NVIDIA’s $3,000 AI Desktop Is Now a $4,699 DGX Spark—What It Really Does

Yes, NVIDIA announced a personal AI computer—but the $3,000 figure is no longer its current price. NVIDIA introduced the machine as Project DIGITS at CES on January 6, 2025, promising a May launch at a starting price of $3,000. It was renamed DGX Spark in March 2025, began shipping in October, and the U.S. Founders Edition now has a $4,699 MSRP after a February 2026 price increase attributed to memory-supply constraints.

DGX Spark is best understood as a compact, Linux-based NVIDIA AI workstation with 128GB of coherent unified memory. It can hold models that will not fit into the 32GB of a single GeForce RTX 5090, but that does not make it the fastest AI computer or a miniature data center. Its strongest case is local CUDA development, large-model inference, fine-tuning, robotics, computer vision, and testing software before moving it to NVIDIA cloud or data-center hardware.

What NVIDIA actually announced

The original announcement was real, but many descriptions of it are now outdated. NVIDIA presented Project DIGITS as a personal AI computer for researchers, data scientists, developers, and students. Its selling point was bringing a Grace Blackwell-based AI development environment to a desk rather than requiring every experiment to run on a rented cloud GPU.

#	Preview	Product	Price
1		Dell NVIDIA Tesla V100 GPU SXM2 32GB NWWWX by DELL	$854.96	Buy on Amazon
2		Gigabyte NVIDIA GeForce RTX 3060 Gaming OC V2 Graphics Card - 12GB GDDR6, 192-bit, PCI-E 4.0,...	$695.00	Buy on Amazon

The product’s name, delivery schedule, and price subsequently changed:

The original $3,000 price should therefore be treated as a historical launch-announcement price, not as the current cost of the NVIDIA Founders Edition. Partner systems can have different prices because of their SSD capacity, cooling, warranty, retailer, region, and inventory. NVIDIA’s February 2026 price announcement specifically applied the adjustment to the Founders Edition and did not state that every partner system would have the identical price.

#1 Best Overall

Dell NVIDIA Tesla V100 GPU SXM2 32GB NWWWX by DELL

• GPU Chipset: NVIDIA
• Memory: HBM2
• Programming Interface: CUDA
• Memory Capacity: 32GB
• Slot Compatibility: SXM2

$854.96

Buy on Amazon

DGX Spark specifications

At roughly 150mm square and 50.5mm tall, DGX Spark is much smaller than a conventional workstation. The Founders Edition weighs 1.2kg and uses a 240-watt external power adapter. Its core specification is the GB10 Grace Blackwell Superchip, which combines an Arm CPU and Blackwell GPU in a unified-memory design.

See NVIDIA’s DGX Spark hardware documentation for the current specifications. OEM systems share the GB10 platform but are not necessarily identical in storage, firmware, cooling, chassis design, warranty, or support.

Why the 128GB unified-memory design matters

DGX Spark does not have 128GB of dedicated video memory. The accurate description is 128GB of coherent unified memory shared by the CPU and GPU. The processor, graphics engine, operating system, applications, caches, and model all draw from the same pool.

That arrangement removes one major limitation of a conventional desktop GPU: the fixed boundary between system RAM and VRAM. A workstation with a 32GB RTX 5090 may have plenty of ordinary system RAM, but a model that needs more than 32GB of fast GPU memory may have to be split, offloaded to system memory, quantized more aggressively, or distributed across multiple GPUs. DGX Spark can keep a much larger model in one shared address space.

The trade-off is bandwidth. DGX Spark’s stated 273GB/s is substantially below the bandwidth available on high-end discrete GPUs and Apple’s larger unified-memory systems. A model fitting in memory is not the same as a model running quickly. During inference, especially token-by-token decoding, memory bandwidth and data movement can matter more than peak tensor-core arithmetic.

There is also less than 128GB available to a model in practice. The operating system, display, processes, runtime, and other workloads consume part of the pool. NVIDIA documents that memory reporting on this platform can differ from conventional discrete-GPU systems: nvidia-smi may report GPU memory usage as unsupported even though per-process memory information is available. The operating system may also reclaim or swap pages, so reported and actually allocatable memory are not always identical. NVIDIA’s known-issues documentation and DGX Spark documentation explain these behaviors.

