NVIDIA DGX Spark Review (2026): A Desktop Built Only for AI

This NVIDIA DGX Spark review looks at the only machine in our lineup designed for nothing but AI. Specifically, the DGX Spark is a small desktop built on the GB10 Grace Blackwell superchip. Also, it pairs 128GB of unified memory with a full Blackwell graphics unit. So it delivers up to one petaFLOP of AI performance, and runs models up to roughly 200B parameters locally.

Its real selling point is not raw speed. Instead, it is compatibility. The DGX Spark speaks native CUDA, the same software stack as NVIDIA’s server hardware. So code you prototype on your desk moves to a data-centre GPU unchanged. In short, that is a genuinely different proposition from the other 128GB machines.

One thing to weigh is its memory bandwidth, which we cover in full below. Indeed, it shapes what this machine does best.

The DGX Spark Founders Edition launched at $3,999. After a memory-driven rise, the NVIDIA DGX Spark price now sits closer to $4,699. So it is the most expensive machine in this lineup by some margin. Ultimately, it is a specialist tool at a specialist price.

You are not limited to NVIDIA’s own box, though. For instance, OEM versions from Asus, Dell, HP and Lenovo sell the same GB10 platform. In practice, they typically fall in the four to five thousand dollar range. So it is worth comparing configurations before you commit.

Strip away the specifications, and this is what you are actually buying. The DGX Spark runs DGX OS and the complete CUDA stack, identical to an NVIDIA data-centre GPU. Essentially, for a developer or research team, that parity is the whole game. So you can build, debug and fine-tune on a quiet box under your desk. Then you deploy to a server cluster without rewriting a line.

That is the practical difference between this and the unified-memory rivals. Notably, the GMKtec EVO-X2 offers the same 128GB for half the money. But it runs AMD’s ROCm, which is improving yet still needs coaxing. So if your world is CUDA, the DGX Spark removes every compatibility question. Naturally, for many buyers, that alone justifies the premium.

Capacity is enormous. Indeed, the 128GB of unified memory holds models up to around 200B parameters on a single unit, far beyond a 32GB RTX 5090. Even so, what you should calibrate is speed, which splits into two very different halves:

• Prompt processing is fast. On a 20B model it reads a prompt at over 2,000 tokens per second, so it digests long context and documents quickly.
• Token generation is modest. That same 20B model generates at around 50 tokens per second. By comparison, a 70B model manages roughly 2 to 3, slow for real-time chat.
• Batching helps a lot. Serve several requests at once and total throughput multiplies. So as a shared prototype server, it makes more sense than the single-user numbers suggest.

So the takeaway is clear. It loads almost anything you throw at it and reads prompts fast. But it is not the machine for the quickest possible single-user chat. It is a development and prototyping box first.

Here is the number to weigh. For all its compute and capacity, the DGX Spark moves memory at 273 GB/s. Notably, that is modest, a fraction of an RTX 5090’s 1,792 GB/s. Because token generation is limited by memory bandwidth, that is why big models generate slowly, even though they fit with room to spare.

So on a shared 20B model, an RTX 5090 generates tokens roughly four times faster. The DGX Spark answers in two ways. First, it holds models the 5090 cannot fit at all. Second, it processes prompts quickly. Still, if your goal is the fastest answers from a model that already fits a 5090, this is not the machine. Buy it for capacity, CUDA and prompt speed, not for token-generation records.

Key NVIDIA DGX Spark specifications:

One feature sets the DGX Spark apart from every other desktop here. Specifically, its built-in ConnectX-7 networking is the same technology NVIDIA uses in data centres. So you can link two units together over a fast, low-latency connection. Then you can run models up to around 405B parameters.

Clearly, that is a serious capability in a machine this size. It also points to who the DGX Spark is really for: teams that prototype large models locally before scaling to a cluster. Indeed, no other box in this lineup offers anything like it. So it is a large part of why the price sits where it does.

Physically, the DGX Spark is tiny, about 150mm square and 50mm tall. Visually, it is a gold-flecked block that looks nothing like a workstation. Meanwhile, it draws around 240W at peak, far less than a discrete-GPU tower. And it ships with a 4TB drive and DGX OS ready to go.

Being an early product, though, it has had some rough edges. Reviewers on pre-production units saw driver instability and networking that did not quite hit its rated throughput. Even so, those should settle with updates. Still, if you buy early, expect to update often. It is a first-generation machine, and it feels like one.

Both carry 128GB of unified memory, but they take very different paths to get there. Here is how the NVIDIA DGX Spark compares to the GMKtec EVO-X2 side by side.

What we liked

• 128GB of unified memory runs models up to about 200B parameters locally
• Native CUDA, identical to NVIDIA data-centre hardware, so code deploys unchanged
• Very fast prompt processing, over 2,000 tokens per second on a 20B model
• Throughput scales strongly under batching, ideal as a shared prototype server
• ConnectX-7 networking links two units to run 405B-class models
• Tiny 150mm desktop form factor drawing only about 240W
• Ships with a 4TB drive and the complete DGX OS software stack

What could be better

• Memory bandwidth of 273 GB/s caps token generation, so big models answer slowly
• A 70B model generates at only about 2 to 3 tokens per second, poor for real-time chat
• Expensive, around $4,000 or more, roughly double a 128GB GMKtec EVO-X2
• A post-launch price rise pushed it from $3,999 toward $4,699
• Early units showed driver instability and below-rated networking that need updates

Buy it if you are a researcher, AI engineer or team that develops on the CUDA stack. Ideally, you want to prototype large models locally before deploying to NVIDIA servers. In that case, its capacity, CUDA parity and fast prompt processing are exactly what you need. And nothing else here matches its clustering.

Skip it if you mainly want the largest model for the least money. Instead, the GMKtec EVO-X2 gives you the same 128GB for half the price. Skip it too if you want the fastest single-user answers. Instead, an RTX machine like the HP OMEN 45L generates far quicker on models that fit.

Overall, the NVIDIA DGX Spark is a specialist tool. Judged as one, it is remarkable: a data-centre CUDA environment, 128GB of memory and 200B-model capacity in a box the size of a paperback. For the developer who needs that parity, nothing else comes close. Still, its honest weakness is bandwidth, which makes token generation slow. So it is a prototyping and prefill machine, not a throughput champion. And it carries a first-generation price and a few first-generation rough edges.

In the end, weighed as the purpose-built AI machine in our best desktop for AI lineup, it is the right answer for a narrow, serious audience. So we rate the NVIDIA DGX Spark 4.2 out of 5. It is a unique CUDA development desktop, held back for general use by modest bandwidth and a specialist price.