RTX Spark superchip laptops go on sale this autumn, and they are not another spin on the Copilot+ badge. Nvidia has taken the Grace CPU and Blackwell GPU pairing it built for desk-side AI boxes, shrunk it into a chassis as thin as 14mm, and persuaded six of the largest PC makers in the world to build around it.
The announcement landed on 31 May 2026 at Computex in Taipei, and the language was deliberately grand. “For forty years, you launched apps. Click. Type,” Jensen Huang said in the Nvidia press release. “With RTX Spark and Microsoft Windows, you ask — the PC does the work.” Satya Nadella’s line was shorter: “unmetered intelligence to every home and every desk.” Those are claims about AI agents replacing the app model, not about frame rates.
We have already looked at one machine built on this silicon — Acer’s small-form-factor desktop concept from IFA — but that was industrial design, not a shipping product. This article is about the RTX Spark superchip itself and the wave of laptops it is about to arrive in: what the architecture actually does differently, what the tipped prices are, and what a leaked prototype revealed weeks before launch.
Table of contents
- What the RTX Spark Superchip Actually Is
- Why the RTX Spark Superchip Is Not Just Another Copilot+ PC
- Every RTX Spark Superchip Laptop Announced So Far
- What an RTX Spark Superchip Laptop Is Likely to Cost
- The Leaked RTX Spark Superchip Benchmarks
- The Windows on Arm Problem the RTX Spark Superchip Inherits
- What Windows Does With the RTX Spark Superchip
- Who Should Buy an RTX Spark Superchip Machine
- What the RTX Spark Superchip Means for Business IT
- RTX Spark Superchip Questions People Are Asking
- References and Further Reading
What the RTX Spark Superchip Actually Is
The word “superchip” is Nvidia’s, and it describes a packaging decision rather than a marketing tier. Two processors that would normally sit apart are fused onto one RTX Spark superchip module and given a single pool of memory.
A Grace CPU and a Blackwell GPU on one package
The laptop configuration pairs a 20-core Nvidia Grace CPU, built on Arm and co-designed with MediaTek, with a Blackwell RTX GPU carrying 6,144 CUDA cores. They are joined by NVLink-C2C, a chip-to-chip interconnect rated at roughly 600 GB/s — an order of magnitude beyond what a PCIe link between a separate CPU and graphics card would carry.
One pool of memory instead of two
This is the part that matters most. A conventional gaming laptop has system RAM and a walled-off block of video memory, and moving data between them costs time. Every RTX Spark superchip instead exposes up to 128 GB of LPDDR5X as unified memory that both processors address directly, at around 300 GB/s. There is no copy step, and no 16 GB VRAM ceiling.
Where the petaflop number comes from
The headline “1 petaflop of AI performance” is measured at FP4 precision on fifth-generation Tensor Cores. That is a legitimate number for the inference work the RTX Spark superchip is sold for, and a meaningless one for anything else. Read it as a throughput figure for running quantised models, not as a general measure of speed.
The specification, laptop versus desktop
| Specification | Laptop | Desktop (N1X) |
|---|---|---|
| Grace CPU cores | 20 | 18 |
| Blackwell RTX GPU cores | 6,144 | 5,120 |
| Unified memory (max) | 128 GB LPDDR5X | 64 GB LPDDR5X |
| Peak AI throughput | 1 petaflop FP4 | 1 petaflop FP4 |
| TDP | 45-80W | 140W |
| Tensor / RT cores | 5th gen / 4th gen | 5th gen / 4th gen |
| Graphics stack | DLSS 5, Reflex 2, CUDA | DLSS 5, Reflex 2, CUDA |
Why the RTX Spark Superchip Is Not Just Another Copilot+ PC
Copilot+ certification requires a neural processing unit above a stated threshold, and the RTX Spark superchip carries a dedicated NPU to earn that badge. But the NPU is the least interesting silicon on the RTX Spark superchip package.
The ceiling an NPU runs into
An NPU is tuned for small, always-on models: background blur, live captions, local search indexing. It is efficient and it is limited. Ask it to hold a 70-billion-parameter model in memory and the question stops being about the processor and starts being about how much RAM the RTX Spark superchip has.
