Nvidia spent three decades selling graphics chips that other companies bolted into their own computers. On May 31, 2026, that changed. At the Taipei Music Center, CEO Jensen Huang introduced RTX Spark, a new class of chip Nvidia calls a “superchip,” built to power Windows laptops and compact desktops directly rather than sit inside someone else’s design. It pairs a 20-core Grace CPU with a Blackwell RTX GPU carrying 6,144 CUDA cores, up to 128GB of unified memory, and a claimed 1 petaflop of FP4 AI performance. “RTX Spark is our new superchip for Windows PCs,” Huang told the Computex 2026 crowd, unveiling a new lineup that spans laptops, compact desktops, and enterprise workstations.
The announcement puts Nvidia in a business it has never directly competed in: selling a full Windows PC platform built around its own CPU design. Six laptop makers, Asus, Dell, HP, Lenovo, Microsoft Surface, and MSI, are already lined up to ship RTX Spark machines this fall, with Acer and Gigabyte expected to follow. That lineup puts Nvidia RTX Spark head to head with Qualcomm, AMD, Intel, and Apple in the AI-PC laptop market for the first time, and it does so with a spec sheet that looks nothing like a typical Copilot+ PC chip.
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 Is Nvidia RTX Spark?
Nvidia showed the chip on stage under two development codenames, N1 and N1X, before settling on its retail name. RTX Spark anchors a family of three machine types Nvidia unveiled at Computex 2026: RTX Spark laptops, RTX Spark compact desktops, and a much larger workstation sibling called DGX Station for Windows. All three share the same underlying idea. Instead of a separate CPU talking to a separate discrete GPU over a narrow bus, RTX Spark fuses a 20-core Grace CPU and a Blackwell GPU on one TSMC 3-nanometer die, connected by Nvidia’s own NVLink-C2C interconnect. Huang told the keynote audience the chip packs 70 billion transistors.
Nvidia and Microsoft announced RTX Spark jointly, and Nvidia’s own announcement calls it the world’s first Windows PCs built specifically for personal AI agents. That framing matters. Rather than pitching RTX Spark as a faster gaming laptop chip, Nvidia is positioning it as infrastructure for AI that runs locally instead of in the cloud, borrowing the same Grace-plus-Blackwell architecture Nvidia already sells to data centers and shrinking it to fit inside a machine that weighs about three pounds. According to Nvidia’s Computex 2026 product page, the platform combines CUDA, RTX, DLSS, FP4, TensorRT, OptiX, Reflex, and G-SYNC into a single chip aimed at slim Windows laptops and compact desktop PCs.
Inside the Silicon: Grace CPU Meets Blackwell GPU
Nvidia and MediaTek co-designed the CPU side of RTX Spark, and the GPU side reuses the Blackwell architecture from Nvidia’s current graphics lineup. A widely-cited recap of the keynote, sourced from Nvidia’s own product presentation, put it this way: “Developed in partnership with MediaTek and built on TSMC’s 3nm process, the RTX Spark (codenamed N1X) is a massive system-on-a-chip (SoC) designed to bring supercomputer-level AI and workstation performance to ultra-slim laptops and ultra-efficient compact desktops.”
The Grace CPU Side
The top-tier RTX Spark configuration uses 20 Arm-based CPU cores, reportedly split between 10 high-performance Cortex-X925 cores and 10 efficiency-focused Cortex-A725 cores, peaking around 4.1GHz according to HotHardware’s breakdown of the announcement. That core mix looks closer to a phone chip’s big.LITTLE design than a traditional laptop CPU, which tracks given Nvidia built this alongside MediaTek, a company that has spent years tuning Arm silicon for phones and tablets.
The Blackwell GPU Side
The GPU half carries 6,144 CUDA cores built on fifth-generation Tensor Cores with FP4 precision support, the same low-precision math format Nvidia uses to squeeze more throughput out of its data center Blackwell chips. The two halves talk over NVLink-C2C, which HotHardware clocked at up to 600GB/s of chip-to-chip bandwidth. Nvidia rates the whole package for a 45W-to-80W typical power envelope, with transient boosts as high as roughly 245W, numbers that will matter a lot once laptop makers start publishing battery life figures.
