The NVIDIA GB10 ecosystem is reeling as NVIDIA makes two changes to the lineup. First, the price of the NVIDIA DGX Spark 128 GB family is increasing. It appears that this model is moving to a level of $6,950, with different manufacturers charging different amounts. Probably realizing that it costs a lot out of the market, NVIDIA offers a solution with a 64 GB model at $4,999.
NVIDIA DGX Spark 64GB launched and significant price increases for 128GB GB10
First, NVIDIA says the new DGX Spark 64GB is available from Acer, ASUS, Dell, Gigabyte, HP, and MSI on Friday, October 23, starting at $4,999. We’ll be interested to see if the partner’s $4,999 price is for 1TB or 4TB when the day comes. While NVIDIA is quick to point out that the 64GB version retains the NVIDIA ConnectX-7 200GbE network we saw in the 128GB model, the price comes as a shock to many, since the NVIDIA DGX Spark 128GB/4TB launched at $3,999 about a year ago. The new price is $6,950 for this model.
NVIDIA says you can easily bundle two new 64GB models for 128GB at around $8,000, and it easily allows for this in its software.
We built a big 8x NVIDIA GB10 128GB cluster earlier this year when we purchased nodes for around $3,000 each.
It’s crazy to think that this is now a $50,000+ option. Yet when we did the Use Dell Pro Max with GB10 to Profit Within 12 Months article, it was easy to generate around $6,000 for a 256GB cluster, and that was before OpenClaw really unleashed the huge advances in agentic AI. To be fair, these nodes were added to the 8x node cluster and we are running them in that pool, but we underestimated the payback period of these systems because they were on the cusp of agentic AI.
The two 128 GB node cluster is great, but we usually tell people to get a single 128 GB node instead of two 64 GB nodes, given our experience running them in clusters following our scaling and then outsourcing advice.
Final words
Ok, let’s get to what you’re probably thinking at this point. $4,999 for a 64GB machine seems pretty steep. $6,950 for a 128GB machine also seems steep. Our team discussed it, and $6,950 puts it in range for some of the AMD Gorgon Halo systems coming out in October. The question then becomes: do you want similar memory bandwidth but AMD with 50% more memory capacity at 192GB, or NVIDIA with better networking but lower capacity at 128GB? Up to 192 GB total capacity, we would prefer to stay on one system rather than scale. You want to grow as much as possible, and then evolve.
Likewise, Apple will soon start shipping the $10,000 256GB Mac Studio M5 Ultra, and that will be a very strong contender if two NVIDIA GB10s cost more. Our lab’s M3 Ultra 512GB unit is a real workhorse.
Memory prices continue to rise, largely due to improving models, and so the value of the GB10 in terms of the work it can do increases, but the cost of the hardware is high for everyone. $6,950 is almost three years of a $200/month OpenAI or Claude subscription, but with cloud subscriptions you send your data to a third party. With the latest versions of OpenAI and Claude, the gap between frontier models and the smaller open source models you can run locally is wider, so that’s another valid point to make.
I discussed this with Patrick, our EiC and lead analyst, before doing the story. His response was, “Yeah, but we’re paying $2,000 for 64GB ECC RDIMMs, so prices should logically go up.” » It is also telling: “We were a long way from how quickly we were able to generate value from GB10, because we got way more out of it than we expected. »
In October, you should see three NVIDIA DGX Station reviews on STH. Systems based on the GB300 currently sell for around $100,000, but at some point, as the price of the GB10 increases, that gap between a GB10 cluster and the larger $100,000 case gets closer. If you saw our GB300 station article in Substack, we have these GB300 stations running in a cluster using 800 Gbps networking, which was fascinating and allowed us to appreciate the debate between scaling and scalability.
Gn tech