What “1 petaflop” means—and what it does not

NVIDIA advertises up to 1 PFLOP, or 1,000 trillion operations per second, of AI performance. That figure needs three important qualifications:

1. It is an FP4 figure. FP4 is a very low-precision numerical format. A peak figure in FP4 should not be compared directly with a GPU’s BF16, FP8, FP16, or general-purpose floating-point rating.
2. It uses sparsity. The advertised number assumes the chip’s sparse-computation capability. It is not necessarily the result for a dense workload.
3. It is theoretical arithmetic throughput. Real applications can be limited by memory bandwidth, model architecture, quantization, software kernels, context length, thermal behavior, and CPU/GPU data movement.

Consequently, “1 petaflop” is not a promise of one petaflop of BF16 training, a particular number of generated tokens per second, image-generation speed, or gaming performance. NVIDIA’s DGX Spark support page identifies the precision and sparsity context behind the specification.

What models can DGX Spark run?

NVIDIA says one DGX Spark can perform inference on models of up to 200 billion parameters and locally fine-tune models of up to 70 billion parameters. Two systems can be linked through their high-speed networking for models of up to 405 billion parameters. These are important capacity claims, but they are not speed guarantees. NVIDIA describes the figures in its shipping announcement.

Whether a particular model fits and performs usefully depends on:

• Quantization format and implementation
• Context length and the resulting KV-cache size
• Operating-system and application overhead
• Dense versus mixture-of-experts architecture
• Inference, fine-tuning, or full training requirements
• Model weights, activations, optimizer state, and temporary buffers
• Whether the model and software kernels are optimized for GB10

A 200-billion-parameter model may fit only after quantization and still generate tokens slowly. A mixture-of-experts model can have a large total parameter count while activating fewer parameters for each token. Conversely, a long context window can consume a substantial amount of memory even when the model weights fit comfortably.

Inference is not fine-tuning, and fine-tuning is not training from scratch

These distinctions matter:

• Quantized inference: Loading a model and generating responses. This is the least memory-intensive of the three activities, although long contexts and concurrent users increase memory requirements.
• Fine-tuning: Adapting an existing model to a task or dataset. Techniques such as parameter-efficient fine-tuning can reduce requirements, but fine-tuning still needs additional memory and compute beyond ordinary inference.
• Training from scratch: Creating a foundation model. This requires vastly more compute, storage, memory, time, and often many accelerators. DGX Spark is not a replacement for a multi-GPU training cluster.

Therefore, “supports up to 200B models” should not be read as “comfortably runs every 200B model.” NVIDIA’s 70B figure refers to its local fine-tuning claim, not to training a 70B foundation model from scratch. The 405B figure applies to two linked systems, not to one box.

What is DGX Spark intended to do?

NVIDIA positions DGX Spark as an AI development node for:

• Prototyping and testing AI models
• Local LLM and multimodal-model inference
• Fine-tuning existing models
• AI-agent development
• Robotics and physical-AI development
• Computer vision
• Image and video generation
• Data science
• Preparing code and models for cloud or data-center deployment

The intended workflow is local first, cloud or data center later:

1. Install or develop the application locally on DGX Spark.
2. Experiment with models, prompts, data pipelines, and agents without sending every iteration to a cloud service.
3. Fine-tune or validate the workflow on the local system where practical.
4. Move the same code, containers, and model workflow to DGX Cloud, an accelerated cloud instance, or a larger NVIDIA system when the workload outgrows the box.

This can reduce iteration delays, data-transfer friction, and repeated cloud usage for development. It does not eliminate the need for cloud or data-center compute when the model, dataset, concurrency, or training job exceeds local capacity.

Software: the main reason to choose NVIDIA

DGX Spark is built around NVIDIA’s AI software stack rather than a generic mini-PC operating system. The documented environment includes:

• DGX OS
• CUDA and the CUDA Toolkit
• PyTorch
• TensorRT-LLM
• NVIDIA NeMo
• NVIDIA RAPIDS
• NIM microservices
• NVIDIA Blueprints
• Jupyter and JupyterLab
• Docker with the NVIDIA Container Runtime
• NVIDIA NGC containers and models

Partner and ecosystem applications include Ollama, LM Studio, ComfyUI, Docker, Anaconda, JetBrains tools, Hugging Face software, and Roboflow. The preinstalled and configured NVIDIA Container Runtime allows Docker containers to access the GPU through Docker’s --gpus mechanism.