128GB of Unified Memory: Why It Changes the Local AI Math
The headline spec isn’t the CPU or the GPU. It’s the memory. RTX Spark supports up to 128GB of LPDDR5X unified memory shared between the CPU and GPU, with roughly 300GB/s of bandwidth according to Tom’s Hardware’s Computex coverage. Nvidia’s own product materials pose the point almost as a rhetorical question: “The major feature? Up to 128GB of LPDDR5X unified memory and a staggering 1 Petaflop of AI compute.” Unified memory means the CPU and GPU draw from the same pool instead of copying data back and forth between separate VRAM and system RAM, which is exactly the bottleneck that keeps most consumer laptops from running large AI models locally in the first place.
The rough math explains why 128GB is the number Nvidia keeps repeating. Large language models need memory roughly proportional to their parameter count and the precision they run at, and FP4 quantization (the format RTX Spark’s Tensor Cores are built for) cuts that requirement dramatically compared to the 16-bit precision most cloud inference still uses.
Rough local-LLM memory math (industry rule of thumb)
memory_needed_GB ≈ parameters_billion x bytes_per_parameter
FP16 (16-bit): 70B params x 2.0 bytes ≈ 140GB
INT8 (8-bit): 70B params x 1.0 byte ≈ 70GB
FP4 (4-bit): 120B params x 0.5 byte ≈ 60GB (plus KV-cache overhead)
RTX Spark: up to 128GB unified LPDDR5X, shared by CPU and GPU
-> leaves headroom for a roughly 120B-parameter model at FP4For readers who want to try this kind of workload on hardware they already own, our guide to running an LLM locally walks through the same memory tradeoffs on existing consumer GPUs.
Bandwidth Math: Why 300GB/s Sets the Real Ceiling
Nvidia’s two headline numbers, 128GB of capacity and 1 petaflop of compute, get most of the attention, but a third figure does more to predict how RTX Spark actually feels to use: memory bandwidth. The Grace CPU and Blackwell GPU talk to each other over NVLink-C2C at up to 600GB/s, but the LPDDR5X memory pool itself delivers roughly 300GB/s to whichever processor is asking for it, per Tom’s Hardware’s and StorageReview’s Computex coverage. That gap isn’t a flaw, unified-memory systems are built this way on purpose, but it does mean the CPU-GPU interconnect was never the bottleneck Nvidia needed to solve. The real constraint on speed is how fast that 300GB/s figure can feed a large model’s weights into the GPU’s Tensor Cores. It’s also why Nvidia leans so heavily on FP4 quantization in RTX Spark’s software stack: per the memory math above, a model quantized to 4-bit weights moves roughly a quarter of the data through that 300GB/s pipe compared to running the same parameter count at 16-bit precision, which does more to offset the bandwidth gap than any hardware change could.
That distinction matters because generating text with a language model is typically a bandwidth problem, not a compute problem, once the model already fits in memory. Every token a model produces requires reading through its active parameters again, so a chip’s realistic token-generation speed tracks memory bandwidth far more closely than it tracks headline petaflop figures. RTX Spark’s 300GB/s trails what dedicated GDDR7 or HBM memory delivers on desktop gaming GPUs, though none of those cards can approach RTX Spark’s 128GB capacity at any price. The tradeoff Nvidia is making is explicit: RTX Spark gives up raw bandwidth in exchange for enough capacity to hold models that would never fit on a conventional laptop GPU at all, regardless of how fast that GPU’s own memory runs. With RTX Spark still months from shipping as of August 2026, whether that tradeoff feels fast enough in daily use remains untested outside Nvidia’s own labs.
Performance Claims: 1 Petaflop and the 120-Billion-Parameter Question
Nvidia rates RTX Spark at 1 petaflop of FP4 AI performance, and outlets present at the briefing pushed further on what that means in practice. KitGuru’s Computex report says RTX Spark can run large language models up to 120 billion parameters locally, positioned as letting “the masses” run personal AI agents at home instead of renting cloud GPU time.
