NVIDIA’s DGX Spark 64GB goes on sale today, Friday, October 23, 2026, at prices starting from $4,999 through six manufacturer partners: Acer, ASUS, Dell, Gigabyte, HP and MSI. StorageReview’s launch report confirmed the date, the price and the feature NVIDIA is pushing hardest, two units that pool into 128GB of memory. The catch sits one tab over on NVIDIA’s own marketplace, where the 128GB DGX Spark lists at $6,950.
That $1,951 gap turns a simple cheaper-model story into a pricing puzzle. NVIDIA first announced the smaller box on October 2 (the 64GB DGX Spark announcement coverage has the original details) and updated the post on October 21 to lock in today’s availability. This piece looks at launch day itself: what you can actually buy, what the numbers say about value, and how the new tier changes the local-AI hardware market heading into 2027.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
DGX Spark 64GB Goes on Sale: What Shipped on October 23
The DGX Spark 64GB is a partner-built desktop AI computer. It uses the same GB10 Grace Blackwell Superchip, DGX OS and NVIDIA AI software stack as the 128GB model, according to NVIDIA’s announcement. The only headline spec that changes is unified memory, which drops from 128GB to 64GB.
NVIDIA says one unit runs models of up to 100 billion parameters on the desk. Two units connected through the new NVIDIA Sync Cluster Assistant pool 128GB and, per NVIDIA, handle models of up to 200 billion parameters. That scaling story is the reason the company built a second tier instead of simply cutting the price of the first.
Availability is partner-only. NVIDIA reportedly will not offer a Founders Edition of the 64GB box, according to coverage of the launch, so buyers who wanted to order direct from NVIDIA still land on the 128GB listing. Lenovo is not among the six names on the 64GB announcement, even though NVIDIA separately says DGX Spark systems are available from Lenovo.
The $4,999 Starting Price Versus the $6,950 128GB Listing
NVIDIA states that prices for the 64GB model start at $4,999. The phrase starting price matters, because the company does not publish separate prices for Acer, ASUS, Dell, Gigabyte, HP or MSI. Each partner sets its own number, and storage, support terms and bundled extras will push some configurations above the floor.
The 128GB figure is firmer. NVIDIA’s marketplace lists the DGX Spark with 128GB of coherent unified memory, 4TB of self-encrypting NVMe storage and ConnectX-7 networking at $6,950, on a page updated October 22. Hardware Busters framed the launch as the 64GB model arriving while the 128GB model jumps to $6,950, and TweakTown pointed out that the new half-memory box costs more than the original 128GB model did at its own launch.
The sources reviewed for this article do not establish the exact date the 128GB price moved, so treat the price-hike framing as the press’s reading rather than a documented NVIDIA announcement. What the sources do establish is the present state: $4,999 and up for 64GB, $6,950 for 128GB, and both sold under the DGX Spark name.
| Spec | DGX Spark 64GB | DGX Spark 128GB |
|---|---|---|
| Unified memory | 64GB | 128GB |
| Starting price | $4,999 (partner prices vary) | $6,950 (NVIDIA Marketplace) |
| Chip | GB10 Grace Blackwell Superchip | GB10 Grace Blackwell Superchip |
| Software | DGX OS, NVIDIA AI stack | DGX OS, NVIDIA AI stack |
| Sales channel | Acer, ASUS, Dell, Gigabyte, HP, MSI | NVIDIA Marketplace and partners |
| Cost per GB of memory | about $78.11 | about $54.30 |
| Two-unit cluster cost | 128GB pooled, about $9,998 at the floor price | 256GB pooled, about $13,900 at list price |
The cost-per-gigabyte rows are simple division on the published starting prices: $4,999 over 64GB and $6,950 over 128GB. They are arithmetic, not NVIDIA claims, and they carry the same caveat as the price floor itself.
What the 64GB Configuration Keeps and Gives Up
Keep first. The silicon is identical. Compute throughput, the Blackwell GPU side of GB10 and the ConnectX-7 networking that makes clustering possible all carry over, and NVIDIA says its October software update supports both the 64GB and 128GB systems, so one DGX OS image serves either box. A developer who prototypes on a 64GB unit can move the same workflow to a 128GB machine later without rewriting anything.