For a researcher whose code already targets CUDA, TensorRT-LLM, NIM, or NeMo, this environment can be more valuable than a faster general-purpose machine that requires porting or replacing the software stack. The current Founders Edition software versions listed in NVIDIA’s July 2026 release notes are:

• DGX OS 7.5.0
• NVIDIA GPU Driver 580.159.03
• CUDA Toolkit 13.0.2
• Canonical Kernel 6.17
• UEFI 1.110.13

Partner systems may receive firmware and software updates on a different schedule. Enterprise support and NVIDIA AI Enterprise entitlements should also be treated separately from the included development software. A buyer should check the DGX Spark support and licensing documentation rather than assume unlimited enterprise support is included with the hardware.

Is it a normal desktop computer?

Only in the broad form-factor sense. It can be connected to a monitor, keyboard, and mouse, but its strongest use case is closer to a small local server or AI workstation.

DGX Spark can run headlessly over SSH and on a network. Its software assumes Linux and CUDA workflows, and its 20-core CPU uses the Arm64 architecture rather than the x86-64 architecture found in most Windows desktops and traditional workstations. Some x86-specific applications, precompiled binaries, and development assumptions may require a native Arm build, porting, emulation, or a compatible container.

NVIDIA’s DGX Spark porting guide specifically addresses the differences between an ARM64/unified-memory system and a conventional x86-64 computer with a discrete GPU. Buyers should verify their essential packages before purchasing.

It is also not primarily a Windows gaming computer. The system has limited conventional desktop expansion compared with a tower, no user-upgradable memory pool, and an integrated GPU rather than a replaceable PCIe graphics card.

Do not confuse DGX Spark with RTX Spark, a separate 2026 NVIDIA/Microsoft platform aimed at Windows PCs and laptops. RTX Spark is a newer Windows-oriented product category; it does not turn DGX Spark into a Windows mini-PC. NVIDIA describes that separate initiative here.

Independent performance evidence

Official specifications explain what DGX Spark is designed to do. Independent testing is more useful for understanding how the platform behaves in practice. The results below are not interchangeable benchmarks: each used different models, software, quantization, context lengths, and test configurations.

StorageReview: GB10 systems are broadly similar, but implementation matters

StorageReview tested the ASUS Ascent GX10 alongside the NVIDIA Founders Edition and other GB10 systems using vLLM online-serving workloads. The systems tracked closely in core model behavior because they share the same GB10 platform. Differences were more visible in storage, chassis design, cooling, and implementation than in the underlying AI compute.

In the ASUS test system, StorageReview recorded burst temperatures of approximately 87.3°C for the CPU and 82°C for the GPU, while reporting controlled sustained behavior. It also found that the included 1TB Gen4 SSD limited storage performance compared with faster configurations. The practical lesson is to compare an OEM system’s SSD, cooling, warranty, support, and price rather than expect a radically different AI processor from each GB10-branded machine. See the StorageReview assessment.

ServeTheHome: the value is capacity and portability

ServeTheHome measured roughly 40–45 watts at idle on its tested unit, around 60–90 watts in several AI inference workloads, and less than 200 watts under combined CPU/GPU load. It measured noise below 40dBA outside extreme stress testing.

Rank #2

Gigabyte NVIDIA GeForce RTX 3060 Gaming OC V2 Graphics Card - 12GB GDDR6, 192-bit, PCI-E 4.0, 1837MHz Core Clock, RGB, 2X DP 1.4, 2X HDMI 2.1, NVIDIA Ampere - GV-N3060GAMING OC-8GD

• NVIDIA Ampere Streaming Multiprocessors: Building blocks for the world's fastest, most efficient GPUs, the all-new Ampere SM brings twice the FP32 throughput and improved energy efficiency
• 2nd Generation RT Cores - Experience 2x the 1st Generation RT Cores throughput, plus competitive RT and shading for a whole new level of ray-tracing performance
• 【3rd Generation Tensor Cores】Get up to 2X the throughput with structural sparsity and advanced AI algorithms such as DLSS
• Core Clock: 1837MHz
• WINDFORCE 3X Cooler

$695.00

Buy on Amazon

Its review emphasized that the 128GB memory pool and 200GbE networking make large models portable and potentially clusterable. It also highlighted the downside: the system’s lower memory bandwidth means a larger discrete GPU can process some workloads faster. This is a “large models in a small box” product, not necessarily the fastest token generator. Read the ServeTheHome measurements and analysis.