That figure is worth watching closely once independent reviewers get hardware in hand. Nvidia’s own DGX Spark desktop, which shares the same 128GB memory ceiling and 1-petaflop rating, was marketed at launch as capable of running models up to 200 billion parameters. Why RTX Spark’s local-model claim lands lower than its desktop sibling isn’t explained in Nvidia’s public materials, and it could come down to differences in how Windows manages memory versus the Linux-based DGX OS, or simply more conservative marketing math the second time around. Either way, nobody outside Nvidia has independently benchmarked RTX Spark yet, so both numbers should be treated as vendor claims until third-party reviewers publish results this fall.
Release Date, Launch Partners, and Pricing
Nvidia confirmed RTX Spark systems ship “this fall,” without naming an exact date, from six launch partners: Asus, Dell, HP, Lenovo, Microsoft Surface, and MSI. Acer and Gigabyte are expected to add models after launch. No official USD pricing has been announced for any configuration, and as of August 2026 that’s still the case — Nvidia hasn’t sharpened “this fall” into a firm date or published pricing for any tier. HotHardware’s reporting describes four distinct silicon tiers rather than a single chip, which suggests Nvidia and its partners plan to spread RTX Spark across multiple price points rather than launch a single flagship SKU.
| Configuration | CPU Cores | GPU (CUDA Cores) | Max Unified Memory | Likely Positioning |
|---|---|---|---|---|
| RTX Spark (top tier) | 20-core Grace (10x Cortex-X925 + 10x Cortex-A725) | 6,144 | 128GB LPDDR5X | Premium laptops & compact desktops |
| Tier 2 | 18-core | 5,120 | 64GB | Mainstream AI laptops |
| Tier 3 | 12-core | 2,560 | 64GB | Value AI laptops |
| Tier 4 | 10-core | 2,048 | 64GB | Entry-level AI PCs |
Source: HotHardware’s Computex 2026 teardown of Nvidia’s RTX Spark announcement.
Pre-Orders and Availability: Where Things Stand
As of August 2026, Nvidia has not opened pre-orders for RTX Spark, and none of the six confirmed launch partners, Asus, Dell, HP, Lenovo, Microsoft Surface, or MSI, have published a price or order page for the hardware. That tracks with Nvidia’s own language: the company has only ever committed to “this fall” for availability, without narrowing that window further since the Computex 2026 announcement. Buyers hoping to reserve a unit ahead of launch don’t have an official channel to do so yet, and Nvidia’s track record with DGX Spark, which shipped roughly five months later than its earliest expected timeline, suggests “fall 2026” itself carries some slip risk. Watching Nvidia’s newsroom and the confirmed OEM partners directly remains the most reliable way to catch an actual order window, since no authorized listings exist anywhere yet.
Should You Wait for RTX Spark, or Buy Now?
With no pricing, no ship date firmer than “this fall,” and rising memory costs all still unresolved as of August 2026, the practical question for most shoppers isn’t whether RTX Spark looks impressive on paper, it clearly does, but whether it’s worth waiting for. The answer depends heavily on what a buyer actually plans to run on the machine.
| Buyer type | Wait for RTX Spark? | Why |
|---|---|---|
| Local AI / CUDA developers | Yes | Native CUDA, TensorRT, and OptiX support plus up to 128GB unified memory has no direct equivalent on Snapdragon X or Copilot+ NPU hardware today |
| General productivity / office use | No | Copilot+ PCs from Qualcomm, AMD, and Intel already ship today at established prices, likely well below RTX Spark’s eventual positioning |
| Gamers | No | RTX Spark’s spec sheet and Nvidia’s own messaging center on AI compute and memory capacity, not gaming performance |
| Creative / engineering software users | Depends | Hinges on whether specific creative or engineering tools ship native Arm builds by fall 2026; unaddressed by launch partners as of August 2026 |
| Enterprise AI teams | Consider DGX Station instead | Up to 748GB of memory and 20 petaflops of FP4 compute for teams that have outgrown a single 128GB machine |
Recommendations based on confirmed specs and pricing as of August 2026. RTX Spark pricing has not been announced for any tier.