Now the loss. Memory capacity decides which models load at all, and unified memory on this class of machine is shared between the CPU side, the GPU side and the operating system. Sixty-four gigabytes sounds generous until a large model, its key-value cache and a few developer tools compete for the same pool.
Storage is the unconfirmed detail
The 128GB listing specifies 4TB of self-encrypting NVMe storage. Partner pages for the 64GB models may ship with different drive sizes, and the launch material reviewed here does not confirm that every partner configuration matches the 4TB figure. Check the exact storage line on the partner’s order page before comparing it to the $6,950 box.
Which Partners Sell It and What We Know
NVIDIA’s announcement names Acer, ASUS, Dell, Gigabyte, HP and MSI. The information that exists for each is uneven, and the table below sticks to what the launch materials confirm.
| Partner | 64GB product | Price status |
|---|---|---|
| Acer | Partner version of the DGX Spark 64GB | Follows NVIDIA’s $4,999 starting price |
| ASUS | Ascent GX10 64GB, announced October 2 (LPDDR5x unified memory) | US price not confirmed in ASUS materials |
| Dell | Dell Pro Precision with GB10, 64GB configuration, available October 23 | No official price in Dell’s release |
| Gigabyte | Named by NVIDIA as a launch partner | Product name and price not published in reviewed sources |
| HP | Named by NVIDIA as a launch partner | Product name and price not published in reviewed sources |
| MSI | Named by NVIDIA as a launch partner | Product name and price not published in reviewed sources |
ASUS sells its version as the Ascent GX10 64GB, an add-on to the existing 128GB Ascent GX10. ASUS positions the 64GB model for roughly 30 to 35 billion parameter models, a far more conservative number than NVIDIA’s 100 billion headline. Dell’s release lists its 64GB deskside system under the Dell Pro Precision name and offers no price.
That missing price column is the real launch-day story for shoppers. A starting price tells you the floor, and partners with different support tiers rarely land on it.
Clustering: Sync Cluster Assistant, 128GB Pools and the 1.7x Claim
NVIDIA Sync Cluster Assistant is the software piece that makes two small boxes behave like one larger one. According to NVIDIA, two clustered 64GB systems pooled 128GB of memory and delivered up to 1.7 times the performance of a single system in an NVIDIA test running Qwen 3.8 27B. That is a vendor-run result on a single model, not an independent benchmark, and it should be read as the best case.
The scaling ceiling is higher on the 128GB hardware. An October 22 NVIDIA update said up to eight 128GB units can be manually clustered for as much as 1TB of unified memory. At the $6,950 list price, eight units come to $55,600 before networking gear, which puts a 1TB Spark cluster in the same budget range as a modest server.
Clustering also reaches beyond NVIDIA’s own boxes. Software that spreads inference across machines you already own, such as NVIDIA PAIR for linking idle PCs, attacks the same memory-wall problem from the other direction, using hardware you have instead of buying a second Spark.
The price of two small boxes
Here the arithmetic gets awkward. Two 64GB units at the floor price cost about $9,998. One 128GB unit costs $6,950. For $3,048 more, the pair delivers the same pooled memory plus, in NVIDIA’s best-case test, higher throughput. The single 128GB machine wins on price, simplicity and no cluster overhead. The pair wins only if you buy one now and the second later, spreading the spend.
What Actually Fits in 64GB of Unified Memory
Model weights are the first memory bill. At 4-bit precision, a 100-billion-parameter model needs roughly 50GB for weights alone, which leaves around 14GB on a 64GB machine for the context cache, the operating system and any other process. NVIDIA’s 100 billion figure is therefore reachable only with aggressive quantization and short contexts.
ASUS’s more cautious 30 to 35 billion parameter guidance sounds closer to everyday use. At 8-bit precision a 30-billion-parameter model takes about 30GB, which leaves space for long prompts and agent tool outputs. NVIDIA also reportedly cited local runs of Nemotron 3.5 Bolt 74B, Qwen3.8 27B and Gemma 4 26B-A4B on the 64GB systems, according to coverage of its materials. That list spans the realistic range: one mid-size model that fits easily and one larger model that stretches the box.