Signal65: efficiency and concurrency can be strong

Signal65’s 2026 testing reported DGX Spark GPU power of 46.6W during GPT-OSS 20B inference, 39.0W during GPT-OSS 120B inference, and approximately 51–52W on its tested Llama 3.3 70B and Qwen3-VL-32B workloads. Total wall power ranged from 106.1W to 123.3W across the tested workloads, compared with 138.4–146.3W for the AMD comparison system.

Signal65 also reported faster prompt processing and higher concurrency performance for DGX Spark in its particular test configuration. Those results support the argument that DGX Spark can be efficient and useful as a local shared inference node, but they should not be treated as a universal ranking against every AMD, Apple, or discrete-GPU system. The full Signal65 report provides the test context.

Tom’s Hardware: Apple can be faster for selected local-LLM workloads

Tom’s Hardware compared an Apple M4 Max Mac Studio with GB10 and AMD Strix Halo systems. In its tested models, the Mac Studio produced higher tokens-per-second throughput than the competing systems, although the size of the advantage varied substantially by model.

On Qwen 3.6-35B-A3B, the M4 Max’s advantage over GB10 was about 25% despite having roughly twice the memory bandwidth. On the dense Gemma 4 12B test, Apple’s advantage was larger. On GPT-OSS 120B, the M4 Max averaged approximately 1.5 times the GB10 throughput in that test. Apple also had compatibility limitations in at least one image-generation workflow using FP8.

This is the central buying distinction: Apple can deliver higher local-LLM throughput in some tests, while DGX Spark may be the more appropriate platform for CUDA-native research, NVIDIA containers, TensorRT-LLM, and a workflow targeting NVIDIA cloud or data-center hardware. See Tom’s Hardware’s comparison for the benchmark-specific results.

Power, thermals, and practical limitations

The supplied 240W adapter is not merely a convenient accessory. NVIDIA says it is required for optimal performance. A lower-rated or different adapter can reduce performance, prevent boot, or cause unexpected shutdowns, according to the hardware documentation.

The adapter’s rating should not be interpreted as a guarantee that every workload will draw 240W. Independent testing has measured lower consumption in many AI workloads, and some retail systems did not reach the full adapter ceiling during combined tests. Power depends on the model, concurrency, CPU work, networking, software, and thermal conditions.

Early user reports and commentary raised questions about power caps, thermal throttling, crashes, and unexpected reboots. However, the cited coverage did not establish one conclusive cause or show that every DGX Spark has the same problem. The fair conclusion is narrower: sustained-load behavior is workload- and implementation-dependent, NVIDIA requires the supplied adapter for best operation, and buyers should review current firmware, cooling, warranty, and return policies. Tom’s Hardware’s reporting documents the concerns without proving a blanket defect.

Storage is another practical consideration. The Founders Edition has one internal M.2 NVMe drive, and OEM systems can use different capacities and SSD models. Large model collections, datasets, checkpoints, and containers can consume 1TB quickly, so a 2TB or 4TB configuration may be preferable even if the compute hardware is otherwise identical.

DGX Spark versus the alternatives

AMD Ryzen AI Max+ 395 and Ryzen AI Halo

AMD’s Ryzen AI Halo developer platform is the closest capacity-oriented alternative. AMD lists a Ryzen AI Max+ 395 with 16 Zen 5 CPU cores and 32 threads, Radeon 8060S integrated graphics with 40 compute units, 128GB of LPDDR5X memory, 256GB/s memory bandwidth, a 2TB SSD, 10GbE, Wi-Fi 7, Linux or Windows 11 support, and a 120W TDP. AMD states a $3,999 retail price for the developer platform and claims support for models up to 200B parameters. The specifications are on AMD’s product page.