The CUDA advantage is the strongest argument for waiting. Nobody building on Nvidia’s AI software stack has ever had that tooling available natively on a Windows-on-Arm laptop before, and that alone closes a gap Qualcomm’s Snapdragon X chips have never been able to. But two real costs sit on the other side of that decision. First, price: the only public reference point is DGX Spark’s $3,999 desktop, and RTX Spark’s four tiers stretch from an entry-level configuration up to that same 128GB ceiling, so a fully specced laptop could land anywhere from noticeably cheaper to noticeably more expensive than that number once OEMs actually price their hardware. Second, timing risk: DGX Spark’s own five-month slip between its expected and actual ship dates is the only precedent Nvidia has set for this exact architecture, and “this fall” still isn’t a firm date more than two months after the Computex 2026 announcement.
For buyers who don’t specifically need local LLM capacity beyond what a 40-to-80 TOPS NPU already covers, an existing Copilot+ PC or a Snapdragon X2 Elite Extreme machine remains the lower-risk purchase today. For buyers whose workload is currently bottlenecked by cloud GPU rental costs or API limits, RTX Spark is the first Windows laptop chip built to actually solve that problem locally, which makes waiting a calculated bet rather than a simple upgrade decision.
RTX Spark vs. DGX Spark vs. DGX Station: Nvidia’s Own Compute Ladder
RTX Spark doesn’t exist in isolation. It’s the middle rung of a three-product ladder Nvidia has now built entirely around the Grace-plus-Blackwell formula, aimed at three very different buyers.
| Product | Announced | Ships | Form Factor | Memory | AI Performance | Price |
|---|---|---|---|---|---|---|
| DGX Spark | Jan. 2025 (as Project DIGITS) | Oct. 15, 2025 | Desktop mini-PC | 128GB LPDDR5X | 1 petaflop FP4 | $3,999 |
| RTX Spark | May 31, 2026 | Fall 2026 | Laptop / compact desktop | Up to 128GB LPDDR5X | 1 petaflop FP4 | Not yet announced |
| DGX Station for Windows | May 31, 2026 | Not yet announced | Deskside workstation | Up to 748GB | Up to 20 petaflops FP4 | Not yet announced |
Sources: Nvidia, Tom’s Hardware, Forbes, StorageReview.
DGX Spark, the desktop version, already proved Nvidia could pack a petaflop of FP4 compute and 128GB of unified memory into a 1.2-kilogram box roughly the size of a small router. RTX Spark takes that same recipe and adapts it for mainstream Windows laptops instead of a niche developer tool. DGX Station goes the other direction entirely, aimed at enterprise teams who want to run frontier-scale models on a desk instead of in a data center rack.
How RTX Spark Stacks Up Against Qualcomm, AMD, and Intel
RTX Spark arrives into a Windows AI-PC market that already has three established chip vendors, each measuring “AI performance” a different way. Microsoft’s Copilot+ PC badge requires a minimum of 40 TOPS from a dedicated neural processing unit, alongside 16GB of RAM and a 256GB SSD. Qualcomm, AMD, and Intel have all cleared that bar comfortably.
| Chip | Maker | CPU Cores | AI Rating | Status |
|---|---|---|---|---|
| RTX Spark | Nvidia | 20-core (Arm) | 1 petaflop FP4 (full SoC via GPU) | Ships fall 2026 |
| Snapdragon X2 Elite Extreme | Qualcomm | Up to 18 | 80 TOPS NPU (INT8) | Shipping H1 FY2026 |
| Ryzen AI 400 series | AMD | Varies by SKU | 60 TOPS NPU | Shipping 2026 |
| Core Ultra Series 3 (Panther Lake) | Intel | Varies by SKU | ~50 TOPS NPU | Shipping 2026 |
| Snapdragon X (original) | Qualcomm | Up to 12 | 45 TOPS NPU | Shipping since 2024 |
Sources: Futurum Group, PCWorld, TechDaily.ai. RTX Spark’s petaflop figure and competitors’ TOPS ratings measure different things and are not directly interchangeable (see note below).