For coding agents, retrieval pipelines and fine-tuning experiments, 64GB covers the middle of the open-weight market. It does not cover the largest open models, which is exactly where the 128GB machine, or the eight-unit cluster, exists to take over.
DGX Spark, RTX Spark and DGX Station: NVIDIA’s Local AI Family
NVIDIA now sells a three-part local-AI lineup under the Spark banner, and the DGX Spark 64GB only makes sense inside it. DGX Spark targets developers on Linux. RTX Spark targets personal agents on Windows PCs. DGX Station targets larger workloads and enterprise Linux use.
| Product | Primary OS and audience | Memory and compute | Price status |
|---|---|---|---|
| DGX Spark 64GB | Linux, developers | 64GB unified, GB10 | From $4,999 |
| DGX Spark 128GB | Linux, developers | 128GB unified, GB10 | $6,950 on NVIDIA Marketplace |
| RTX Spark systems | Windows, personal agents | Up to 128GB unified, up to 1 petaflop AI | ASUS ProArt GR1X MSRP from $2,799.99, due later in 2026 |
| Dell Pro Precision with GB300 | Windows deskside AI | Up to 748GB coherent, 20 petaflops FP4 | Coming soon, no price |
RTX Spark launched on October 16, and the RTX Spark launch analysis against Apple Silicon covers the Windows side. The Linux workstation tier is moving too. Dell announced a Windows-based GB300 deskside system with up to 748GB of coherent memory and 20 petaflops of FP4 compute, and NVIDIA previewed a Windows version of DGX Station with 748GB of memory. The DGX Station for Windows is still a preview, not a shipping product.
Microsoft’s own take on the category shows how crowded the middle is. The Surface Dev Box built on RTX Spark offers a Windows route to a similar memory ceiling, and that pushes the 64GB DGX Spark to justify itself as the Linux-first, CUDA-first entry point.
Market Impact: Why NVIDIA Opened a Cheaper Tier Now
The timing tracks the memory market. ASUS’s own site refers to the 2026 memory shortage when it tells gamers that an 8GB graphics card can still deliver at 1080p. A box that carries 128GB of unified memory is exposed to the same shortage, and halving the memory is the cheapest lever NVIDIA and its partners have to hold a sub-$5,000 sticker.
Futurum Group’s analysis of the launch argues that the lower price point aims to accelerate local AI inference, and the logic holds. A $4,999 floor puts the machine within reach of a team’s equipment budget rather than a capital request, and it widens the funnel for developers who would later buy 128GB boxes or DGX Station hardware.
Pricing pressure shows up at the data-center end too, as the report on NVIDIA B200 pricing explains. For teams that rent cloud GPUs, ASUS argues on its site that a $3,000 to $4,000 agent computer can pay for itself in under two years against cloud AI bills. That is a vendor’s marketing claim, but it explains why the $4,999 line matters: local hardware competes with a recurring cloud invoice. The same pressure appears in coverage of how RTX Spark undercuts large AI workstations.
Competitive Comparison: Mac Studio, Framework Desktop and Strix Halo Boxes
NVIDIA’s rivals in this segment sell on memory per dollar and on ecosystem comfort. Apple’s Mac Studio line scales unified memory far above anything a single Spark offers, and AMD’s Strix Halo mini-PCs undercut NVIDIA on sticker price. NVIDIA answers with CUDA compatibility, ConnectX-7 networking and a software stack aimed at developers who deploy to NVIDIA data-center GPUs.
Our head-to-head pieces cover those trade-offs in more depth. The DGX Spark vs Framework Desktop vs Mac Studio comparison and the earlier DGX Spark versus Mac Studio breakdown both predate the 64GB tier, so their price tables need one new row.
For budget buyers, the GMKtec, Beelink and Minisforum mini-PC comparison shows what Strix Halo machines offer instead. The 64GB model is NVIDIA’s first real attempt to meet those rivals at the low end of its own lineup. The cheaper box still carries CUDA and ConnectX-7, but it also carries the smaller memory pool, and memory is exactly where Apple and AMD compete hardest.