What 128 GB of RTX Spark superchip memory unlocks
Nvidia’s claim is that an RTX Spark superchip can run a 120-billion-parameter model with a one-million-token context window locally. Whatever the real-world speed turns out to be, the capability is a memory-capacity story, and no NPU-first design shipping today has anything close to the same headroom.
The CUDA argument
The other differentiator is software the company did not have to build for this launch. CUDA has fifteen years of tooling, and llama.cpp founder Georgi Gerganov, ComfyUI creator Yannik Marek and Adobe all appeared in the launch materials. Adobe says it is rearchitecting Photoshop and Premiere for roughly 2x faster AI and graphics performance on the RTX Spark superchip.
| Approach | Memory for models | Local AI ceiling | Ecosystem |
|---|---|---|---|
| RTX Spark | Up to 128 GB unified | ~120B parameters | CUDA, mature |
| NPU-first Copilot+ | Shared system RAM | Small on-device models | Windows ML |
| Apple M-series | Up to 128 GB unified | Large, proven | Metal, no CUDA |
| AMD Strix Halo | Up to 128 GB unified | Large | ROCm, x86 native |
| Discrete RTX 5070 laptop | Separate VRAM block | VRAM-capped | CUDA, mature |
Every RTX Spark Superchip Laptop Announced So Far
Six vendors are in the first wave: ASUS, Dell, HP, Lenovo, Microsoft Surface and MSI. Acer and GIGABYTE have been named as following later, which is what the IFA concept desktop was previewing.
The six RTX Spark superchip launch partners
Nvidia’s own framing is that these are the world’s first Windows PCs purpose-built for personal agents. In practice they are premium creator laptops between 14 and 16 inches, in machined aluminium, with tandem OLED or mini-LED panels and G-SYNC.
Microsoft’s own hardware
Microsoft announced two devices: the Surface Laptop Ultra and a Surface RTX Spark Dev Box. The Laptop Ultra is a 15-inch mini-LED machine at 2,000 nits and 262 PPI, under 18mm thick and under 4.5 lb, with what Microsoft describes as 2.5x the thermal capacity of the seventh-edition Surface Laptop.
| Model | Display | Max memory | Notable |
|---|---|---|---|
| Surface Laptop Ultra | 15in mini-LED, 2,000 nits | 128 GB | Under 4.5 lb |
| ASUS ProArt P16 | 16in 4K OLED 120Hz | 128 GB | 1.77 kg, 99.9Wh |
| ASUS ProArt P14 | 14in 3K OLED 120Hz | 128 GB | 1.48 kg, 90Wh |
| Dell XPS 16 Creator | 16in tandem OLED | 128 GB | True Black HDR 600 |
| HP OmniBook Ultra 16 | 16in | Not stated | 15.73mm rear height |
| Lenovo Yoga Pro 9n | 15in OLED | Not stated | Haptic touchpad |
| MSI Prestige N16 Flip | 16in UHD+ tandem OLED | Not stated | 2-in-1, 99.9Wh |
When they actually ship
Nvidia said “fall 2026” in May and has not sharpened it since. TechRadar reported in the run-up that machines are expected in October. As of early September there is still no firm on-sale date and no published price for any configuration from any vendor.
What an RTX Spark Superchip Laptop Is Likely to Cost
No manufacturer has named a price. Everything below is a tipped figure or an analyst projection, and should be read as a range rather than a number.
The tipped bands
A Morgan Stanley note reported in early June put entry-level N1 machines from around $1,799 and N1X-equipped models at a floor of about $2,899 — and those base configurations carry 16 or 32 GB of memory and a 512 GB or 1 TB SSD, not the 128 GB headline. Analysts have projected the Surface Laptop Ultra between roughly $2,800 and $4,500.
How that lands against a MacBook Pro
For context, a 128 GB MacBook Pro with an M4 Max sits near $5,099. If the tipped numbers hold, a maximum-memory RTX Spark superchip machine competes directly with Apple’s flagship rather than undercutting it.
The memory tier is the real price lever
Because memory is unified and soldered, the configuration you buy is the configuration you keep. Nvidia has also said cheaper, lower-memory variants will follow after launch, which means the first machines on shelves are unlikely to be the cheap ones.
