In my DGX Spark versus Strix Halo comparison I told you to wait for October. October is close enough to see now, NVIDIA has confirmed the configurations, and I owe you the follow-up.
The short version: everything is official except the price, and the reason it isn't official is the most interesting thing in this story.
What's Actually Confirmed
NVIDIA detailed two RTX Spark N1X configurations at IFA in early September. Both are Arm — Arm v9.2, ten Cortex-X925 performance cores and ten Cortex-A725 efficiency cores on the larger part — running Windows 11, with the CPU co-designed with MediaTek and linked to the GPU over NVLink-C2C.
| 20-core | 18-core | |
|---|---|---|
| CPU | 20-core Grace | 18-core Grace |
| GPU | 6,144 Blackwell CUDA cores | 5,120 Blackwell CUDA cores |
| Memory | 24–128GB unified LPDDR5X | 24 or 32GB (see below) |
| Bandwidth | 300GB/s, 256-bit | not documented |
| AI compute | ~1 petaflop FP4 | — |
| TDP | 45–80W laptop / 140W desktop | 45–80W |
| Form factor | Laptops and compact desktops | Laptops at launch |
Eight OEMs are in: Asus, Dell, HP, Lenovo, Microsoft, MSI at launch, with Acer and Gigabyte following. Named machines so far include the Dell XPS 16 Creator Edition, a Microsoft Surface Laptop Ultra, the Lenovo Yoga 9n and an Acer small-form-factor desktop. Laptops get as thin as 14mm and as light as about three pounds.
One spec I can't fully resolve
Tom's Hardware, Windows Central and implicator.ai all independently reported from IFA that the 18-core part tops out at 24 or 32GB — not 128GB. NVIDIA's live product page appears to list "up to 128GB" against both laptop columns. It's a tabbed configurator that doesn't render cleanly, so this may be a reading artifact rather than a contradiction. Three reputable outlets landed on the same number independently, so I'd trust 24–32GB on the smaller part until NVIDIA says otherwise — but if 128GB matters to you, wait for a spec sheet you can hold.
The 600GB/s Number Is Wrong
You will see RTX Spark's memory bandwidth reported as 600GB/s. Multiple outlets have run it. It's wrong, and since bandwidth is the number that determines how fast a local model actually talks, it's worth being precise.
The confirmed figure is 300GB/s on a 256-bit bus, from NVIDIA's own Windows on Arm porting guide, which lists "Up to 128 GB of LPDDR5, shared by the CPU and iGPU" at a bandwidth of "300 GB/s."
The 600GB/s figure traces back to an NVIDIA slide reading "600GB/s GPU to CPU." That's the NVLink-C2C interconnect between the CPU and GPU dies — how fast the two halves of the chip talk to each other — not how fast either reaches system memory. The same relationship holds on the DGX Spark: 600GB/s chip-to-chip, 273GB/s to DRAM. Two different numbers measuring two different things, and the bigger one keeps getting printed.
If you read the local LLM stack piece, you know why this matters: token generation is bandwidth-bound. Compute decides how fast the machine reads your prompt; bandwidth decides how fast it writes the answer. A petaflop of FP4 doesn't help the second one.
So Where Does 300GB/s Land It?
| Max memory | Bandwidth | Price | |
|---|---|---|---|
| RTX Spark N1X | 128GB | 300GB/s | unannounced |
| DGX Spark (GB10) | 128GB | 273GB/s | $4,699 |
| Ryzen AI Max+ 395 | 128GB | 256GB/s | $3,449–$4,349 |
| Apple M5 Max | 128GB | 614GB/s | from $2,499 |
| Apple M5 Ultra | 512GB | 1.2TB/s | from $5,499 |
That's a 10% bandwidth improvement over the DGX Spark, about 17% over Strix Halo, and roughly half an M5 Max. For pure token generation, Apple still wins the bandwidth argument comfortably — as it has all year.
The genuinely notable thing is what N1X is: near-identical silicon to the DGX Spark. Same CPU topology, same 6,144 CUDA cores, same petaflop of FP4, same 140W desktop envelope, same 256-bit bus. The differences are 300 versus 273GB/s, and Windows 11 instead of DGX OS. NVIDIA is putting a DGX Spark-class part into consumer laptops you can buy at retail, which is a real shift regardless of what it costs.
