RTX Spark Laptops for Local AI: Which Config Is Worth It, and Which Is a Trap

Short answer: buy an RTX Spark laptop only in the 6,144-core, 128GB configuration, and only if the machine has to be a Windows CUDA laptop for a model that does not fit a 27B-class 4-bit pool. The 24GB and 32GB configs are the trap. StorageReview (Brian Beeler, 7 October 2026) says those pools leave room for 27B-class models at 4-bit and not much above, and that the large-model lines in the launch material assume the 128GB SKU. On the Surface Laptop Ultra, Windows Central the same day lists the baseline at $2,599 with 24GB, a 32GB step on the larger chip at $3,699.99, and a top at $5,899.99 with 128GB and a 1TB SSD. First shipments for that laptop start 16 October 2026. Tom’s Hardware (Jeffrey Kampman, 15 October 2026) says lead times for the Surface it reviewed, listed at $5,899 with 1TB, are already out to late December.

Preorders opened 7 October 2026. The embargo on independent reviews lifted 15 October 2026, the day before first shipments. The speed numbers below are community-cited from those reviews. LocalRig has not run an RTX Spark laptop. On the 128GB Surface, Windows Central’s review the same day says the default GPU allocation is 64GB, and that it can be raised when a model does not fit.

Which laptop is actually on preorder, and when does it ship?

The name “RTX Spark” covers two laptop chips and a wide price floor. The ship date is not the same for every OEM.

NVIDIA’s blog on 7 October 2026 says laptop preorders opened that day and the laptops are available 16 October, with compact desktops on sale in November. The same post says systems are coming from Acer, ASUS, Dell, HP, Lenovo, Microsoft, MSI, and Gigabyte. It does not publish a price table. Prices below are journalism, checked by LocalRig against those articles on 9 October 2026. LocalRig did not open Microsoft, Dell, HP, ASUS, Lenovo, or MSI carts.

Machine	Price as reported 2026-10-07	Ship window in that report	Memory tied to the price
Surface Laptop Ultra, baseline	$2,599 (config article). The same-day roundup says prices start at $2,599.99	16 October 2026	24GB, 512GB SSD, 18-core CPU, 5,120-core GPU
Surface Laptop Ultra, larger chip, lowest RAM	$3,699.99	16 October 2026	32GB (lowest RAM on this chip), 1TB (lowest SSD on this chip), 20-core CPU, 6,144-core GPU
Surface Laptop Ultra, top named	up to $5,899.99	16 October 2026	128GB, 1TB, the larger Spark
ASUS ProArt P14	from $2,599.99	16 October 2026	Not stated
ASUS ProArt P16	from $2,799.99	16 October 2026	Not stated
Dell XPS 16 Creator Edition	from $3,799.99	16 October 2026	Windows Central does not state RAM. StorageReview’s related headline the same day describes this XPS as 128GB at $3,799.99. This page did not fetch that Dell article body
HP OmniBook X 14	from $2,999.99	4 November 2026	Not stated
HP OmniBook Ultra 16	from $3,199.99	4 November 2026	Not stated
Lenovo Yoga 9n 2-in-1	from $3,099.99	16 October 2026	Not stated
MSI Prestige N16 Flip AI+	from $3,299.99	Preorders “might” start 8 October. No ship date in the roundup	Not stated
Surface RTX Spark Dev Box (desktop, not a laptop)	$5,999, 128GB, Microsoft.com US only	November 2026	128GB. No ConnectX-7 in Microsoft’s port list, per StorageReview

A $4,999.99 HP ceiling was not on the Windows Central roundup fetched 9 October 2026. This page does not use it. Acer and Gigabyte are on NVIDIA’s vendor list and have no price in the four sources above.

Tom’s Hardware on 15 October 2026 says the Surface Laptop Ultra it reviewed lists for $5,899 with 1TB of storage, and that lead times for that system are already out to late December. That is a review-dated availability note for the top Surface, community-cited, not a new cart check of the other rows. The 16 October ship window in the table is the 7 October report. priceDate on this page stays 2026-10-09.

Windows Central’s config article is the only place in this set that ties a Surface dollar amount to a memory size. The roundup’s “from” prices for ASUS, Dell, HP, Lenovo, and MSI are floors. A low floor can still hide a 24GB or 32GB pool. Read the RAM line on the cart before you treat any of those floors as the 128GB machine.

