The NVIDIA GB10 ecosystem is getting a shock as NVIDIA makes two changes to the lineup. First, the NVIDIA DGX Spark 128GB family is increasing in price. It looks like that model is moving up to a $6950-ish tier, with various OEMs charging different amounts. Likely realizing this is pricing many out of the market, NVIDIA has a solution with a 64GB model at $4,999.
NVIDIA DGX Spark 64GB Launched and Big 128GB GB10 Price Increases
First off, NVIDIA says that the new DGX Spark 64GB is available from Acer, ASUS, Dell, Gigabyte, HP, and MSI on Friday, Oct. 23, starting at $4,999. We will be interested to see if the partner $4999 pricing is for 1TB or 4TB when that day comes. While NVIDIA is quick to point out that the 64GB version retains the NVIDIA ConnectX-7 200GbE networking that we have seen in the 128GB model, the price is coming as a shock to many, as the NVIDIA DGX Spark 128GB/ 4TB launched at $3999 about a year ago. The new price is $6950 for that model.
NVIDIA says you can easily cluster two new 64GB models for 128GB at around $8000, and it is enabling this easily in its software.
We built a big 8x NVIDIA GB10 128GB cluster earlier this year when we were buying nodes for around $3000 each.
It is crazy to think that it is now a $50K+ option. Still, when we did the Using the Dell Pro Max with GB10 to Profit within 12 Months piece, $6000 or so for a 256GB cluster was easy to drive value from, and that was before OpenClaw really set off the huge advancements with agentic AI. To be fair, those nodes got added to the 8x node cluster, and we run them in that pool, but we underestimated the payback period with those systems because it was just on the cusp of agentic AI.
The two 128GB node cluster is great, but we generally tell folks to get a single 128GB node instead of two 64GB nodes, given our experience running these in clusters following our scale up then out advice.
Final Words
Ok, let us get to what you are probably thinking at this point. $4999 for a 64GB machine feels quite high. $6950 for a 128GB machine feels lofty, too. Our team has been chatting about this, and $6950 puts it in the range of some of the AMD Gorgon Halo systems coming out in October. So the question then becomes: do you want similar memory bandwidth but AMD with 50% more memory capacity at 192GB, or NVIDIA with the better networking but the lower 128GB capacity point? Up to 192GB of total capacity, we would prefer staying on a single system instead of scaling out. You want to scale up as much as you can, then scale out.
Likewise, Apple will begin shipping the $10,000 256GB Mac Studio M5 Ultra soon, and that will be a very strong competitor if two NVIDIA GB10s cost more. Our lab’s M3 Ultra 512GB unit is an absolute workhorse.
Memory prices are still increasing, largely due to the models getting better, and so the value of the GB10 in terms of work that it can do is going up, but the hardware cost is rough for everyone. $6950 is almost three years of a $200/mo OpenAI or Claude subscription, but with the cloud subscriptions you are sending your data to a third party. With the latest OpenAI and Claude releases, the gap between frontier and open-source smaller models you can run locally is wider, so that is another valid point to make.
I was chatting with Patrick, our EiC and Chief Analyst, about this before doing the story. His response was, “Yikes, but we are paying $2000 for 64GB ECC RDIMMs, so pricing logically must increase.” Tellingly also, “We were far off in how fast we have been able to generate value from the GB10’s as we have been getting much more out of them than we projected.”
In October, you should see three NVIDIA DGX Station reviews on STH. The GB300-based systems currently sell for around $100K, but at some point, as the GB10’s increase in price, that gap between a GB10 cluster and the $100K bigger box gets closer. If you saw our GB300 station piece in Substack, we have those GB300 stations running in a cluster using 800Gbps networking, which was fascinating and gave us an appreciation for the scale-up versus scaling-out debate.
