Nvidia's Cheaper DGX Spark Goes On Sale Today: a $4,999 Box That Runs Big AI Models at Home

So basically: Nvidia’s DGX Spark is a small desktop box that it calls a personal AI supercomputer. Today, Oct. 23, a new cheaper version with 64GB of memory goes on sale starting at $4,999. It’s sold through Acer, ASUS, Dell, Gigabyte, HP and MSI, not by Nvidia directly.

What you get

It’s the same guts as the original 128GB model, just with half the memory:

• Nvidia’s GB10 Grace Blackwell superchip
• Nvidia’s own DGX OS with its AI software already set up
• A built-in high-speed ConnectX-7 network port

Nvidia says one 64GB unit can run AI models up to about 100 billion parameters right on the box. That’s big enough for a capable chatbot or coding helper that never sends your data to the cloud.

Two is better than one

This is the fun part. You can connect two of them with a single cable, and they pool their memory into 128GB. That bumps you up to models around 200 billion parameters. Nvidia says a new app called Sync Cluster Assistant sets the whole thing up for you.

In Nvidia’s own test with a model called Qwen 3.8 27B, two clustered boxes ran up to 1.7 times as fast as one. Nvidia also says a Model Launcher is coming at the end of the month that’ll let you start a model with a few clicks.

Why a smaller one now

Memory has gotten really expensive lately because AI data centers are buying up everything. Tom’s Hardware called this version a “lifeline” for local AI fans during the “RAMpocalypse.” Cutting the memory in half is how Nvidia keeps the starting price under five grand.

Same speed, less room

Tom’s Hardware says the 64GB version keeps the same 20-core Arm processor and the same 273 GB/s of memory bandwidth as the original. Memory bandwidth is how fast the chip can feed data to the AI, so on smaller models it should feel about the same. You just can’t fit models as big.

Why’s that OK now? AI models have gotten smaller and smarter. Tom’s Hardware points out that a strong model like Qwen 3.8 27B can fit in 32GB of memory now, with limited context. A year ago, the thinking was you needed 128GB to run the good stuff.

Is it actually cheaper?

It depends on what you compare it to. Tom’s Hardware says that if you can even find a 128GB GB10 system in stock right now, it runs about $7,000 to $9,000. Next to that, $4,999 looks good. Some readers weren’t buying it, though, and pointed out the 128GB model sold for around $4,000 not long ago. With memory prices moving this fast, Tom’s warned that the $4,999 price might not last long.

The 128GB version is sticking around for people who need bigger models or want to fine-tune their own.

Who it’s for

This isn’t a gaming PC. It’s for developers, researchers and hobbyists who want to:

• Run an AI agent around the clock on their own hardware
• Keep private data off the cloud
• Let the box do the heavy AI work while their regular laptop stays free

My take

Five grand is still a lot of money, so this isn’t for most people. But it’s wild that a box this small can run AI models that needed a server room a few years ago. The “buy one now, add a second later” idea is smart too. I think we’re going to see a lot more people running their own AI at home instead of paying monthly forever.