That last line matters more than it looks. TOPS figures describe a small, low-power NPU block designed to run background Windows AI features efficiently. Nvidia’s 1-petaflop figure describes the entire Blackwell GPU running FP4 tensor math flat out, a much larger and more power-hungry piece of silicon. Comparing them directly is a bit like comparing a car’s city fuel economy to a truck’s towing capacity. Both numbers are real, but they answer different questions, and RTX Spark’s advantage really comes down to memory capacity rather than the raw AI-rating comparison. For more on how Nvidia’s other current chips stack up on paper, see our breakdown of the RTX 5090 vs. RTX 5070 Ti for AI workloads, and for the handheld side of the AI-chip race, our look at Intel’s Panther Lake handheld chips.
RTX Spark’s Four Tiers Against the Copilot+ PC Baseline
Every RTX Spark configuration clears Microsoft’s Copilot+ PC baseline by a wide margin, but the size of that margin changes depending on which of the four confirmed tiers ends up in a given laptop. Lining up Microsoft’s minimum spec against Nvidia’s own tier breakdown shows how much headroom Nvidia is building in even at the lowest published configuration.
| Spec | Copilot+ PC Minimum | RTX Spark Entry Tier | RTX Spark Top Tier |
|---|---|---|---|
| GPU (CUDA cores) | Not applicable (NPU-based) | 2,048 | 6,144 |
| CPU cores | Not specified | 10-core Grace | 20-core Grace |
| Unified/system memory | 16GB | 64GB | 128GB |
| Rated AI performance | 40 TOPS (NPU) | Not broken out by Nvidia | 1 petaflop FP4 |
Sources: Microsoft Copilot+ PC requirements; Nvidia/HotHardware RTX Spark tier breakdown, cited earlier in this article. Nvidia has published a single petaflop rating for the flagship configuration and hasn’t broken out AI-performance figures for Tiers 2 through 4.
The memory gap is the number that stands out most. Even RTX Spark’s entry-level tier ships with four times the unified memory Microsoft requires for base Copilot+ certification, and Nvidia hasn’t published any configuration near the 16GB floor. That tracks with Nvidia’s broader pitch: RTX Spark isn’t aiming to squeak past a Copilot+ checkbox, it’s treating the badge as a floor and building every tier well above it, which is also why none of the four configurations have been positioned by Nvidia or its partners as budget hardware.
Why Nvidia Just Became a Windows PC Chip Maker
Nvidia’s core business is still data center GPUs, and that isn’t changing. What is changing is where Nvidia thinks the next growth curve sits. The company has spent 2026 pulling back from parts of the traditional gaming GPU cycle, our earlier coverage detailed how Nvidia skipped a new generation of gaming GPUs entirely this year, breaking a three-decade release cadence. RTX Spark helps explain where some of that engineering attention went instead.
By building its own CPU-GPU combo chip instead of just selling discrete graphics silicon to OEMs, Nvidia captures more of the bill of materials in every RTX Spark machine sold, the CPU margin, the GPU margin, and the platform licensing that comes with owning the reference design. It also ties RTX Spark to Nvidia’s CUDA software stack in a way a plain discrete GPU never could on a Windows-on-Arm machine, since previous Arm-based Windows laptops from Qualcomm couldn’t run CUDA software at all. That software lock-in, more than the raw silicon, may end up being Nvidia’s actual competitive moat here.
The Memory Crunch Behind Every AI PC Launch in 2026
RTX Spark’s 128GB memory promise lands at an awkward moment for memory pricing. Industry reporting from July 2026 points to significant DRAM price increases throughout the year, driven largely by AI data center demand pulling memory supply away from consumer devices, with laptops, workstations, and servers all absorbing higher costs. LPDDR5X, the specific memory type RTX Spark uses, isn’t immune to that same supply pressure.