Surface Laptop Ultra search on Amazon →

That link is an affiliate search (tag localrig-20). The prices in the table are not Amazon prices.

What is the difference between the two chips?

128GB exists only on the larger chip. The cheaper Surface cannot be configured into the keynote machine.

StorageReview’s reading of NVIDIA’s RTX Spark spec table, published 7 October 2026:

	Laptop, larger	Laptop, smaller	Desktop N1X	GB10 DGX Spark
GPU	6,144-core Blackwell RTX	5,120-core Blackwell RTX	6,144-core	Blackwell
CPU	20-core Grace	18-core Grace	20-core Grace	20 Arm cores
TDP	45–80W	45–80W	140W	140W
Unified memory	Up to 128GB LPDDR5X	Up to 64GB LPDDR5X	Up to 128GB LPDDR5X	64GB or 128GB, 273 GB/s
OS	Windows 11	Windows 11	Windows 11	DGX OS
High-speed cluster NIC	Not in the laptop columns	Not in the laptop columns	Not in the desktop N1X column	ConnectX-7

StorageReview also says that of 15 marketplace laptop configs they read, seven use the 5,120-core chip, and the memory sizes they saw were 24GB, 32GB, 48GB, and 128GB. The smaller chip’s ceiling is 64GB. A 64GB laptop SKU was not in that list of sizes. The $2,599 Surface is the smaller chip at 24GB. Ordering that unit does not put you on a path to 128GB. The sources do not say the RAM is socketed.

The larger chip’s Surface ladder starts at 32GB / $3,699.99, not at 128GB. You can order that chip with 128GB. The 32GB computer you receive is still a 32GB computer.

Laptop power is a separate cut. StorageReview says a laptop N1X at 45–80W runs at 32% to 57% of the 140W desktop rating, while the family is still sold with the same “up to 1 petaflop” of FP4. Microsoft’s Surface footnotes, as reported there, call that petaflop a theoretical FP4 number using sparsity. Do not read it as a tokens-per-second result, and do not paste a desktop figure onto a 45W laptop.

Windows Central’s 15 October 2026 spec table lists configurable SoC TDP as 45–80W on the 5,120-core part and 45–140W on the 6,144-core part. That 140W top conflicts with StorageReview’s laptop column. Tom’s Hardware the same day describes the tested Surface’s power adapter as 140W, against a 96W charger on the 14-inch MacBook Pro. Adapter wattage is not a measured SoC TDP. This page keeps StorageReview’s 45–80W laptop range and does not adopt 140W as a verified laptop TDP.

Memory bandwidth is not the DGX Spark’s 273 GB/s. StorageReview says Dell’s XPS spec sheet lists LPDDR5X at 9,400 MT/s, which on a 256-bit bus works out to 300.8 GB/s, in line with a 300 GB/s figure NVIDIA quoted in June, about 10% above DGX Spark. That calculation is theirs, from a Dell sheet, not a LocalRig measurement of a Surface Laptop Ultra. Tom’s Hardware on 15 October 2026 states the platform maximum as 300 GB/s of shared memory bandwidth, and it lists the reviewed Surface at 128GB of LPDDR5X-9400. Windows Central’s spec table the same day says up to 9,400 Mbps, a 256-bit bus, and up to 300 GB/s. Those are review statements of the spec, community-cited, not a LocalRig bandwidth row. The CPU and GPU are connected at 600 GB/s NVLink-C2C in the Windows Central table, matching the NVIDIA blog. LocalRig has no Surface bandwidth measurement of its own.

Which memory size holds which model?

Fit the file under the GPU ceiling, not under the number on the sticker. 24GB and 32GB stop at the same class as a 24GB discrete card.

StorageReview’s line on the small configs is the working rule: 24GB and 32GB leave room for 27B-class models at 4-bit and not much above. The large-model claims assume 128GB. Use the VRAM calculator as an estimator for a dense model. It is not an RTX Spark measurement.

