manual also exposes the SSD as a field-replaceable component, which matters more for a long-lived workstation than a marketing benchmark does.

As of August 10, 2026, Dell’s US store listed a 2 TB configuration at $5,688.18 and a configurable 4 TB path above $6,000. Prices and promotions are volatile, so treat those as current US-store observations, not fixed global pricing.

If your organization already standardizes on Dell service and procurement, that support path may justify paying more than the bare reference system.

ASUS Ascent GX10: the most flexible published storage menu

ASUS’s current product specifications list the Ascent GX10 with:

• NVIDIA GB10 Grace Blackwell;
• 128 GB LPDDR5X coherent unified memory;
• 1 TB PCIe 4.0, 2 TB PCIe 4.0 or 4 TB PCIe 5.0 SSD options;
• 10 GbE;
• ConnectX-7;
• three USB 3.2 Gen 2x2 Type-C ports plus USB-C power input;
• HDMI 2.1;
• up to 240 W power-adapter output, with the device specified for up to 180 W input;
• 150 × 150 × 51 mm dimensions.

ASUS publicly announced availability in October 2025, so this is not a future or concept system.

The GX10 is therefore attractive if you want a GB10 box but prefer ASUS’s storage configuration, chassis and support ecosystem. It should not be bought under the assumption that its 1 PFLOP headline represents more GB10 compute than DGX Spark; that figure comes from the common platform capability.

HP ZGX Nano G1n: the software-workflow differentiator

HP’s current US store confirms the same fundamental GB10 recipe: 20-core Arm CPU, Blackwell GPU and 128 GB coherent unified memory.

HP’s main differentiator is ZGX Toolkit. HP describes it as a workflow layer with open-source frameworks, MLflow tracking, Ollama testing, system discovery, synchronization/export and local serving. The current product page says the toolkit is free, but its client requirements include an x86 Windows 11 or Ubuntu 24.04-or-later system with Visual Studio Code.

HP offers current store configurations with 2 TB and 4 TB SSDs, while its broader product description documents 1 TB or 4 TB self-encrypting storage choices depending on region/configuration.

This makes the ZGX Nano particularly interesting for teams that want the GB10 device to behave more like a network-attached AI development appliance managed from an existing workstation.

Do not confuse that software convenience with more unified memory or a faster GB10 GPU. The underlying model-capacity class remains shared.

Lenovo ThinkStation PGX: enterprise workstation integration

Lenovo’s product guide is unusually detailed. It confirms:

• 20-core NVIDIA Grace Arm CPU;
• Blackwell GPU;
• 128 GB LPDDR5X unified memory;
• 256-bit memory bus;
• 273 GB/s bandwidth;
• 1 TB or 4 TB self-encrypting NVMe storage;
• 10 GbE;
• two ConnectX-7 QSFP ports;
• three USB-C 20 Gb/s ports plus a USB-C power connection;
• HDMI 2.1a;
• Wi-Fi 7;
• TPM 2.0 and Secure Boot;
• 240 W power supply;
• 150 × 150 × 50.5 mm chassis;
• starting weight of about 1.2 kg.

Lenovo also positions PGX as part of its broader Hybrid AI workstation and infrastructure portfolio and offers ThinkStation-style support services such as Premier Support and Keep Your Drive.

For corporate buyers already standardized on Lenovo, those operational details can matter more than a minor chassis or storage difference.

Which one should you buy?

Choose DGX Spark if you want the cleanest reference platform

It is the simplest choice when you value first-party NVIDIA documentation, reference firmware/software timing and the least ambiguity about what the GB10 platform is supposed to look like.

Choose Dell Pro Max with GB10 if serviceability and enterprise support matter

Dell’s configurable storage, documented replaceable SSD and familiar enterprise support stack make it a practical workstation purchase. The tradeoff is price: the current US configurations can sit well above NVIDIA’s Founders Edition MSRP.

Choose ASUS GX10 if storage flexibility and ASUS hardware support appeal to you

The 1/2/4 TB official storage ladder is useful, and ASUS’s 4 TB option uses PCIe 5.0 in its published specification. It remains fundamentally a GB10 system, so choose it for implementation details rather than imagined extra AI silicon.

Choose HP ZGX Nano if HP’s software workflow is valuable

ZGX Toolkit is the most distinct vendor software layer in this group. If your intended workflow involves managing a dedicated AI station from an existing laptop or workstation, that can be more important than small hardware differences.