That backdrop makes Nvidia’s decision to lead with a 128GB spec even bolder. Every gigabyte of unified memory Nvidia packs into RTX Spark is a gigabyte OEMs have to source at 2026 prices, not 2024 prices. Our recent coverage of the broader RAM shortage hitting gaming hardware prices found the same AI-driven demand squeeze pushing costs up across five separate gaming platforms this year. RTX Spark is effectively betting that buyers will pay a premium for local AI capability precisely when the memory that capability depends on has gotten more expensive to source.
What This Means for Qualcomm’s Snapdragon X and Copilot+ PCs
Qualcomm spent the past two years building the Windows-on-Arm business almost single-handedly, positioning Snapdragon X chips as the default silicon behind Microsoft’s Copilot+ PC branding. RTX Spark complicates that story in an unusual way, because Microsoft is a launch partner on both sides. Microsoft Surface will ship RTX Spark hardware this fall, the same company that helped define what a Copilot+ PC is supposed to look like around Qualcomm silicon in the first place.
Qualcomm isn’t standing still. Its Snapdragon X2 Elite Extreme, unveiled with up to 18 CPU cores and an 80 TOPS NPU, is aimed at H1 FY2026 availability and remains the far cheaper option for OEMs building mainstream Copilot+ laptops. RTX Spark isn’t trying to win that volume segment. It’s aimed higher, at buyers who specifically want to run bigger local models than a 45-to-80 TOPS NPU can realistically handle, which is a narrower but potentially more defensible niche.
Will Your Existing Software Run on RTX Spark?
RTX Spark’s Grace CPU raises a question that has shadowed every Arm-based Windows laptop before it: how well does existing Windows software actually run on Arm silicon? Windows 11 handles that gap with an emulation layer that translates x86 and x64 instructions for Arm processors, the same approach Qualcomm’s Snapdragon X chips have relied on since their launch. Native Arm apps run at full speed, but any program that hasn’t been recompiled for Arm still pays a performance tax through emulation, and plenty of niche business software, older games, and specialized creative tools still haven’t made that jump.
RTX Spark complicates this picture in a way Qualcomm’s chips never could: CUDA. Nvidia’s software stack, the same CUDA, TensorRT, and OptiX libraries that already run on Nvidia’s data center and desktop GPUs, works on RTX Spark’s Blackwell GPU natively, per Nvidia’s own Computex product materials. That’s a meaningfully different situation than a Snapdragon X machine, where GPU-accelerated software written for Nvidia or AMD hardware has no equivalent path to run well at all. For AI developers specifically, tools built around CUDA, likely the largest existing ecosystem of GPU-accelerated AI software, should carry over to RTX Spark without the emulation penalty that hits general-purpose Windows applications. Whether mainstream creative and productivity software makers close that gap by the time RTX Spark actually ships is a separate question, and one none of the six launch partners has addressed publicly as of August 2026. For buyers who spend most of their time in specialized creative or engineering software rather than CUDA-based AI tools, that emulation question is arguably more important to next fall’s purchase decision than any petaflop or memory figure Nvidia has published so far.
The Developer and Enterprise Angle: Local AI Agents
Nvidia’s messaging around RTX Spark leans heavily on “personal agents,” AI systems that run continuously on-device rather than firing off a cloud API call every time a user asks a question. That’s a meaningfully different pitch than the Copilot+ PC generation, which mostly used its NPU for narrow features like live captions, background blur, and Windows Studio Effects.
A 128GB unified memory pool changes what’s realistic to run at the edge. Developers who currently split workloads between a local machine and a rented cloud GPU, the exact tradeoff we cover in our step-by-step guide to running an LLM locally, get a meaningfully larger local memory budget to work with once RTX Spark hardware ships. Whether software actually catches up to that hardware capability by fall 2026 is a separate question. Nvidia’s CUDA and TensorRT tooling already runs on DGX Spark today, which gives RTX Spark a head start that a brand-new architecture wouldn’t have.
