NVIDIA DGX Spark vs Lenovo ThinkStation PGX: Which GB10 System Fits?

Direct answer: DGX Spark is the safer choice when you want NVIDIA’s first-party software image, an identified 128 GB configuration, and a purchase path that is already documented. ThinkStation PGX is compelling when Lenovo’s service relationship, procurement process, or regional support matters more than owning the NVIDIA-branded enclosure. Both systems target the same broad GB10 class, but a shared chip does not make their software, warranty, shipping, or final configuration identical.

This is a specification analysis, not a benchmark review. Neither system was physically tested for this article. Lenovo configurations and availability can vary by country and sales channel, so the final quote—not a headline configuration—should be treated as the procurement record.

The short comparison

The table separates known platform facts from configuration details that require a quote. Memory, storage, ports, service, and operating-system images can be changed by an OEM without changing the GB10 name.

The common ground: GB10 and 128 GB are the starting point

NVIDIA describes DGX Spark as a desktop system based on the GB10 Grace Blackwell Superchip. Its published specification includes a 20-core Arm CPU, a Blackwell GPU, 128 GB of coherent LPDDR5x system memory, 273 GB/s of memory bandwidth, and 4 TB of NVMe M.2 storage. NVIDIA positions the system for local model development, inference, fine-tuning, data science, and autonomous-agent workloads.

Lenovo’s ThinkStation PGX materials describe a workstation based on the same GB10 family. The practical appeal is familiar: a Lenovo-branded workstation can fit an existing purchasing, asset-management, and service process. The exact experience still depends on the system configuration and the software image Lenovo ships in the buyer’s region.

FACT: A common GB10 platform makes the systems related.

ANALYSIS: It does not establish equal application performance, identical firmware, identical thermals, or identical support terms. Treat the model name as a starting point for due diligence.

Software is the first buying gate

DGX Spark’s strongest differentiator is the first-party NVIDIA environment. NVIDIA ships DGX OS and an AI software stack that includes CUDA-oriented libraries, containers, tools, and models. A team that already deploys to NVIDIA data-center systems can use the same conceptual ecosystem during local development. This can reduce translation work between a desktop prototype and a production target.

The caveat is architecture. GB10 uses an Arm CPU, so every important dependency must be checked for Linux Arm64 support. Review Python wheels, native extensions, compilers, database clients, monitoring agents, licensed binaries, and container base images. A repository that says “CUDA supported” may still fail because an adjacent package only publishes x86_64 artifacts. NVIDIA’s DGX Spark dependency guide should be part of the acceptance checklist.

ThinkStation PGX may arrive with a Lenovo-selected software configuration rather than exactly the same DGX OS lifecycle. Confirm the following in writing:

• the operating system and its update channel;
• whether NVIDIA drivers and CUDA libraries are preinstalled;
• which container runtime and toolkit versions are supported;
• how Arm64 package compatibility is handled;
• whether Lenovo support covers the AI software stack or only the hardware.

If the project is sensitive to reproducibility, record the image version, firmware version, driver version, and a tested container digest. That record is more useful than assuming two GB10 systems are interchangeable.

Memory capacity is similar, but the workflow can differ

Both products are aimed at local AI developers who value a large unified memory pool. Unified memory lets the CPU and GPU work from a shared system allocation, which can be useful when a model’s weights and runtime overhead would not fit in an ordinary desktop GPU.

Do not turn 128 GB into a universal model-size promise. Weights, quantization metadata, key-value cache, context length, batch size, temporary tensors, and the serving engine all consume memory. A vendor statement about a model category is not a guarantee of acceptable token throughput for your context window or concurrency.

DGX Spark’s 4 TB storage standard is useful for model files, container layers, datasets, and local logs, although a serious team still needs a backup plan. Lenovo’s listed PGX material identifies 1 TB NVMe storage, which may be adequate for a focused development image but can become restrictive when several quantized checkpoints and datasets coexist. Ask whether the quoted unit can be configured differently, whether storage is replaceable, and how that affects warranty.

Networking and team use

NVIDIA lists 10 GbE and ConnectX-7 networking for DGX Spark, with a high-speed path intended to connect systems and support larger model workflows. That matters when the box is used as a shared target, when datasets live on a fast network, or when multiple systems are combined for experimentation.

For PGX, verify the exact Ethernet, Wi-Fi, ConnectX, and cable options in the Lenovo regional quote. Do not infer the port layout from a photograph or from a different market’s PSREF. If the workstation will sit in a small office, check switch compatibility, noise, rack or desk placement, and remote-management capabilities before treating networking as a solved problem.

Procurement, warranty, and availability

NVIDIA’s product page directs buyers to its Marketplace and authorized partners. That gives DGX Spark a recognizable reference configuration, but regional stock, tax, shipping, and partner terms still vary.

Lenovo can be the better operational choice for an organization that already has a Lenovo purchasing account, standard warranty process, imaging workflow, and local field-service relationship. Those advantages are real only if the AI workstation is included in the same support scope. Confirm response times, parts replacement, firmware delivery, and escalation ownership. A workstation warranty that excludes the software environment may leave the team responsible for the hardest compatibility problem.

Do not compare an NVIDIA Marketplace price with a Lenovo quote until the two configurations include equivalent memory, storage, support, shipping, and tax. A lower initial quote may omit an extended warranty, the desired network adapter, or the setup work needed to make a production-like environment.

Which buyer should choose which?

Choose DGX Spark when:

• CUDA and NVIDIA’s software stack are non-negotiable;
• a documented 128 GB / 4 TB configuration fits the workload;
• you want a dedicated local target that resembles an NVIDIA reference environment;
• you expect to connect with other DGX Spark systems or fast storage;
• the team can own Arm64 dependency validation.

Choose ThinkStation PGX when:

• Lenovo is already your approved workstation supplier;
• regional service, asset management, or procurement simplicity has high value;
• the quoted configuration includes the memory, storage, OS, and networking you need;
• Lenovo can document support boundaries for NVIDIA software;
• you prefer an OEM workstation relationship to a direct NVIDIA product channel.

Choose neither yet when:

• the workload is bursty and a cloud GPU experiment would answer more questions;
• the team has not tested its exact model, context, and serving stack on Arm64;
• the project requires more than 128 GB in one node;
• the purchase depends on a price or availability claim that is not in a current quote.

A practical acceptance test

Before purchasing either system, prepare one reproducible test package: the model and quantization format, a representative prompt set, target context length, concurrency, container digest, expected time to first token, generation throughput, and peak memory. Run it on the actual configuration or request a partner demonstration. Also test cold-start time, storage throughput, remote access, update behavior, and recovery after a power interruption.

The result should be a small procurement record rather than a marketing score. It should answer whether the system can complete the real job, whether the software can be maintained, and whether the support path is acceptable.

OPINION: DGX Spark is the clearer default for a technically self-sufficient team. ThinkStation PGX becomes the stronger choice when Lenovo’s service and procurement advantages are material and the quote proves software and configuration parity. GB10 is the common ingredient; the ownership experience is still a system-level decision.

For the next comparison in this series, see DGX Spark vs ASUS Ascent GX10. For a broader view of the category, compare DGX Spark vs Mac Studio and DGX Spark vs cloud GPUs.

Sources and verification note

DGX Spark specifications and positioning were checked against NVIDIA’s DGX Spark page and NVIDIA’s Arm64 dependency guide on August 30, 2026. Lenovo product information was checked against Lenovo PSREF and Lenovo Press on the same date. Verify the current regional configuration, price, availability, warranty, and software support before purchase.