Choose Lenovo PGX if you want ThinkStation integration

PGX combines the same GB10 platform with detailed enterprise support, security and Lenovo infrastructure integration. It is a sensible choice for organizations already buying ThinkStation systems.

What GB10 is good at

The unusual part of GB10 is not the vendor logo. It is the 128 GB coherent memory pool.

A conventional consumer GPU may deliver far more memory bandwidth and excellent token throughput but expose only 24–32 GB of VRAM. GB10 trades some of that raw discrete-GPU bandwidth for a much larger shared memory space in a compact system.

That makes this class particularly interesting for:

• local inference with models that do not fit comfortably in normal gaming-GPU VRAM;
• model prototyping and fine-tuning workflows that value capacity over maximum tokens/second;
• private RAG/agent workloads with large model footprints;
• teams that want an always-on local inference appliance;
• ARM64-native AI development using NVIDIA’s software stack.

It is not automatically the fastest local-AI choice for models that already fit in a high-end discrete GPU. Model throughput depends heavily on memory bandwidth, kernels, quantization, batching and runtime maturity.

The 200B and 405B claims need context

All five vendors repeat some version of NVIDIA’s model-size positioning. The useful interpretation is addressable model capacity, not guaranteed interactive performance.

For example, a heavily quantized model can have a weight footprint far below its original BF16 size, while KV cache, runtime workspaces and context length add memory pressure back. A model that technically loads is not necessarily a pleasant interactive experience.

A two-node 405B configuration also does not turn two 128 GB systems into one transparent 256 GB desktop memory pool for arbitrary software. The runtime must support distributed execution across the high-speed interconnect.

So use vendor parameter counts as a compatibility ceiling to investigate, not a benchmark.

GB10 vs an RTX workstation

This is the decision many buyers actually need to make.

GB10 advantage: memory capacity. A 128 GB coherent pool lets you attempt models and contexts that simply cannot reside fully in a 24–32 GB discrete GPU.

RTX workstation advantage: for models that fit, a high-end discrete GPU can provide much higher dedicated-memory bandwidth and very strong mature desktop software support.

Other differences matter too:

• GB10 is Arm64, not x86-64;
• some third-party tools, Python wheels or containers may need Arm64 support;
• the GB10 devices are compact appliances rather than highly expandable towers;
• consumer/workstation RTX systems offer more CPU, PCIe, storage and upgrade flexibility depending on the build.

The right purchase therefore depends less on the headline 1 PFLOP number and more on whether your limiting resource is memory capacity or per-token throughput.

Buying checklist

Before buying any GB10 system, verify:

1. Your actual model footprint. Include quantization, KV cache, context length and runtime overhead.
2. Arm64 compatibility. Check the exact inference engine, container images, Python packages and plugins you use.
3. Storage requirement. Large model collections can consume terabytes quickly.
4. Network plan. If you expect to pair two units, budget for the required high-speed cabling and confirm runtime support.
5. Local support and warranty. Partner systems may be easier to service in your region than a Founders Edition unit.
6. Current price. OEM and regional pricing moves independently; do not assume the cheapest vendor globally is cheapest locally.
7. Whether you actually need 128 GB. If your workloads fit inside a conventional high-end GPU, compare real inference throughput before paying for unified-memory capacity.

Bottom line

The five major GB10 desktop systems are variants of one compute platform, not five fundamentally different AI accelerators.

Choose DGX Spark for the reference NVIDIA experience, Dell for business support/serviceability, ASUS for its storage/chassis implementation, HP for ZGX Toolkit, or Lenovo for ThinkStation integration. But do not let vendor branding obscure the main architectural decision: GB10 is primarily compelling when 128 GB of coherent memory solves a model-capacity problem that a conventional discrete GPU cannot.

If your models already fit in ordinary GPU VRAM, compare real application throughput and total system cost instead of assuming that the words “AI supercomputer” make the GB10 box automatically faster.

Sources

• NVIDIA DGX Spark hardware guide
• NVIDIA DGX Spark support specifications
• NVIDIA DGX Spark release notes
• NVIDIA February 2026 DGX Spark price-change announcement
• Dell Pro Max with GB10
• Dell Pro Max with GB10 owner documentation
• ASUS Ascent GX10 availability and specifications
• HP ZGX Nano G1n
• Lenovo ThinkStation PGX product guide