Historical Context: From One 128GB Box to a Two-Tier Lineup
DGX Spark began as a single product, a 128GB GB10 desktop. Over the past three weeks NVIDIA has turned it into a family. The 64GB tier arrived with the October 2 announcement, the clustering guidance grew to as many as eight units by October 22, and the Spark name now stretches across Windows PCs and workstations.
The roadmap behind it is also longer. On August 31, NVIDIA and MediaTek extended their partnership to cover multiple future generations of RTX Spark and DGX Spark chips, from consumer PCs to AI developer supercomputers and enterprise workstations. No successor to the current GB10 box has been announced, and none of the sources reviewed carry a release date for one.
Discontinuation worries have circulated in developer forums. NVIDIA’s materials reviewed for this article do not show any retirement or renaming of the 128GB DGX Spark, and the 64GB tier is presented as an addition, not a replacement.
Who Should Buy the 64GB Model, and Who Should Skip It
Buy the 64GB version if your daily models sit between 7 billion and roughly 35 billion parameters, if you want a CUDA development box that mirrors NVIDIA’s data-center stack, or if you plan to add a second unit later. The $4,999 floor also suits teams buying several seats where every unit does not need to run the largest model.
Skip it if your plan involves 70-billion-class models with long contexts on one machine. At $1,951 more, the 128GB box gives you twice the memory for roughly 39 percent more money, and the per-gigabyte math favors it by about $24. Skip it too if you work mainly on Windows, since the RTX Spark systems target that audience. Readers tracking the wider category can follow the latest hardware coverage.
Five Predictions for DGX Spark Through Early 2027
- Partner prices will scatter above $4,999. With six vendors setting their own numbers, expect a spread of configurations, and expect the cheapest ones to sell out first.
- The 128GB listing stays expensive while memory is tight. Nothing in NVIDIA’s materials hints at a cut, and the shortage that justified the 64GB tier has not eased.
- NVIDIA ships more playbooks for the 64GB box. The company said more playbooks are coming to 64GB devices, and clustering guides are the likeliest next release.
- The 8-unit cluster becomes the showcase demo. NVIDIA will keep promoting the 1TB pool because it makes a $6,950 box look like a building block.
- Windows and Linux tiers converge in marketing. RTX Spark and DGX Station for Windows will push NVIDIA to explain which Spark a given developer should buy.
These are analysis, not reported plans. None of them rests on a confirmed NVIDIA roadmap date.
Frequently Asked Questions
When does the DGX Spark 64GB go on sale?
Today, Friday, October 23, 2026. NVIDIA announced it on October 2 and confirmed the date in an update on October 21.
How much does the DGX Spark 64GB cost?
Prices start at $4,999 through Acer, ASUS, Dell, Gigabyte, HP and MSI. Each partner sets its own final price, and NVIDIA publishes only the floor.
What does the 128GB DGX Spark cost now?
NVIDIA’s marketplace lists it at $6,950, on a page updated October 22. That is a separate configuration from the 64GB launch model.
Can two DGX Spark 64GB units work together?
Yes. NVIDIA Sync Cluster Assistant connects two systems into a 128GB pool. NVIDIA reports up to 1.7 times the performance of a single system in its own Qwen 3.8 27B test.
Is the DGX Spark discontinued?
No official source reviewed shows that. NVIDIA still lists the 128GB DGX Spark and added the 64GB tier as a new configuration.
Does the 64GB model run different software than the 128GB?
No. Both use the GB10 chip, DGX OS and NVIDIA’s AI software stack, and the October software update supports both memory configurations.
Which models run on a single 64GB unit?
NVIDIA says models of up to 100 billion parameters, while ASUS positions its 64GB Ascent GX10 for roughly 30 to 35 billion. Quantization level and context length decide where a given model lands between those figures.
How is DGX Spark different from RTX Spark?
NVIDIA positions DGX Spark as a Linux system for developers, and RTX Spark as the chip for personal agents on Windows PCs.
