Nvidia's RTX Spark Laptops Land at $7,000 During a Memory Panic. The Chip Is the Least Interesting Part.

Nvidia's RTX Spark laptops went on sale this week. On the benchmark numbers, they are the fastest Windows notebooks you can buy for running AI models locally. They are also $2,599 at the bottom of the range and $6,999 at the top, and they landed in the same seven trading sessions in which investors stripped roughly $2 trillion of market value out of semiconductor stocks.

Both facts are on the record. The thing worth understanding is what connects them: the price on these machines is no longer set by the chip inside them.

What shipped, and what it costs

RTX Spark is Nvidia's first processor built for Windows PCs. The N1X part pairs up to 20 Grace Arm CPU cores, supplied via MediaTek, with a Blackwell GPU of up to 6,144 CUDA cores and up to 128GB of unified memory shared between them, per The Register's teardown. A cheaper bin with 18 CPU cores and 5,120 CUDA cores exists alongside it.

The launch roster covers Microsoft's Surface Laptop Ultra, HP's OmniBook Ultra 16 and OmniBook X, Asus's ProArt P14 and P16, Lenovo's Yoga 9n and Pro 9n, Dell's XPS 16 Creator Edition and MSI's entry. The floor is $2,599.99 for 18 cores, 24GB of memory and 512GB of storage. Microsoft's top Surface Laptop Ultra is $5,899.99 with 128GB. Asus's maxed ProArt P16 is $6,999.99 with 128GB and 2TB. The Register's review unit, an HP OmniBook Ultra 16 with the full 20-core part, came in at $6,229 against a roughly $3,200 base.

The chip underneath is a year old

This is the detail that got buried under the launch coverage. The N1X is based heavily on the GB10, the chip Nvidia put in its Linux-based DGX Spark more than a year ago. It is not new silicon. The headline improvement is memory bandwidth, lifted from the DGX Spark's 273 GB/s to around 300 GB/s, roughly a ten percent gain.

That matters because of how these chips actually run models. Token generation is memory bound. How fast a local model answers you is governed by memory bandwidth divided by how much of the model gets read for each token. Extra cores speed up prompt processing. They do not buy the one thing that makes a local agent feel usable.

On raw CPU, the N1X does not beat Apple, either. The Verge's benchmark table puts its Geekbench 7 single-core score around 2,700 against 3,693 for Apple's M5 Max, and its Cinebench 2026 multi-core result at 6,638 against 8,952. Antonio Di Benedetto's summary is blunt: a 12-core M6 Mac Mini competes with Nvidia's 20-core part in multicore, and that Mac Mini starts at $900.

Where it does win is graphics and AI throughput in a chassis you can carry, and it wins clearly. Computerworld's Dave Altavilla measured the Surface Laptop Ultra at 2,703 in UL Procyon's AI text generation test against 2,027 for a Lenovo Intel Core Ultra X9 machine, about 33 percent ahead, and at 26.97 images per minute on MLPerf's Flux 2 Klein 4B image workload against 9.96, roughly 2.7 times the throughput. Ars Technica's Nick Indge ran a 43GB Llama 3.3 70B model on the GPU at 6.26 tokens per second, which is not fast, but it is a model most laptops cannot load at all.

The price is set by memory, not silicon

Now the part that decides whether this category survives.

Nvidia raised the price of its 128GB DGX Spark, built on the same GB10 silicon, to $6,950 on Friday 2 October, an increase of nearly 75 percent from a year earlier, per The Register. The 64GB variant it launched alongside sits at around $4,999, which is still more than the 128GB model cost at launch. Project Digits, the concept this all descends from, was briefed at around $3,000.

That is not margin. That is memory. TrendForce raised its first-quarter forecast for conventional DRAM contract prices to a 90 to 95 percent quarter-on-quarter increase, up from an earlier estimate of 55 to 60 percent, and said PC DRAM prices would more than double. They landed at 93 to 98 percent, and TrendForce then projected another 58 to 63 percent rise in the second quarter. Reuters carried the February revision when it landed.

The mechanism will not self-correct. Samsung, SK hynix and Micron have moved the bulk of their wafer capacity toward high-bandwidth memory for AI accelerators, where margins are far higher. That leaves less capacity for the LPDDR5X that goes into a notebook. The Register noted that each of Nvidia's NVL72 racks carries between 36 and 54 terabytes of the stuff, and that is CPU memory alone. Every rack shipped is memory that does not go into a laptop.