The Leaked RTX Spark Superchip Benchmarks
In late July a TechPowerUp forum user going by Fouquin said they had spent a month with a pre-release Surface Laptop Ultra and published numbers. It is the only independent data anyone has.
What the RTX Spark superchip prototype scored
The RTX Spark superchip unit carried the full N1X — 6,144 CUDA cores, the 20-core Arm CPU — with 24 GB of unified memory and a 512 GB SSD. Cinebench 2024 multi-thread returned 1386, behind a MacBook Pro with an M4 Pro. In 3DMark, Port Royal managed 31 fps and Steel Nomad at 4K about 22 fps, which puts it nearer an RTX 3060 Ti than the RTX 5070 Nvidia has been comparing it to. Phoronix AI tests landed in desktop RTX 4070 territory.
Why one prototype is not a verdict
A single pre-release unit tested by an enthusiast is not a review, and it is worth saying so plainly. The sample was one machine, on one driver build, weeks before launch, with 24 GB of memory rather than the 128 GB the marketing leads with. It is evidence about that specific RTX Spark superchip prototype and nothing more. What makes it useful is simply that it exists: no reviewer has been given a retail unit, so the alternative to imperfect data is no data at all.
The genuinely good result
One number stood out for the right reasons: performance was identical on battery and plugged in. On a thin creator laptop that is unusual, and it is the clearest evidence so far that the efficiency claims are real.
What was broken
The list is long. Phoronix CUDA tests failed outright. G-SYNC misbehaved, local dimming zones malfunctioned, touch input was noisy, power profiles did not work, and gaming showed stutter with unstable GPU clocks. Idle draw jumped from 7-8W to 12-23W after a driver update — the wrong direction.
How much of it is fixable
Most of it, probably. Driver instability and broken power profiles on an unreleased machine are ordinary. The Cinebench result is harder to wave away, because it points at CPU design rather than software maturity, and IEEE Spectrum’s reporting already noted that Grace cores lag competitors on raw speed.
The Windows on Arm Problem the RTX Spark Superchip Inherits
This is an Arm platform, and every Arm Windows launch since 2018 has run into the same wall.
Prism and the app estate
Native Arm builds run at full speed. Everything else goes through Microsoft’s Prism emulator, and the leaked prototype needed Prism for most Windows games. Microsoft says it has improved Prism for this generation, but emulation overhead is a tax that x86 machines simply do not pay.
Why this attempt has better odds
Nvidia brings weight Qualcomm never had. Ryan Shrout of Signal65 put it as “more clout and more industry weight” for getting developers to do the work, and Nvidia controls more than 90% of the discrete GPU market with famously mature drivers. Xbox on PC, NetEase and Remedy were all named at launch.
The dissenting read
Not everyone is convinced the AI framing is the point. Anshel Sag of Moor Insights & Strategy argued the machines will sell to creators and gamers, and that the AI positioning is “mostly to appease investors.” That is worth holding onto when reading the petaflop number.
What Windows Does With the RTX Spark Superchip
Microsoft’s contribution is the part that gets least attention and may matter most, because it is the difference between a fast laptop and a machine that can safely run autonomous software.
OpenShell and containment
Nvidia’s OpenShell runtime executes agents inside new Windows security primitives covering identity, containment, policy and end-to-end protection. The point is that an agent with local file access is a different risk class from a chatbot, and the operating system now has somewhere to put it.
Routing local or cloud
Windows adds policy controls that decide whether a query is answered by an on-device model or sent to a cloud one, with personal information masked before anything leaves the RTX Spark superchip. For regulated work, that switch is the whole argument for buying local compute rather than renting it. Windows ML also exposes TensorRT natively, so a developer targeting local inference or computer vision workloads is not writing against a proprietary shim.
The agent surface
Agents are reachable from the Windows taskbar. Jeff Fisher, an Nvidia senior vice president, framed the partnership as a shared view that “agents are the future of personal computing” — which is a product bet, not a specification.
Who Should Buy an RTX Spark Superchip Machine
The honest answer in September 2026 is that almost nobody should buy one yet, because nobody can: there is no price and no date. But the shape of the decision is already clear.