The Real Story Is a Memory Shortage
Here's the part that explains the blank price tag, and it's bigger than one product.
When I wrote the DGX-versus-Strix comparison, I quoted Strix Halo boxes at $3,300 to $4,349 and noted the DGX Spark had gone from $3,999 to $4,699 — an 18% increase NVIDIA attributed explicitly to memory supply constraints.
That wasn't an NVIDIA problem. It was the category. As of this week, a comparison of 128GB Strix Halo machines notes that every price has roughly doubled since June, driven by DRAM contract pricing: the 128GB configuration that anchored this category near $2,000 now runs $3,449 from Framework, $3,649 from GMKtec, and $4,349 from Beelink.
Read that against the specs. These machines are, essentially, a large pile of LPDDR5X with a processor attached. The entire product category — DGX Spark, Strix Halo boxes, RTX Spark, Apple's high-memory configurations — is a bet that unified memory is the thing worth buying. When memory contract prices double, the bet gets expensive all at once.
Which is why nobody will name a price
Lenovo announced the Yoga 9n at IFA and deferred pricing while pricing other products the same day. Microsoft's Andrew Hill: "We're not ready to talk about pricing yet until we get closer to availability." A month out from launch, with a memory market that repriced an already-shipping product upward by 18%, that silence reads less like marketing strategy and more like nobody knowing what their bill of materials costs in October. That's my read, not reporting — but it fits the evidence better than anything else.
For what it's worth, the one number circulating is a Morgan Stanley estimate relayed secondhand: N1X systems needing to price at about $2,899, with the smaller N1 around $1,799. No methodology beyond "checks with PC brands." Treat it as a rumor with a good haircut.
Three Things To Know Before You Want One
It's Windows on Arm, and NVIDIA documents the rough edges itself. Its porting guide warns that interlocked atomics "may incur heavy penalties when memory address is not naturally aligned," that x87 floating point may not return bit-accurate results under emulation, and that "instruction timings may vary under emulation and even from run to run." NVIDIA also conceded to press that CPU-intensive games could "suffer larger drops." Mainstream apps are largely fine on Arm in 2026; specialist tools are where it still bites, and I could find no authoritative answer on CUDA, PyTorch wheels, Docker or WSL2 status on this platform. For an AI development machine that's not a small gap.
You can't use all 128GB for a model. NVIDIA's unified memory documentation splits memory into a GPU carveout, a shared system pool and CPU-only memory, and states the shared pool is "clamped to a minimum of 50% and a maximum of 80%" of post-carveout capacity. So the marketing line about running 120-billion-parameter models with a million tokens of context is doing work against a number smaller than 128. How much smaller is configuration-dependent and nobody has tested it.
No independent benchmarks exist. At IFA, reporters weren't allowed to run them — one noted some demo units were simply powered off. The only data points are leaked engineering samples: a Geekbench entry at 2,593 single-core and 23,424 multi-core, and a Cinebench 2026 run at 540 single and 5,771 multi from a "Surface Laptop Ultra Pre-Sample" reporting a 1.01GHz clock. Multi-core looks strong; single-core looks weak next to Apple's M5 and Snapdragon X2 Elite, which complicates the "fastest Windows laptop" framing. And there is not one published LLM tokens-per-second figure for this chip anywhere.
What I'd Do
Still waiting, and now for a better reason
Don't buy anything at 128GB this quarter if you can avoid it. The category repriced upward on a DRAM shortage, not on improved products. You'd be paying a memory-market premium for silicon that hasn't changed.
If you need local AI now, the advice from the earlier comparison holds and has gotten stronger: a used Mac with high unified memory is the value play, and Apple's bandwidth advantage is still the largest single number on the board.
When N1X reviews land in October, the two numbers to check are sustained memory bandwidth under load — not the spec sheet — and real tokens per second on a mixture-of-experts model. Everything else is marketing. I'll run that piece when there's something real to run it on.
The thing worth holding onto: NVIDIA just put DGX Spark silicon into consumer laptops from eight manufacturers. In a normal market that's a story about local AI going mainstream. In this market it's a story about how much a pile of memory costs — and until someone prints a price, that's the only chapter we have.
Mentioned in this article
Every machine mentioned here, with today's price and measured throughput, is in the local AI hardware table.