From Project DIGITS to RTX Spark: A Short History
RTX Spark didn’t appear out of nowhere. Nvidia first teased the underlying idea in early 2025 under the name Project DIGITS, then detailed it further at its March 2025 GTC keynote and renamed it DGX Spark. The desktop version was originally expected to ship around May 2025, according to Tom’s Hardware’s reporting, but actually reached customers on October 15, 2025, at $3,999, roughly five months later than first suggested.
That delay is worth remembering now. DGX Spark used the same core promise RTX Spark makes today, a petaflop of FP4 compute and 128GB of unified memory in a small form factor, and it still slipped its original timeline before landing at a higher starting price than early previews suggested. Forbes described the finished DGX Spark as data center AI packed into a desktop box roughly 150mm square and weighing 1.2 kilograms, capable of handling models up to 200 billion parameters locally, with up to 70 billion parameters supported for fine-tuning. RTX Spark inherits that legacy directly, for better and for worse.
| Date | Milestone |
|---|---|
| Early 2025 | Nvidia first teases the architecture as Project DIGITS |
| March 2025 | Nvidia details the chip at GTC, renames it DGX Spark, targets a May 2025 ship date |
| October 15, 2025 | DGX Spark actually ships, roughly five months later than first suggested, at $3,999 |
| May 31, 2026 | Nvidia unveils RTX Spark at Computex 2026, alongside DGX Station for Windows |
| Fall 2026 | RTX Spark laptops and desktops targeted to ship from six launch partners |
Timeline compiled from Nvidia’s own announcements and Tom’s Hardware’s reporting, cited earlier in this article.
Laid out chronologically, the pattern is hard to miss. DGX Spark’s own internal target slipped by roughly five months, from an expected May 2025 ship date to its actual October 15, 2025 launch. RTX Spark is now more than two months past its own Computex 2026 unveiling with still no date firmer than “this fall” as of August 2026, the same open-ended window Nvidia used for DGX Spark before that slip. That doesn’t guarantee RTX Spark repeats the delay, but it’s the closest precedent available, and it’s Nvidia’s own.
Expert and Industry Reactions
Coverage of the Computex 2026 keynote has been broadly consistent on the facts, even where outlets disagree on how significant the moment is. Jensen Huang used the keynote itself to frame RTX Spark as a platform shift rather than a single product launch, telling the Computex 2026 audience that the new lineup spans desktop, laptop, and workstation form factors built on the same Grace-plus-Blackwell architecture.
Trade press reaction has focused on the same tension that runs through this entire launch: genuinely new hardware capability, paired with an unproven claim about how many people actually want to run AI agents locally instead of through a browser tab. Nvidia’s own product description, again pulled from its Computex 2026 keynote coverage, keeps circling back to the same two numbers regardless of which outlet is reporting it: 128GB of memory, and 1 petaflop of compute. Whether that combination is enough to justify a new PC purchase cycle is the open question every OEM partner is now betting real inventory on.
Market Impact: What RTX Spark Signals for the AI PC Race
The clearest market signal here isn’t a stock chart, it’s a strategic one. Nvidia has spent the past two years selling nearly every data center GPU it can manufacture, and demand hasn’t meaningfully cooled. Entering the Windows laptop CPU business at the same time is a statement that Nvidia sees the AI compute story extending well past the data center and into every device people carry, echoing the same broader AI infrastructure buildout our coverage of AMD’s Advancing AI push has tracked on the rival side of the chip market this year.
For OEMs, RTX Spark represents both an opportunity and a risk. It’s a genuinely differentiated spec sheet that no existing Windows laptop chip can match on paper, which gives Asus, Dell, HP, Lenovo, and MSI something new to market. But it also means committing shelf space and marketing budget to hardware built on a completely unproven consumer use case, running large local AI models on a laptop, at a moment when memory costs are already rising and pricing hasn’t even been announced. That’s a meaningfully different risk profile than adding another 2 or 3 TOPS to next year’s NPU refresh.