Workload	Cited size	24GB laptop	32GB laptop	48GB laptop	128GB laptop (64GB default GPU allocation on the 15 October review; 110GB total GPU on one earlier Surface screenshot)
Qwen 3.8 27B, Unsloth 4-bit	16–19GB before long context; UD-Q4_K_XL file 17.6GB (card and docs, accessed 2026-09-06, on the 27B page)	Weights fit. Long-context KV: not yet measured. This file is the wrong reason to buy a Spark laptop	Weights fit. Same caveat	Weights fit	Weights fit. You are paying for memory you do not need for this file
70B Q4 weights	About 35–40GB of weights, with a 46–48GB comfortable floor once KV is included (70B page)	No	No. 32GB is under the weight file	The weight file can sit inside 48GB. OS plus KV on a shared Windows pool: not yet measured. No Surface price for 48GB was in the config article	Weights fit. Context on top of them: not yet measured
MAI Code 1.1 Flash	137B total, 6.8B active. About 3.3 bits per weight. 53GB weights; 75.5GB peak at 256K context (Microsoft, via StorageReview)	No	No	53GB weights do not fit in 48GB	53GB weights can sit inside a 64GB allocation with little spare. The 75.5GB peak at 256K does not. That peak needs the allocation raised, and it fits under the older 110GB total. This is a vendor footprint, not a LocalRig load
Qwen 3.8 Flash-Next GGUF	About 75GB of RAM (Unsloth, accessed 2026-09-07, on the Flash page)	No	No	No	The file is over the 64GB default and under the older 110GB total, so it needs a raised allocation. Whether it loads on this laptop: not yet measured. NVIDIA’s blog names a 125B Qwen 3.8 Flash Next as a model the platform can run. That sentence does not name the quant or the SKU
DeepSeek-V4-Flash, Unsloth 3-bit	About 103GB on the 0731 page	No	No	No	Over the 64GB default. Under 110GB on file size, with little spare, if the allocation can be raised that far. Microsoft names a 284B DeepSeek V4 Flash as coming to RTX Spark and does not state the footprint. The 103GB figure is Unsloth’s 3-bit band, not a Spark load

Parameter slogans do not close that table. StorageReview lists the spread: NVIDIA’s product page says models up to 120 billion parameters with a 1 million token context; NVIDIA’s blog cites Qwen 3.8 Flash Next at 125B; Microsoft says models exceeding 120 billion and then names MAI Code 1.1 Flash at 137B and DeepSeek V4 Flash at 284B; Dell quotes 200 billion. The spread is bits per weight. A 284B model at a high bit width does not fit under a 110GB GPU ceiling. StorageReview’s own arithmetic is that DeepSeek V4 Flash at 284B needs about 3.1 to 3.3 bits per weight or fewer to fit under that ceiling before context. That is their estimate. It is not a published file size.

How much of the 128GB can the GPU use?

Installed RAM is not the GPU budget. The 15 October review default is 64GB, and it is adjustable. An earlier Surface capture showed 110GB total GPU memory.

Windows Central (Zac Bowden, 15 October 2026) says the 128GB configuration is set to split that memory evenly between system RAM and the GPU. In the default configuration, and for most of the benchmarks in that review, the GPU had 64GB of video RAM. Models that do not fit can be given a larger allocation. The review does not publish the maximum that control will allow, and it does not say that every benchmark used the 64GB split. This is community-cited. LocalRig has not opened that control.

StorageReview describes a different Task Manager screenshot, from a Microsoft and GitHub engineering post, on a Surface Laptop Ultra with the 6,144-core N1X:

• Dedicated GPU memory: 74.9GB
• Shared GPU memory: 35.5GB
• Total GPU memory: 110GB
• System memory: 51.5GB
• Driver: 32.0.16.1630, dated 31 July
• The NPU is listed as a separate device

Dedicated plus system is 126.4GB, so Windows is presenting the unified pool the way it presents a discrete card. On that unit the GPU ceiling is 110GB, which is 86% of the installed 128GB. Weights, the KV cache, and the runtime all have to fit under it. The operating system and other applications draw from the same pool.

That is one screenshot of one Surface, with a July driver, reported on 7 October. The 15 October reviews do not repeat it. Plan a 128GB order as a machine whose default GPU allocation is 64GB, which can be raised, and whose only published total-GPU figure is still that 110GB capture. It is not a promise that every 128GB RTX Spark laptop exposes 110GB, and it is not a promise that a future driver will expose 128GB.

