Direct answer: NVIDIA says the 64 GB DGX Spark will be sold only through partners (Acer, ASUS, Dell, Gigabyte, HP and MSI) from October 23, 2026, starting at $4,999. The 128 GB model is the one NVIDIA Marketplace listed at $6,950 and out of stock when we last checked it on October 6. Both use the same GB10 Grace Blackwell chip and DGX OS; the practical difference is memory, and with it the size of model each can hold. NVIDIA’s own sizing is up to 100B parameters for 64 GB and up to 200B for 128 GB. Choose 64 GB if your models are small and cost matters most; choose 128 GB if you want headroom and cannot easily cluster later.
This is a dated comparison, not a purchase quote. The 64 GB units had not shipped on the day we wrote this, so every 64 GB price below is an announced starting price, not an observed shelf price.
What changed on October 2
On October 2, 2026, NVIDIA announced that DGX Spark would be available with 64 GB of unified memory from partner manufacturers. Its product page now shows the 64 GB model as “Coming Soon” and says the configuration is available exclusively through participating OEM partners. Until now, the platform’s headline configuration was the 128 GB model. The change is therefore less about a new chip and more about a cheaper entry point into the same platform.
Side-by-side: what is the same and what differs
| Item | DGX Spark 64 GB | DGX Spark 128 GB |
|---|---|---|
| Chip | GB10 Grace Blackwell Superchip | GB10 Grace Blackwell Superchip |
| Unified memory | 64 GB LPDDR5X | 128 GB LPDDR5X |
| Memory bandwidth | 273 GB/s (NVIDIA’s listed figure) | 273 GB/s |
| Peak AI compute | Up to 1 PFLOP FP4 (NVIDIA’s theoretical figure) | Up to 1 PFLOP FP4 |
| NVIDIA-stated single-unit model size | Up to 100B parameters | Up to 200B parameters |
| Operating system and stack | DGX OS and NVIDIA AI software | DGX OS and NVIDIA AI software |
| Where to buy | OEM partners only | NVIDIA Marketplace and partners |
| Price evidence | Announced from $4,999 (October 23) | $6,950 observed on October 6, out of stock |
NVIDIA’s blog states that the 64 GB model keeps the same chip, operating system and software stack as the 128 GB model. NVIDIA’s product page lists one bandwidth figure for the platform, so we do not claim any bandwidth difference between the two. We also do not claim any speed difference: NVIDIA published no single-unit speed comparison between them, and memory size alone does not set tokens per second.
What the memory difference means for models
The model-size limits NVIDIA gives are about what fits, not about how fast it runs. On the product page, one 64 GB unit is rated for models up to 100B parameters, and one 128 GB unit for up to 200B. These are vendor-stated limits that depend on quantization and workload, so treat them as upper bounds. Our VRAM guide explains why context length and concurrent users eat memory on top of the weights.
In practice, think about three situations:
- Models up to a few tens of billions of parameters: 64 GB is likely to be enough for one user, with room for context.
- Models near NVIDIA’s 100B line, or long contexts, or several concurrent users: 64 GB leaves little margin. The 128 GB model gives breathing room.
- Models between 100B and 200B: only the 128 GB model, or two 64 GB units clustered, is within NVIDIA’s stated range.
The price math, with the caveats
The announced starting price of $4,999 compares with $6,950 for the 128 GB model, a gap of $1,951, or about 28% of the 128 GB price. Per gigabyte of memory, that is roughly $78 per GB for 64 GB versus $54 per GB for 128 GB. The cheaper box is not the cheaper memory.
Treat that comparison cautiously for three reasons. First, “starting at” is a floor: partner configurations may differ in storage and price. Second, the $6,950 figure is a single dated observation of a sold-out Marketplace listing that had shown $4,699 on September 11 and no price on September 29. Third, taxes, shipping and region are not included in either figure.
Clustering: two 64 GB units versus one 128 GB unit
NVIDIA says two DGX Spark units can be connected directly with a QSFP cable through the built-in ConnectX-7 networking. Its Sync Cluster Assistant configures the link, and the pooled memory of two 64 GB units is 128 GB, with model support up to 200B parameters. NVIDIA reported up to 1.7× the performance of a single system in its own Qwen test with two clustered 64 GB units; that is a vendor test on one model, not a general rule.
The cost trade-off is simple. Two units at the announced starting price come to $9,998, about $3,048 more than one 128 GB unit at $6,950, before any cable or other partner differences. Clustering mainly makes sense as a growth path: start with one 64 GB unit and add a second only if the workload outgrows it. If you already know you need 128 GB of memory today, one 128 GB unit costs less and avoids a cluster to operate.
Who should buy which
Choose 64 GB if you are learning, building agents around small or mid-size models, or want the cheapest official GB10 platform and can accept a hard memory ceiling. It is also a reasonable first unit if you may add a second later.
Choose 128 GB if you want to run larger models or long contexts, serve a few users at once, or do not want a memory-driven rebuild after the first project grows. Availability is the catch: the 128 GB Marketplace listing was out of stock when we checked.
Neither may be right if your usage is occasional. A few GPU-hours a month are usually cheaper to rent; see our DGX Spark versus cloud GPUs comparison. If you want a working private AI system rather than hardware to operate, our bare versus turnkey guide shows what a managed deployment adds and costs. For partner-built alternatives on the same chip, compare the HP and Acer systems, the Dell Pro Max with GB10 and the ASUS Ascent GX10.
What we could not verify
- Live Marketplace price and stock today: NVIDIA Marketplace did not load for us on October 7, so the $6,950 and out-of-stock status are from October 6.
- Final 64 GB partner prices and configurations: only the October 2 starting price of $4,999 is published. Storage, partner pricing and regional availability are unconfirmed until the units ship on October 23.
- Real-world performance: we have not benchmarked either configuration. The model-size and clustering numbers are NVIDIA’s.
We plan to update this page after the October 23 launch. Check the dates in the sources note below before relying on any figure.
Verdict
The 64 GB DGX Spark lowers the cost of entering the platform; it does not make the memory cheaper, and it caps you at roughly half the model size NVIDIA rates for the 128 GB model. If your models fit, it is the sensible low-cost start, with clustering as an upgrade path. If you are unsure, the 128 GB model is the safer buy when you can find it at a price you accept, and waiting for real partner prices on October 23 costs little.
Sources and verification note
The NVIDIA DGX Spark product page and NVIDIA’s October 2, 2026 announcement were read on October 7, 2026. The NVIDIA Marketplace figures ($6,950, out of stock, $4,699 on September 11, no displayed price on September 29) and Aradia’s pricing page were last checked on October 6. Prices, stock and partner configurations can change; confirm with the seller before you buy.
Affiliate disclosure: AI Compute Scout may earn a commission if you purchase through a marked Aradia referral link. Aradia sells a separate turnkey deployment, not a bare Spark, and this does not affect the analysis.