Nvidia has cut the price of entry into its desktop AI supercomputer lineup, unveiling a 64GB configuration of the DGX Spark priced at $4,999 as a memory supply squeeze reshapes the AI hardware market. The new tier, confirmed by PCMag, Tom’s Hardware, The Register, and Hardware Busters, arrives through OEM partners rather than as a standalone Nvidia-branded box, and it lands at a moment when the existing 128GB DGX Spark has climbed to roughly $6,950, up from its original $3,999 launch price in October 2025. Nvidia had already raised that same configuration once before, to $4,699 in February 2026, citing memory-supply constraints, before the price climbed further to its current level. All told, that is a jump of nearly 75% in about a year. For a product line Nvidia pitched as a way to put “a petaflop of AI compute” on a desk, the sudden existence of a cheaper, lower-memory variant tells its own story about where component costs are heading in late 2026.
The timing is not a coincidence. Memory manufacturers have been absorbing relentless demand from data center AI buildouts for the better part of two years, and that pressure has finally worked its way down to prosumer and workstation-class hardware. Nvidia’s own reporting on the memory crunch and Micron’s supply commitments running through 2027 already signaled that DRAM scarcity was going to show up in finished-product pricing. The DGX Spark’s new 64GB tier is one of the clearest consumer-facing examples yet.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Nvidia actually announced
According to coverage gathered by PCMag, Tom’s Hardware, and Hardware Busters, Nvidia announced its new DGX Spark configuration on October 2, 2026. The 64GB model carries 64GB of coherent unified LPDDR5X memory and is scheduled to become available on Friday, October 23, 2026, just three weeks after the announcement. Unlike the original DGX Spark, which Nvidia sold directly, the 64GB model will ship through six OEM partners, Acer, ASUS, Dell, Gigabyte, HP, and MSI, according to reporting cited by those outlets. That distribution shift matters: it signals Nvidia is treating the lower-memory Spark as a channel product meant to widen the funnel of buyers who can afford a local AI development box, not as a flagship.
Both the 64GB and 128GB variants are still built on Nvidia’s GB10 Grace Blackwell Superchip, which pairs a Blackwell-generation GPU with a 20-core Arm CPU complex co-designed with MediaTek. The two configurations reportedly share identical memory bandwidth of 273GB/s, which is notable on its own: normally, halving a system’s memory capacity also cuts the number of active memory channels, reducing bandwidth along with it. Hardware Busters’ analysis suggests Nvidia instead used lower-density LPDDR5X packages to hit the 64GB target while keeping the memory bus fully populated, preserving throughput. That is a deliberate engineering choice, not a side effect, and it tells you Nvidia wanted the cheaper Spark to still feel like a full-bandwidth machine rather than a bandwidth-starved budget model.
The tradeoff shows up elsewhere. The 64GB configuration reportedly ships with half the storage of the original 128GB model, though the exact capacity figure has not been independently confirmed across every outlet tracking the launch. Nvidia’s own product page remains the authoritative source for the final storage specification once the SKU goes live on October 23.
Why a cheaper DGX Spark exists at all
Coverage from PCMag, Tom’s Hardware, The Register, and Hardware Busters converges on the same explanation: Nvidia built the 64GB tier in direct response to a sharp increase in memory costs, driven in large part by AI data center demand competing for the same advanced memory supply that feeds consumer and workstation products. PCMag’s framing is particularly pointed, noting that even the “discount” 64GB Spark at $4,999 still lands $1,000 above the original 128GB model’s launch price of $3,999. In other words, buyers get less memory for more money than Nvidia charged a year ago for twice the capacity. That is not a typical product-tiering story. It is a pricing story forced by the raw materials market, a dynamic that research firms like IDC have tracked broadly across AI infrastructure spending throughout 2026.
This pattern lines up with what Tech Insider has already reported on the broader DRAM squeeze. Micron’s own account of the crunch described supply growth of roughly 20% failing to keep pace with demand, and separately, Micron’s leadership has said memory supply for 2027 is already around 75% committed. When the companies making the chips are telling investors they are sold out more than a year in advance, it should not be surprising that a device like DGX Spark, which leans on large pools of unified LPDDR5X memory, ends up more expensive to build than it was at launch.