This cap has a cousin on the AMD side. AMD’s Variable Graphics Memory FAQ, cited on the mini PC guide, documents a 128GB Strix Halo system at 96GB dedicated and 112GB total graphics-addressable. Different silicon, same lesson: the sticker is not the weight budget. The Strix Halo vs Mac Studio page is the platform comparison. This page does not re-measure either pool.

What speed numbers exist, and which cells are empty?

Independent reviews published 15 October 2026. The tok/s are Procyon, on small models, with quantization unstated. The large-model comparison is AgentPerf-local completion time, not decode tok/s. Both are community-cited. LocalRig has not run this hardware.

Procyon text generation, Windows Central

Windows Central (Zac Bowden, 15 October 2026) ran UL Procyon text generation on a Surface Laptop Ultra with the 20-core CPU, 6,144 GPU cores, and 128GB of RAM. The benchmarks section says every test ran on power, with the Best Performance profile, unless stated otherwise. The same review says most benchmarks used the default 64GB GPU allocation. It does not attach that sentence to this table alone.

Quantization, context length, batch size, and the Procyon version are not stated. Do not paste these rates onto a 24GB or 32GB laptop, and do not treat them as a llama.cpp decode of a 27B model.

Model	Overall score	tok/s	Time to first token	Model load
Phi-3.5-mini	2,615	72.21	0.27 s	0.66 s
Mistral 7B	2,276	53.07	0.47 s	0.96 s
Llama 3.1 8B	2,251	49.44	0.40 s	1.50 s
Llama 2 13B	2,108	27.44	0.84 s	1.69 s

The body text says time to first token was 0.39 seconds on “LLaMA3”. The results table, printed twice, says 0.40 seconds for Llama 3.1 8B. This page uses 0.40. Windows Central also says sustained decoding trails higher-wattage discrete GPUs. The comparison-score row for an RTX 5060 on “LLaMA 2” prints 249 against thousands on the other GPUs, so this page does not adopt the competitor score table.

AgentPerf-local completion time, Tom’s Hardware

Tom’s Hardware (Jeffrey Kampman, 15 October 2026) did not publish decode tok/s. It published AgentPerf-local time-to-completion, replaying agent traces. The inference backend was the latest llama.cpp, built from source on each system, with CUDA, Metal, or ROCm.

The Surface was the top configuration: 20 CPU cores, 6,144 CUDA cores, 128GB LPDDR5X-9400. The Mac was a 14-inch MacBook Pro with M5 Max, a 40-core GPU, an 18-core CPU, and 128GB. The AMD system was a Ryzen AI Halo mini desktop with a Ryzen AI Max+ 395, not a notebook. Battery life, build, and display were outside this embargo and are not in the review.

Qwen 3.8 27B, Unsloth Q4_K_M. DFlash2 speculative decoding on the RTX Spark and the MacBook Pro. The Ryzen AI Halo used that model’s built-in MTP. Tom’s Hardware says the M5 Max took about 25% longer than the Surface. It says the Ryzen AI Halo needed about twice as long as either, at 40 minutes. Absolute minutes for the Surface and the M5 Max on this model were not printed. This page does not divide 40 minutes into a Spark tok/s or a Spark minute count. “About twice as long as either” sits next to a 25% gap between the Surface and the Mac, so the 40 minutes is the review’s Halo figure.

Qwen 3.6 35B A3B, Unsloth Q4_K_M, MTP enabled on the run. Tom’s Hardware says the Surface finished in eight minutes, the 14-inch MacBook Pro in ten minutes (which the review calls about 25% faster), and the Ryzen AI Halo in 13 minutes.

Context length and batch size for AgentPerf-local were not stated. Speculative decoding and MTP mean these are not plain llama.cpp decode rates.

The vendor chart is still a vendor chart

Microsoft’s MAI Code 1.1 Flash numbers, as StorageReview reports them, remain the vendor packet. They are not replaced by the Procyon rows:

• Model: MAI Code 1.1 Flash, 137B total, 6.8B active, mixed precision at about 3.3 bits per weight.
• Weights: 53GB. Peak memory at 256K context: 75.5GB.
• Machine: a Surface Laptop Ultra. The 75.5GB peak does not fit the 24GB, 32GB, or 48GB steps, and it does not fit the 64GB default GPU allocation, so this demo needs a raised allocation on a 128GB-class pool. Microsoft’s chart does not, in StorageReview’s account, label the RAM size.
• Runtime: Windows Arm64 llama.cpp, CUDA backend, plus what Microsoft calls DFlash2 sliding-window speculative decoding.
• Workload: synthetic code generation.
• Quality, vendor: 70.8% SWE-Bench verified versus 72.6% for the BF16 cloud version. Terminal-Bench 2.1 at 66.29% versus 62.9%. StorageReview notes the second gap is about three tasks on an 89-task set.