AMD is compelling when the lower stated price, Windows 11 option, x86 compatibility, and general-purpose PC flexibility matter. DGX Spark is more compelling when CUDA, NVIDIA containers, TensorRT-LLM, NIM, NeMo, ConnectX networking, or a direct path to NVIDIA deployment are central. AMD’s own performance comparisons are vendor testing; use independent benchmarks for a purchase decision.

Apple Mac Studio

Apple’s Mac Studio offers a more conventional desktop experience and much higher memory bandwidth. Apple lists M4 Max configurations with up to 128GB of unified memory, M3 Ultra configurations with up to 512GB, and up to 819GB/s of memory bandwidth on M3 Ultra systems. Apple’s U.S. store listing gives the M3 Ultra a $3,999 starting price with 96GB of memory and 1TB of storage. See the Mac Studio specifications.

Choose Apple for high memory bandwidth, strong CPU and media performance, large-memory configurations, and a polished general desktop. Choose DGX Spark for CUDA-native development, NVIDIA-specific libraries and containers, and software that will ultimately run on NVIDIA cloud or data-center GPUs. Apple’s higher tokens-per-second result in selected tests does not remove its compatibility limitations for every AI framework or precision mode.

An RTX 5090 workstation

The GeForce RTX 5090 has 32GB of GDDR7 and an official NVIDIA-listed Founders Edition price of $1,999, although the cited NVIDIA Marketplace listing showed it out of stock. Check NVIDIA’s current listing for availability and pricing.

A complete RTX 5090 tower can offer much higher raw GPU throughput, better gaming performance, and standard x86 Windows or Linux compatibility. It is the better choice when the model fits within 32GB of VRAM and maximum throughput matters. It will generally be larger, hotter, and more power-hungry.

DGX Spark is the better fit when a model does not fit comfortably in 32GB of VRAM, when compact size and power efficiency matter, or when 128GB of coherent memory is more valuable than peak performance on smaller models. The comparison is not simply “which GPU is faster”; it is also “can the desired model fit without complicated offloading?”

Cloud GPUs

Cloud compute remains preferable for intermittent use, large-scale training, models that exceed local capacity, multi-GPU jobs, and workloads requiring high-bandwidth HBM. It also avoids hardware maintenance and the upfront cost of a $4,699 workstation.

DGX Spark is most defensible when the owner will use it frequently for local iteration, private data, repeated experiments, or a shared development and inference node. It is not automatically cheaper than the cloud if it sits idle most of the time.

Who should buy DGX Spark?

DGX Spark is a strong fit if you:

• Need to run large models locally and privately.
• Value 128GB of coherent memory more than maximum tokens per second.
• Use CUDA, PyTorch, TensorRT-LLM, NIM, NeMo, RAPIDS, or NVIDIA containers.
• Want to prototype locally before deploying to NVIDIA cloud or a data center.
• Develop robotics, computer vision, agents, image generation, or video-generation systems.
• Need a compact, portable, relatively efficient AI development or inference node.
• Are comfortable with Linux and the software implications of Arm64.

It is a poor fit if you:

• Mainly want to play games.
• Expect a normal Windows mini-PC experience.
• Primarily perform office work, browsing, or conventional video editing.
• Care more about maximum tokens per second than model capacity.
• Want to train frontier models from scratch.
• Expect all 128GB to behave like high-bandwidth dedicated VRAM.
• Need multiple discrete GPUs, extensive PCIe expansion, or user-upgradable memory.
• Use AI only occasionally and can tolerate cloud latency.
• Already own a 128GB AMD or Apple system that supports your required software.

A practical buying checklist

1. Identify the actual workload. Write down the models, quantization, context length, concurrency, and whether you need inference, fine-tuning, or training.
2. Check memory, not just parameter count. Include weights, KV cache, activations, optimizer state, operating-system overhead, and temporary buffers.
3. Verify Arm64 support. Confirm that your Python packages, compilers, CUDA extensions, containers, and development tools work on NVIDIA DGX OS and ARM64.
4. Compare the complete system price. The $4,699 figure is the U.S. Founders Edition MSRP after February 2026. Check OEM SSD capacity, warranty, support, tax, shipping, and regional availability.
5. Choose storage deliberately. A 1TB OEM configuration can fill quickly with models, datasets, checkpoints, and containers. The Founders Edition provides 4TB, while partners may offer 1TB, 2TB, or 4TB.
6. Plan the network. ConnectX-7 and up to 200Gb/s networking are useful for linking systems or accessing fast storage, but the required switches, cables, and adapters can add cost.
7. Use the supplied power adapter. NVIDIA warns that a lower-rated or different adapter can cause reduced performance, boot failure, or shutdowns.
8. Benchmark your own model. Independent reviews disagree on which platform is fastest because results change with model architecture, quantization, context, runtime, and concurrency.
9. Check enterprise licensing. Preinstalled CUDA and development tools are not the same as an automatically included, unlimited NVIDIA AI Enterprise support entitlement.