It also echoes a trend Tech Insider covered on the storage side: Samsung’s P9 SSD line saw NAND pricing climb 243% even as the drive itself pushed sequential speeds to 4,000 MB/s, illustrating that this is not an isolated DRAM problem. Memory and storage markets are moving together, and any product that depends heavily on either is exposed to the same cost pressure DGX Spark is now passing on to buyers.
DGX Spark 64GB vs. 128GB: full spec comparison
The table below lays out what is confirmed across PCMag, Tom’s Hardware, The Register, Hardware Busters, and Nvidia’s own product listings for both DGX Spark memory configurations.
| Specification | DGX Spark 64GB | DGX Spark 128GB |
|---|---|---|
| Price | $4,999 | ~$6,950 (up from $3,999 at original launch) |
| Availability | October 23, 2026 | Already shipping |
| Processor | GB10 Grace Blackwell Superchip | GB10 Grace Blackwell Superchip |
| CPU complex | 20-core Arm (co-designed with MediaTek) | 20-core Arm (co-designed with MediaTek) |
| Unified memory | 64GB LPDDR5X | 128GB LPDDR5X |
| Memory bandwidth | 273GB/s | 273GB/s |
| Networking | ConnectX-7, 200Gbps | ConnectX-7, 200Gbps |
| AI performance (per unit) | Up to 1 petaflop AI compute | Up to 1 petaflop AI compute |
| Max local model size | ~100 billion parameters (approximate, precision-dependent) | Larger models supported with headroom |
| Storage | Reportedly half of 128GB model | Full capacity (exact figure per Nvidia spec sheet) |
| Sales channel | OEM partners: Acer, ASUS, Dell, Gigabyte, HP, MSI | Nvidia direct and partner channels |
What stands out in that comparison is the middle section: processor, bandwidth, and networking are all unchanged. Nvidia did not build a stripped-down chip for the cheaper tier. It built the same silicon with less memory attached, which is exactly the kind of move a company makes when the memory itself, not the compute, is the scarce and expensive part of the bill of materials.
What 64GB actually buys an AI developer
Nvidia has positioned DGX Spark as a way for developers, researchers, and small AI teams to prototype and fine-tune models locally before moving workloads to a data center or cloud GPU cluster. According to Nvidia’s figures reported in the Perplexity-aggregated coverage, the smaller 64GB configuration is intended to run AI models with up to approximately 100 billion parameters locally, depending on quantization, precision, and the specific software stack in use. That is still a meaningful capability. A 100-billion-parameter class model, run at reduced precision, covers a large share of the open-weight models developers are currently experimenting with, from mid-size reasoning models to multimodal systems that do not require frontier-scale parameter counts.
Where the memory ceiling bites is in the upper range: workloads that assume the full 128GB pool, whether that is a single large model loaded at higher precision or several models running concurrently for a multi-agent pipeline, will not fit on the 64GB box without aggressive quantization or offloading. For many solo developers and small teams, that is an acceptable tradeoff for saving roughly $2,000 off a comparably configured 128GB unit’s new street price. For teams doing serious fine-tuning work on frontier-scale open models, the 128GB variant, expensive as it has become, remains the only option.
Clustering changes the math
One detail that outlets including The Register emphasized is that both Spark variants retain the onboard ConnectX-7 NIC rated at 200Gbps, and that DGX Spark systems can be clustered together using that networking. Nvidia now says DGX Spark can support clusters of up to eight units in manual configurations, potentially providing up to 1TB of aggregate unified memory. PCMag’s coverage notes that two 64GB systems linked together can present a combined 128GB memory pool, effectively replicating the capacity of a single 128GB unit using two of the cheaper boxes.