The decode chart has no data labels. StorageReview estimated, from the drawing: about 63 tok/s at a 2K prompt, 61 at 8K, 57 at 32K, 52 at 64K, 55 at 120K, and 39 at 256K. Those tilde figures are a journalist’s read of a vendor chart. LocalRig will not promote them to a benchmark row. Speculative decoding means they are not a plain llama.cpp decode rate.

Two prompt-processing figures are labeled, and they are still vendor: 923.5 tok/s at 64K context and 769.8 tok/s at 128K. StorageReview converts that to about 71 seconds before the first token on a cold 64K prompt, and about 170 seconds on a cold 128K prompt. Prefill, not decode.

The Mac comparison in the same article is a different test and a smaller pool. Microsoft’s footnotes, via StorageReview: up to 2.1× faster time to first token versus a 16-inch MacBook Pro with M5 Pro and 64GB, in llama.cpp, Qwen3.5 27B at Q4_K_M, fixed 8,192-token prompt. The RTX Spark side was a preproduction 64GB system, tested in September. There is no M5 Max in that comparison, and neither company published a decode tok/s comparison. Image and video multiples in the same footnote (4.3× and 6.2×) are NVIDIA ComfyUI tests. They are not LLM decode.

LocalRig tok/s on any RTX Spark laptop: not measured. The independent cells that are now filled are the community-cited Procyon rates and the AgentPerf-local completion times above. Battery life under inference and sustained watts in a long agent session are still absent from both 15 October reviews, which say that embargo withheld laptop-specific results. Decode tok/s for Qwen 3.8 27B, a 70B, MAI Code, or a Flash-class file on this laptop are still unpublished.

How does the 128GB laptop compare with a Mac, a DGX Spark, and a Strix Halo box?

Pay the laptop premium only for the laptop. The memory class is already on sale in other shapes.

Per configured gigabyte below divides the whole reported computer price by the RAM sticker. It includes the chip and the SSD. It is not a price for the 64GB default GPU allocation or the 110GB screenshot, and it is not a marginal RAM upgrade unless both rows share the chip and the drive.

Configuration	Reported price	$ per configured GB	What you are actually buying
Surface Laptop Ultra 24GB / 512GB, 5,120-core	$2,599	$108.29	A 27B-class 4-bit pool on the chip that cannot reach 128GB
Surface Laptop Ultra 32GB / 1TB, 6,144-core	$3,699.99	$115.62	The large chip, still a 27B-class 4-bit pool
Surface Laptop Ultra 128GB / 1TB, 6,144-core	$5,899.99	$46.09	The only Surface laptop in this set that matches the large-model claim. The 15 October review default is a 64GB GPU allocation, adjustable. The earlier screenshot’s total is 110GB
Step from that 32GB row to that 128GB row	+$2,200 for 96GB	$22.92 per extra GB	Same larger chip, same 1TB floor, in the config article. Intermediate Surface RAM prices were not listed

The 32GB row costs more per configured gigabyte than the 24GB row because the chip and the SSD changed. It is the expensive way to stay in the same model class.

Mac Studio M5 Max, 128GB. The M5 configuration guide records $5,099 for M5 Max, 40-core GPU, 128GB, 512GB SSD, from Apple’s US store on 6 October 2026, at 614 GB/s. This laptop page did not re-open the configurator on 9 October or on 15 October. That list is $800.99 under the Surface’s $5,899.99 128GB / 1TB figure, on a different SSD. Bandwidth on the M5 Max spec is about double the 300 GB/s class attributed to the N1X laptop. The Mac is MLX and Metal. It is not CUDA. M5 decode tok/s on that guide are still empty. Tom’s Hardware’s AgentPerf-local result is a 14-inch M5 Max MacBook Pro with 128GB, not this Studio: on Qwen 3.8 27B Q4_K_M the MacBook Pro took about 25% longer than the Surface. That is completion time, community-cited, not a Studio decode rate. Microsoft’s 2.1× prefill claim does not apply to this Studio. It was a 64GB M5 Pro laptop, on a preproduction 64GB RTX Spark.