The verdict

The original headline is historically accurate but currently misleading. NVIDIA did announce a roughly $3,000 desktop AI computer in January 2025, but that product was then called Project DIGITS. The shipping product is DGX Spark, and NVIDIA’s Founders Edition now costs $4,699 in the U.S.

Its unusual value is the combination of 128GB of coherent unified memory, NVIDIA’s CUDA software ecosystem, a compact 1.2kg chassis, high-speed networking, and a local-to-cloud development workflow. Its weaknesses are equally important: 273GB/s of memory bandwidth, an Arm64/Linux environment, limited desktop expansion, and performance that can lag Apple or discrete GPUs on some workloads.

Frequently Asked Questions

Is NVIDIA DGX Spark the same product as Project DIGITS?

Yes. Project DIGITS was the January 2025 announcement name. NVIDIA renamed it DGX Spark on March 18, 2025. The $3,000 price and May 2025 availability date refer to the original announcement, not the current Founders Edition.

How much does DGX Spark cost now?

NVIDIA’s U.S. Founders Edition MSRP is $4,699 after a February 2026 increase from $3,999. Partner systems may have different prices depending on storage, warranty, cooling, retailer, region, and stock.

Can DGX Spark run a 200-billion-parameter model?

NVIDIA claims up to 200B-parameter inference on one system, but this is a capacity claim rather than a speed guarantee. Quantization, context length, KV-cache size, model architecture, software overhead, and available memory determine whether a specific model fits and performs usefully.

Does DGX Spark have 128GB of VRAM?

No. It has 128GB of coherent unified LPDDR5X memory shared by the CPU and GPU. This can provide more model capacity than a 32GB discrete GPU, but its 273GB/s bandwidth is much lower than that of many high-end discrete or data-center GPUs.

Can DGX Spark train a 70B AI model?

NVIDIA describes support for local fine-tuning of models up to 70B parameters. Fine-tuning an existing model is not the same as training a 70B foundation model from scratch, which requires substantially more compute and memory.

Is DGX Spark good for gaming or Windows?

It is not designed primarily as a gaming PC or conventional Windows desktop. It runs DGX OS on an Arm64 CPU and is intended for Linux, CUDA, containers, and AI development. Buyers seeking Windows, gaming, or broad x86 compatibility should consider an RTX workstation or an AMD Ryzen AI system instead.

The Bottom Line

Bottom line: DGX Spark is a compact local AI development workstation—not a current $3,000 consumer desktop and not a universal replacement for a powerful discrete-GPU tower or cloud cluster. Buy it for CUDA compatibility, large local models, portability, and local-to-cloud development. Choose AMD for lower-cost Windows/x86 flexibility, Apple for high memory bandwidth and general desktop performance, an RTX 5090 tower for maximum throughput and gaming, or cloud compute for occasional use and large-scale training.

Quick Recap

Bestseller No. 1

Dell NVIDIA Tesla V100 GPU SXM2 32GB NWWWX by DELL

GPU Chipset: NVIDIA; Memory: HBM2; Programming Interface: CUDA; Memory Capacity: 32GB; Slot Compatibility: SXM2

$854.96

Bestseller No. 2

Gigabyte NVIDIA GeForce RTX 3060 Gaming OC V2 Graphics Card - 12GB GDDR6, 192-bit, PCI-E 4.0, 1837MHz Core Clock, RGB, 2X DP 1.4, 2X HDMI 2.1, NVIDIA Ampere - GV-N3060GAMING OC-8GD

Core Clock: 1837MHz; WINDFORCE 3X Cooler

$695.00

Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.