HP told investors in February that memory had reached roughly 35 percent of the cost of building a PC, up from 15 to 18 percent a quarter earlier, per Tom's Hardware. New fabs from the three big suppliers are under construction, and none reaches volume production before late 2027 at the earliest. The Register reports the memory makers expect DRAM supply to stay constrained through at least 2028.

The week it landed in

You can read the sticker prices as a product decision. The market read them as something else.

The Philadelphia Semiconductor Index fell 3.8 percent on Friday to close an eighth consecutive down session, its longest losing streak since an eight-session decline in October 2018 by Investor's Business Daily's count. Barron's and MarketWatch compared the run to the eight-day slide of October 2008. The index shed more than 12 percent over the previous seven sessions, wiping out roughly $2 trillion in market value on TradingView data, its fifth correction of at least 10 percent this year. Short sellers booked an estimated $32 billion in paper gains on the way down, according to S3 Partners data reported by Bloomberg.

Then there was CoreWeave. Bonds the company issued in June at a 9.675 percent yield were trading to yield around 15 percent this week, past the level that normally signals distress. Its credit default swap spreads hit roughly 1,000 basis points, implying a near 63 percent chance of default within five years, up from about 30 percent a year ago, per Reuters. Max Frumes of 9fin told Reuters that CoreWeave is "absolutely the canary in the coal mine" for AI-linked credit. CEO Michael Intrator told staff the market sometimes "misunderstands and acts irrationally" and that the company's backlog more than covers its debt obligations.

The piece that matters more for hardware came on 14 October, when Barclays downgraded Dell, HPE and NetApp to equal weight from overweight, writing that it does not "see a meaningful AI inflection yet" in enterprise spending. Dell fell about 17 percent across two sessions. The analysts' core argument is the one worth keeping: part of the recent revenue strength at hardware makers reflects higher component prices rather than more units shipped. When memory costs rise, revenue rises without a single extra customer.

Add the demand side. Ars Technica cites industry figures showing total PC shipments fell 20 percent year over year in the third quarter of 2026.

What the reviewers actually found

Strip out the price framing and the hardware mostly delivers, with caveats that read as structural rather than early.

The Register found the OmniBook Ultra 16 would leave you "hard pressed to find a faster Windows notebook for local AI," then recorded GPU temperatures climbing past 90 degrees Celsius over about an hour of sustained agentic load, producing occasional lockups and one instance of graphical artifacts its reviewer attributed to overheating memory. Its closing line on whether these systems save you money: "So far that argument isn't looking good."

The Verge put three writers on five machines for a week. Tom Warren's verdict on local AI was "not yet," describing it as "a beta aimed at developers." Sean Hollister asked Hermes running Qwen3.8-27B for a news briefing. It took ten minutes and 10 percent of a 99.9 watt-hour battery to produce a list that Claude Opus 5.5 produced in half the time and better. Di Benedetto burned 30 percent of a Lenovo Yoga 9n's battery in 36 minutes resizing five product images, which the model then cropped wrong. Ars found Microsoft's own AI Dev Gallery crashed twice during testing, and that Copilot+ features will not fully arrive until late 2026 or early 2027.

So the hardware is real and the software is not finished, which is roughly what you would expect from a platform launch in any year. The unusual part is that the software gaps are not what limits this product. The price is.

What this means if you build hardware

This launch gets filed as a chip story. It is a bill-of-materials story wearing a chip story's clothes.

Three things are now clear. The differentiator in this class is memory bandwidth and memory capacity, not core counts. The price of that memory is set by an allocation fight with AI datacentres that no OEM wins by negotiating harder. And the supply outlook runs past the life of the current product generation, which means any roadmap that assumes memory costs normalise in the next two years has a hole in it.

That reframes the questions worth asking. What is the memory floor your product needs to do its job, and can you design just under it rather than up to the spec sheet? Which tiers could ship at a lower capacity without breaking the use case? Nvidia's own guidance for the 64GB systems points at 26 to 35 billion parameter models as good enough for local agents, not the largest model that fits in the box. And what does a two-year bill of materials look like when LPDDR5X contracts move 60 percent in a quarter?

The supply chain complexity behind shipping a compute product through a memory crunch is the kind of problem that never shows up in a launch keynote. At DMC we work with hardware companies on sourcing strategy, cost modelling and production ramp planning when component markets move faster than the product cycle. Working out how to spec, price and build through a three-year memory squeeze? let's talk.