DGX Spark. The 64GB vs 128GB page is the desktop split: NVIDIA’s 2 October 2026 blog says the 64GB configuration ships from OEM partners on 23 October 2026, starting at $4,999. The Register the same day reported the 128GB list at $6,950. StorageReview on 7 October uses that $4,999 64GB figure when it prices the Dev Box. This page did not re-fetch the NVIDIA product page on 9 October. DGX OS, ConnectX-7, and 273 GB/s are the platform. A 64GB Spark does not hold the 75GB-class files. A reported $6,950 128GB Spark is $1,050.01 above the Surface laptop’s $5,899.99 top and adds the cluster port the laptop does not have.

Surface RTX Spark Dev Box. $5,999, 128GB, November, US Microsoft.com only. StorageReview says that is $1,000 above the 64GB DGX Spark, with twice the memory, and that Microsoft lists USB-C, USB-A, HDMI, Ethernet, and a headphone jack, with no ConnectX-7. Against the laptop’s verified 128GB price, the Dev Box is $99.01 more and it is not a laptop. Microsoft’s Dev Box page, via StorageReview, quotes a 100W thermal envelope, which is 40W under the 140W desktop N1X TDP NVIDIA lists. Those may be different measurements. Neither is a tok/s. If you do not need a battery, this is the Windows 128GB box to price against the laptop, and the DGX Spark is the box to price if you need the cluster NIC.

Strix Halo and Gorgon Halo mini PCs. The mini PC guide, manufacturer listings checked 6 October 2026, has a Framework Desktop DIY Max+ 395 at 128GB from $3,449, a Minisforum MS-S1 MAX 128GB edition on sale at $3,799, and a Beelink GTR9 Pro 128GB at $4,349. Those are not laptop prices and they are not CUDA. Platform bandwidth on that page is 256 GB/s for Strix Halo, under the N1X laptop’s reported ~300 GB/s class and far under the M5 Max’s 614 GB/s. The GPU-addressable slice is the AMD figure already cited: 96GB dedicated, 112GB total. A 128GB Strix Halo box is the lower bill when the file fits that slice and you do not need NVIDIA’s stack. The Framework page covers the desktop specifically: Framework Desktop for local AI. Tom’s Hardware’s Ryzen AI Halo result is a separate, community-cited timing on that class of chip: 40 minutes on the Qwen 3.8 27B AgentPerf-local run, and 13 minutes on Qwen 3.6 35B A3B against the Surface’s eight. It is not a re-check of those mini PC prices.

Used 24GB NVIDIA, when the file is the 27B Q4. Unsloth’s 4-bit band for Qwen 3.8 27B is 16–19GB. A 24GB card is the matching pool. The used RTX 3090 guide owns the street price. The DGX Spark page’s CompSniper median of $1,499 is a 6 October 2026 fetch across 204 completed sales. This page did not re-pull eBay on 9 October. Even that October median is far under a $2,599 24GB Spark laptop, and the laptop adds Windows on Arm, a 45–80W envelope, and no extra model class.

Browse used RTX 3090 24GB on eBay →

Rent, if you have not proved the model needs a $2,600–$5,900 laptop. The rent vs buy break-even is the check. A rental is not an RTX Spark rental. It is how you learn that the real job was a 27B Q4.

Who should buy which config?

Match the order to the file. Then decide whether it has to travel.

Buy the 6,144-core, 128GB laptop when all of these are true:

• The weights plus the context you will use sit above 32GB. At the Windows Central default they also have to fit a 64GB GPU allocation, or you raise that allocation. The only published total-GPU figure above that default is still the 110GB Surface screenshot from 7 October.
• You need CUDA on Windows, in a laptop. First shipments are 16 October 2026. Tom’s Hardware on 15 October says the top Surface it reviewed is already quoting lead times into late December.
• You can use the community-cited 15 October numbers: Procyon tok/s on Phi-3.5-mini, Mistral 7B, Llama 3.1 8B (49.44 tok/s), and Llama 2 13B, with quantization unstated, plus AgentPerf-local completion times on two Qwen quants. You do not yet have a LocalRig run, a stated Procyon quant, or battery-under-load.
• The cart you are paying is the 128GB line. On the Surface, the named top is $5,899.99 with 1TB in the 7 October config article. Tom’s Hardware prints that class of machine as $5,899. If Dell’s XPS floor of $3,799.99 is the 128GB machine in StorageReview’s headline, confirm that RAM line on Dell’s page before you treat the headline as the SKU.