Do the arithmetic on that and the pricing story gets more interesting. Two 64GB Sparks at $4,999 apiece comes to $9,998, which is well above the $6,950 now being asked for a single 128GB unit. So clustering is not a way to undercut the 128GB model on price for equivalent memory. What it does offer is a path to scale compute incrementally, since two clustered Sparks also roughly double the available AI performance, moving from a single petaflop-class system toward 2 petaflops of combined compute, according to the aggregated reporting. For teams that want to grow capacity over time rather than committing to a single large purchase upfront, buying a 64GB Spark now and adding a second unit later is a more flexible, if ultimately pricier, path than it first appears.
How DGX Spark pricing compares to rival mini AI systems
Nvidia is not the only company selling compact, memory-rich machines aimed at local AI work. Tech Insider has already broken down the fuller three-way comparison in its DGX Spark vs. Framework Desktop vs. Mac Studio analysis, and the categories worth tracking for context are AMD’s Ryzen AI Max platform and Apple’s unified-memory Mac lineup, both of which compete for the same “run a large model on your desk” buyer.
AMD’s approach, detailed in Tech Insider’s coverage of the Ryzen AI Max PRO 400’s 192GB of unified memory, leans on x86 compatibility and a memory ceiling that actually exceeds DGX Spark’s 128GB flagship, aimed at running up to 300-billion-parameter class models locally according to that reporting. Apple’s Mac Studio line, by contrast, trades raw memory bandwidth and Nvidia-ecosystem compatibility (CUDA, TensorRT, the broader Nvidia AI software stack) for power efficiency and macOS integration. On that last point, Nvidia’s own software stack has kept moving alongside the hardware: its October 2026 DGX Spark software release began rolling out on October 21 and 22, adding Linux 7.0 kernel support and Wi-Fi setup in appliance mode, and it supports both the 64GB and 128GB systems. None of these systems are direct apples-to-apples competitors on spec sheets alone. The decision for a buyer mostly comes down to which software ecosystem they are already building on: CUDA favors DGX Spark, ROCm and general x86 tooling favor AMD’s platform, and Apple Silicon’s MLX framework favors the Mac Studio.
| Platform | Starting price | Max unified memory | Primary ecosystem |
|---|---|---|---|
| Nvidia DGX Spark 64GB | $4,999 | 64GB (273GB/s bandwidth) | CUDA / Nvidia AI stack |
| Nvidia DGX Spark 128GB | ~$6,950 (current) | 128GB (273GB/s bandwidth) | CUDA / Nvidia AI stack |
| AMD Ryzen AI Max PRO 400 class systems | Varies by OEM | Up to 192GB | ROCm / x86 general compute |
| Apple Mac Studio (unified memory) | Varies by configuration | Varies by configuration | Apple Silicon / MLX |
The broader competitive picture is also shaped by who else is even in the GPU and AI silicon business at all. Tech Insider’s rundown of which companies make GPUs besides Nvidia and AMD is a useful reminder that Nvidia’s dominance in the discrete GPU market, reported at roughly 90% share, is exactly why it can set terms like this: a mid-cycle price hike on an existing SKU and the introduction of a cheaper, memory-reduced tier, without much competitive pressure forcing its hand. There simply are not many other vendors building anything comparable to the GB10 Grace Blackwell Superchip at this size and price point.
Not to be confused with RTX Spark
Buyers researching this launch should be careful not to conflate DGX Spark with a separate Nvidia product, RTX Spark, which Tech Insider covered when it arrived at $2,899 with 128GB of RAM as a Windows-on-Arm PC platform. The naming overlap is a genuine source of confusion in search results and retailer listings. RTX Spark targets Windows desktop and creative workloads on Arm silicon, while DGX Spark is a Linux-based, CUDA-first machine built specifically for AI model development and fine-tuning. The two share a brand name and a general “compact AI-capable desktop” positioning, but they are built for different buyers running different software stacks.
The memory crunch behind the price hike
The supply-side story driving all of this is bigger than any single product. Industry trackers including TrendForce have spent much of 2026 documenting how data center AI buildouts are consuming advanced memory supply (both HBM for accelerators and conventional DRAM for everything else) faster than manufacturers can expand capacity. Micron’s own disclosures, which Tech Insider covered in its reporting on the 20% supply growth failing to outpace demand, describe a market where every incremental unit of memory capacity added to the supply chain gets absorbed almost immediately by AI infrastructure buyers willing to pay a premium.