The clear yes
If you run large models locally today and keep hitting a VRAM wall, this is the first Windows laptop that removes that wall. The same is true for 12K video work, 90 GB 3D scenes and anyone whose CUDA tooling makes an Apple machine a non-starter.
The clear no
If your workload is Office, browsers and video calls, an RTX Spark superchip is an expensive answer to a question you do not have. If you depend on niche x86 software, drivers or peripherals, Arm emulation is a real risk rather than a theoretical one.
The wait-and-see middle
Everyone else should wait for reviewed silicon at a published price. Nothing in the leak suggests the RTX Spark superchip is bad; it suggests it was unfinished in July.
| If you are | Verdict | Why |
|---|---|---|
| Running local LLMs | Strong yes | 128 GB unified memory |
| Video or 3D creator | Yes, after reviews | 12K and 90 GB scene claims |
| Competitive gamer | Wait | Emulation and clock stability |
| Standard office user | No | Cost with no matching workload |
| Tied to x86 software | No | Prism overhead and edge cases |
What the RTX Spark Superchip Means for Business IT
For anyone buying fleets rather than a single machine, this is a procurement question before it is a performance one.
RTX Spark superchip questions to ask a vendor
Ask which memory tier is actually available at your budget, because the 128 GB figure anchors expectations that a 32 GB order will not meet. Ask for the emulation position on every line-of-business application you run. Ask whether the warranty and imaging process differ from the x86 fleet, because a soldered unified memory pool cannot be upgraded later.
The security and management angle
The containment primitives are genuinely new, and they are also unproven. Treat a machine running local AI agents with file access as a new endpoint class, not as a laptop refresh. Our write-up on Nvidia’s MediaTek partnership covers where the CPU design came from, and the Perplexity local-agent portable shows the same zero-token-cost argument in a different chassis.
The refresh-cycle question
There is also a timing argument. Most fleets replace laptops on a three- to five-year cycle, and a first generation RTX Spark superchip bought in October 2026 will be judged against a second generation that arrives well inside that window. Buying into a new architecture at launch means paying the early-adopter premium for hardware whose software stack is still settling. Unless a specific workload is blocked today, the disciplined move is to pilot a handful of units, measure them against your actual applications, and hold the fleet decision until the drivers and the prices have both stopped moving.
The cheaper alternative nobody mentions
Not every organisation needs local inference on every desk. The JioPC approach to ageing hardware is the opposite bet — keep the endpoint thin and put the intelligence elsewhere — and for most fleets it is still the cheaper answer. If you are weighing either path, our AI strategy and managed IT services teams can model the difference.
RTX Spark Superchip Questions People Are Asking
Is the RTX Spark superchip the same as DGX Spark?
They share the Blackwell GB10 lineage, but DGX Spark is a desk-side developer box and the RTX Spark superchip is the Windows PC implementation of that design, with a consumer graphics stack and a Copilot+ NPU attached.
Can it really run a 120-billion-parameter model?
Nvidia says yes, with a one-million-token context, and the 128 GB unified memory pool makes it arithmetically plausible. No independent test has confirmed the speed at which it does so.
Will my existing Windows software run?
Native Arm applications will. Everything else runs under Prism emulation, which works but costs performance, and the leaked prototype needed it for most games.
Is the memory upgradeable?
No. Unified memory is soldered to the RTX Spark superchip package. The tier chosen at purchase is permanent, which makes the configuration decision unusually consequential.
When can I actually buy one?
Nvidia has committed only to autumn 2026, with October reported as the expected month. No vendor has opened orders or published a price as of early September 2026.
References and Further Reading
NVIDIA and Microsoft Reinvent Windows PCs for the Age of Personal AI
Introducing a powerful new chapter for Windows PCs, accelerated by NVIDIA RTX Spark
RTX Spark Brings Nvidia AI Muscle to Windows PCs
Nvidia’s RTX Spark laptops are coming in October, but price information is still MIA
Nvidia RTX Spark laptop prices tipped
Microsoft laptop with Nvidia RTX Spark leaked and benchmarked before launch
Tech enthusiast delivers first RTX Spark Surface Laptop Ultra review
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.