Do not buy the $2,599 Surface for local LLMs beyond a 27B-class 4-bit model. It is 24GB on the 5,120-core chip. The keynote sentence about models that do not fit a traditional machine is the 128GB configuration. Pavan Davuluri’s line on NVIDIA’s blog — up to 128GB and up to a petaflop — describes that ceiling, not the baseline SKU.

Do not buy the $3,699.99 Surface for those larger models either. Windows Central’s config article treats that 32GB / 6,144-core / 1TB configuration as the place advanced users start. For a local LLM whose size is the constraint, it is the sharper trap: you paid for the chip that can be ordered with 128GB, and the computer in the cart still holds the 27B-class pool. The step to 128GB on that same chip and 1TB floor is +$2,200 in the config article. Pay it, or buy a different class of machine. Do not pay $3,699.99 and expect the 53GB MAI weights or a ~75GB Flash GGUF to fit. On the 128GB machine, the ~75GB file and the 75.5GB MAI peak still do not fit the 64GB default allocation. Raising that allocation is part of the job. Ordering 32GB does not create the allocation to raise.

48GB is on NVIDIA’s marketplace in StorageReview’s read and has no Surface price in the config article. A 70B Q4 weight file can be inside 48GB and still fail once Windows and the KV cache share the pool. That failure is not yet measured. 48GB does not hold the 53GB MAI weights or the ~75GB Flash file.

Buy a desk machine instead when the laptop is not the constraint. Dev Box at $5,999 if you want Windows, 128GB, and no cluster NIC. DGX Spark 128GB at the reported $6,950 if you want ConnectX-7 and DGX OS. DGX Spark 64GB from $4,999 on 23 October only when the file fits in 64GB. M5 Max 128GB at the 6 October list of $5,099 when you can live on MLX or Metal and you want the 614 GB/s pool. Strix Halo 128GB from the 6 October manufacturer checks when CUDA is not required and the AMD addressable slice holds the file.

Who this is NOT for

• Anyone buying “an RTX Spark” without reading the RAM line. Seven of the fifteen marketplace configs StorageReview read use the chip that cannot be configured to 128GB. The $2,599 Surface is that chip, at 24GB.
• Anyone whose model is Qwen 3.8 27B at Q4. The file fits a 24GB card. A Spark laptop’s price, at every named Surface step, is the wrong bill for that download.
• Anyone who needs a LocalRig-measured tokens-per-second number, a stated Procyon quantization, or battery life under inference before they pay. Independent reviews published 15 October 2026. They fill Procyon tok/s for four small models and AgentPerf-local completion times for two Qwen quants. They do not state the Procyon quant, and both reviews say battery and chassis results were outside the embargo. Waiting for a review that pins quant, context, and watts is still a coherent choice.
• Anyone who does not need a laptop, or does not need CUDA. A Strix Halo mini PC, an M5 Mac Studio, a DGX Spark, or the Surface Dev Box is the comparison. The laptop tax is portability plus Windows CUDA, not a larger memory pool than those desks.
• Anyone treating 128GB installed as 128GB of weights. Windows Central’s 15 October review says the default GPU allocation is 64GB. The earlier Surface capture shows 110GB total GPU memory. A ~103GB 3-bit file does not fit the default, and it is already tight under that 110GB capture before context.