Note: the exact percentage increases in HBM and DDR5 spot pricing for October 2026 have not been independently confirmed across the outlets reviewed for this report, and this piece avoids citing specific numeric price-per-gigabyte figures that are not verified in named sources. What is confirmed, repeatedly, across PCMag, Tom’s Hardware, The Register, and Hardware Busters, is the qualitative effect: Nvidia built a lower-capacity DGX Spark specifically because sourcing enough LPDDR5X for every unit at the original 128GB specification has become more expensive, and it raised the price of the existing 128GB model by nearly 75% over its original launch price in the process.
This is consistent with what Micron’s leadership has said about 2027 supply commitments already running around 75% sold out, a figure Tech Insider reported from Micron’s CEO remarks on the memory supply outlook. When a major memory supplier is telling the market its capacity is largely spoken for more than a year ahead, hardware vendors downstream, including Nvidia, are left with three options: raise prices, reduce specifications, or both. DGX Spark’s October 2026 lineup shows Nvidia doing exactly that, simultaneously.
Historical context: how we got here
DGX Spark launched as Nvidia’s answer to a specific gap in its product lineup: a machine small enough to sit on a desk, but with enough unified memory and AI-specific compute to let individual developers and small labs work with large models without renting cloud GPU time. At its original debut in October 2025, the 128GB configuration’s $3,999 price point was positioned as disruptive, a fraction of the cost of a multi-GPU workstation while still offering Grace Blackwell-class AI performance. Nvidia revisited that price just four months later, raising the same 128GB unit to $4,699 in February 2026 and citing memory-supply constraints, a first warning sign of the repricing that would follow.
Less than a year later, that pricing has moved substantially, and Nvidia has felt compelled to introduce a cheaper, reduced-memory SKU just to keep a sub-$5,000 entry point available at all. That is a notably fast repricing cycle for a flagship compute product, and it mirrors broader patterns Tech Insider has tracked across the AI hardware stack this year, from GPU resale values holding unusually high (as seen in reporting on the Nvidia B200’s resale value hitting 158% of launch price) to component shortages reshaping product roadmaps across the industry. The common thread in all of these stories is the same: AI infrastructure demand is pulling resources, attention, and pricing power away from more conventional computing products, and local AI development hardware like DGX Spark sits right at that intersection.
Market impact: who gains and who gets squeezed
For Nvidia, the 64GB DGX Spark is a way to defend a price point. A company that built its reputation partly on making AI development accessible to smaller teams cannot afford to let its entry-level local AI box drift toward $7,000 without offering some alternative underneath it. Introducing a $4,999 tier, even one with less memory and presumably thinner margins given component costs, keeps DGX Spark inside a psychologically important bracket for individual developers, university labs, and startups working with limited hardware budgets.
For OEM partners, including Acer, ASUS, Dell, Gigabyte, HP, and MSI, selling the 64GB model gives them a product to put in front of customers who were priced out of the 128GB unit’s new cost structure. It also diversifies Nvidia’s DGX Spark distribution away from direct sales, which should, in theory, improve availability and reduce the kind of scalper-driven scarcity pricing that has plagued other high-demand Nvidia hardware.
For buyers, the picture is mixed. Developers who only need to run sub-100-billion-parameter models get a genuinely useful machine at a price close to the original DGX Spark’s debut cost. But anyone who already budgeted for a 128GB unit at or near its original $3,999 price is now facing a nearly 75% cost increase for the same specification, with no cheaper path to that exact memory capacity other than clustering two of the new 64GB units, which costs even more in aggregate. The memory crunch, in other words, is not creating losers and winners cleanly. It is raising the floor price of entry across the board while giving Nvidia room to claim it still offers a “budget” option.
What to watch next
A few open questions will shape how this story develops between now and the October 23 launch date and beyond.