Predictions: Where RTX Spark Goes From Here
- Pricing lands in tiers, not one number. With four confirmed silicon configurations, expect entry RTX Spark laptops well under $2,000 and the full 128GB configuration priced closer to, or above, DGX Spark’s $3,999 desktop price.
- More OEMs join within months. Acer and Gigabyte are already confirmed as coming, following the same pattern Nvidia used when it expanded DGX Spark availability through partners after its initial launch.
- The 120-billion-parameter claim gets tested fast. Independent reviewers will benchmark real local-model performance within weeks of shipping, and results will likely depend heavily on quantization settings and context length rather than matching Nvidia’s headline number cleanly.
- Rivals respond with memory, not just TOPS. Expect Qualcomm, AMD, and Intel to start emphasizing unified-memory capacity in their own 2027 roadmaps rather than continuing to compete purely on NPU TOPS ratings.
- Memory costs pressure the final price. Rising DRAM and LPDDR5X pricing could push RTX Spark laptops higher than Nvidia’s internal targets, mirroring the price movement DGX Spark saw between its 2025 preview and its eventual launch.
Frequently Asked Questions
What is Nvidia RTX Spark?
RTX Spark is Nvidia’s first chip built specifically for Windows laptops and compact desktops, announced at Computex 2026. It combines a 20-core Grace CPU with a Blackwell GPU, 6,144 CUDA cores, and up to 128GB of unified memory.
When does RTX Spark release?
Nvidia says RTX Spark laptops and desktops ship this fall, meaning fall 2026, from Asus, Dell, HP, Lenovo, Microsoft Surface, and MSI, with Acer and Gigabyte expected to follow.
How much does RTX Spark cost?
Nvidia has not announced official pricing. Its closest sibling, the 2025 DGX Spark desktop, launched at $3,999, which offers a rough reference point for what a fully specced RTX Spark laptop might eventually cost.
Can I pre-order Nvidia RTX Spark yet?
No. As of August 2026, Nvidia has not opened pre-orders or published pricing, and none of its six launch partners have listed the hardware for sale. The company has only confirmed a “fall 2026” launch window.
What is the difference between RTX Spark and DGX Spark?
DGX Spark is a 2025 desktop mini-PC aimed at AI developers. RTX Spark brings similar Grace-plus-Blackwell silicon to mainstream Windows laptops and compact desktops for a broader consumer and creator audience.
Can RTX Spark run large language models locally?
Nvidia and early coverage cite local LLM support up to roughly 120 billion parameters, though that figure has not been independently benchmarked and may vary based on quantization and context length.
Which laptops will use the RTX Spark chip?
Confirmed launch partners are Asus, Dell, HP, Lenovo, Microsoft Surface, and MSI, with Acer and Gigabyte expected to add models after launch.
How does RTX Spark compare to Qualcomm Snapdragon X2 or Intel Panther Lake?
Those chips use dedicated NPUs rated in TOPS, with Qualcomm’s Snapdragon X2 Elite Extreme claiming 80 TOPS and Intel’s Panther Lake rated around 50 TOPS. RTX Spark’s 1-petaflop figure measures its entire Blackwell GPU running FP4 workloads, a different and much larger measurement, so the two specs aren’t directly comparable.
Does RTX Spark run Windows or Linux?
RTX Spark is built around Windows 11. Nvidia and Microsoft describe it as a platform for on-device AI agents running inside the standard Windows environment, distinct from DGX Spark’s Linux-based developer tooling.
Related Coverage
- Nvidia Skips New Gaming GPUs, Breaks 30-Year Streak [2026]
- RTX 5090 vs RTX 5070 Ti for AI: 2x VRAM, 2.7x Price [2026]
- Intel Panther Lake Handhelds Claim 77% Gaming Gains [2026]
- AMD RX 9070 GRE Ends China Exclusivity at $549 [2026]
- RAM Shortage 2026: AI Chips Hit 5 Gaming Platforms’ Prices [2026]
- How to Run an LLM Locally: 13 Steps, 90 Min [2026]
- AMD Advancing AI: Helios Hits $5.25M, 12GW Booked [2026]