- Nvidia has not published the exact storage capacity for the 64GB DGX Spark; that figure should appear on Nvidia’s official product page closer to launch and will matter for developers planning local dataset storage alongside model weights.
- Whether the 128GB DGX Spark’s price holds at roughly $6,950 or moves again before year-end will be a useful signal of whether the memory market is stabilizing or still tightening heading into 2027.
- OEM retail listings from Acer, ASUS, Dell, Gigabyte, HP, and MSI should clarify whether regional pricing and bundling (storage upgrades, software packages) vary meaningfully across partners.
- Competing unified-memory platforms from AMD and Apple may respond with their own pricing or configuration changes if the DRAM and LPDDR5X supply crunch continues to tighten through early 2027.
- Clustering adoption, specifically whether developers actually buy multiple 64GB units to build combined memory pools rather than simply paying more upfront for a single 128GB unit, will be worth tracking as a real-world indicator of how buyers are responding to the new pricing structure.
Predictions for the rest of 2026 and into 2027
Based on the confirmed facts around this launch and the broader memory market context Tech Insider has tracked throughout the year, a few outcomes look likely. First, expect more vendors across the “local AI workstation” category to introduce reduced-memory, lower-price SKUs over the next two to three quarters, following the same playbook Nvidia just used with DGX Spark, rather than holding firm on existing specifications at higher prices. Second, given Micron’s reported 75% presold 2027 capacity, further price increases on memory-dependent AI hardware are more likely than price stabilization before mid-2027. Third, OEM-channel distribution for niche AI hardware, rather than direct-to-consumer sales, is likely to become more common as companies like Nvidia look to spread availability risk and let retail partners absorb some of the logistics of a tight supply environment. Fourth, clustering features like DGX Spark’s ConnectX-7 networking will likely get more marketing emphasis industry-wide, since vendors benefit from presenting “buy two smaller units” as a feature rather than a workaround for memory shortages. Fifth, expect continued scrutiny of Nvidia’s nearly 90% discrete GPU market share, as reduced competition gives the company unusual latitude to reprice existing hardware well after launch without losing customers to rivals.
Frequently asked questions
When does the Nvidia DGX Spark 64GB go on sale?
According to PCMag, Tom’s Hardware, and Hardware Busters, the 64GB configuration is scheduled for availability on Friday, October 23, 2026, through OEM partners.
How much does the Nvidia DGX Spark 64GB cost?
The 64GB DGX Spark starts at $4,999, according to reporting aggregated from PCMag, Tom’s Hardware, and Hardware Busters.
Why did the 128GB DGX Spark get more expensive?
Outlets covering the launch attribute the price increase, from roughly $3,999 at original launch to about $6,950 currently, to a broader surge in memory component costs driven by AI data center demand for advanced memory supply.
Is the memory bandwidth lower on the 64GB model?
No. Reporting from Tom’s Hardware, The Register, and Nvidia’s own product listings indicates both the 64GB and 128GB DGX Spark configurations share the same 273GB/s memory bandwidth figure.
Can I cluster multiple DGX Spark units together?
Yes. Both configurations include a ConnectX-7 NIC rated at 200Gbps, and Nvidia says up to eight units can be clustered in manual configurations, with PCMag noting that two 64GB units can present a combined 128GB memory pool.
What is the difference between DGX Spark and RTX Spark?
DGX Spark is a Linux-based, CUDA-first machine built for local AI model development. RTX Spark is a separate Nvidia product, a Windows-on-Arm PC platform that launched at $2,899 with 128GB of RAM, aimed at general desktop and creative workloads rather than AI-specific development.
How much local AI model capacity does the 64GB DGX Spark support?
Nvidia’s figures, as reported in coverage of the launch, indicate the 64GB configuration is intended to run AI models with up to approximately 100 billion parameters locally, depending on precision and memory usage.
Will DGX Spark prices come down again?
That depends on the broader memory market. With Micron reporting roughly 75% of its 2027 memory supply already committed, further near-term price relief looks unlikely, though this report does not speculate beyond what named sources have confirmed.