What this page will not fill in

• A LocalRig tok/s row. The 15 October figures are community-cited reviews of the top Surface Laptop Ultra, published the day before first shipments. They are not a LocalRig run, and they are not a review of every OEM. Tom’s Hardware says other RTX Spark laptops are in its lab and that those reviews are still to come.
• Procyon quantization, context, batch, or benchmark version. Not stated. The Llama 3.1 8B rate of 49.44 tok/s is still usable as a cited Procyon result with that gap left visible.
• AgentPerf-local minutes for Qwen 3.8 27B on the Surface and on the M5 Max. Not printed. The review gives the M5 Max as about 25% longer, and the Ryzen AI Halo as 40 minutes.
• Decode tok/s for a 27B, 70B, MAI, or Flash-class file on this laptop. AgentPerf-local is completion time, with DFlash2 or MTP. The MAI chart is still vendor, preproduction, and partly unlabeled.
• Battery life, fan noise, thermals, and display quality. Both 15 October reviews say the embargo withheld those results.
• A Microsoft, Dell, or HP cart screenshot. The price table is Windows Central and StorageReview on 7 October 2026, re-read here on 9 October. Tom’s Hardware on 15 October states a $5,899 list and late-December lead times for the Surface it reviewed. That sentence is not a refresh of the other rows.
• An HP price of $4,999.99. Not in the roundup fetched for this page.
• RAM sizes on the ASUS, HP, Lenovo, and MSI floors. Unknown. The Dell 128GB-at-$3,799.99 reading is a StorageReview headline, not a configurator capture.
• A single GPU ceiling for every 128GB laptop. The review default is a 64GB allocation, adjustable, with no published maximum. The 110GB total is one earlier Surface screenshot.
• A 6,144-core laptop TDP of 140W. Windows Central’s spec table prints 45–140W. StorageReview’s laptop column is 45–80W. Tom’s Hardware’s 140W figure is the Surface power adapter. Unresolved, so this page does not adopt 140W as SoC TDP.
• A 48GB Surface price, a 64GB laptop SKU, and any claim that RAM is upgradable.
• Gaming as a reason to pick the config. Tom’s Hardware did publish two rows, community-cited, and treats gaming as the weak side of the platform. Gears of War: E-Day at 1908×1272, High, Mega Geometry on, MFG 2×: 77.7 average fps, 39.9 1% low, 83.5 ms average PC latency. Control Resonant at 2560×1440, High, ray tracing off, MFG 3×: 109.2 average fps, 51.6 1% low, 78.3 ms. Windows Central’s game notes are not in a results table, so this page does not adopt them. These rows do not change the memory-config decision.

Sources

• StorageReview, Brian Beeler, 7 October 2026: preorders open that day, laptops 16 October, desktops in November. Two laptop N1X parts (6,144-core up to 128GB, 5,120-core up to 64GB), both 45–80W; desktop N1X 140W. Marketplace read: 24/32/48/128GB, seven of fifteen configs on the smaller chip. Task Manager capture: 110GB total GPU memory on a 128GB Surface Laptop Ultra, driver 32.0.16.1630 dated 31 July. 24GB/32GB described as 27B-class at 4-bit. MAI Code 1.1 Flash vendor packet, including the unlabeled decode chart StorageReview estimated at about 39–63 tok/s, and labeled prefill of 923.5 tok/s at 64K and 769.8 tok/s at 128K. Dell LPDDR5X-9400 arithmetic to 300.8 GB/s. Surface Dev Box $5,999 / 128GB / November / no ConnectX-7 mentioned. Related headline on that page: Dell XPS 16, 128GB, $3,799.99. Accessed 2026-10-09.
• Windows Central, Cale Hunt, 7 October 2026, Surface configs: baseline $2,599 with 18-core CPU, 5,120-core GPU, 24GB, 512GB. Larger chip at $3,699.99 with 32GB and 1TB. Top $5,899.99 with 128GB and 1TB. Shipments 16 October. Accessed 2026-10-09.
• Windows Central, Cale Hunt, 7 October 2026, OEM roundup: Surface from $2,599.99; ASUS ProArt P14 from $2,599.99 and P16 from $2,799.99; Dell XPS 16 from $3,799.99; HP OmniBook X 14 from $2,999.99 and OmniBook Ultra 16 from $3,199.99, both shipping 4 November; Lenovo Yoga 9n from $3,099.99; MSI Prestige N16 Flip AI+ from $3,299.99. Accessed 2026-10-09.
• Tom’s Hardware, Jeffrey Kampman, 15 October 2026: community-cited review of a Surface Laptop Ultra with 20 CPU cores, 6,144 CUDA cores, and 128GB LPDDR5X-9400. AgentPerf-local via llama.cpp built from source. Qwen 3.8 27B Unsloth Q4_K_M with DFlash2 on the Surface and a 14-inch M5 Max MacBook Pro (40-core GPU, 18-core CPU, 128GB): the M5 Max took about 25% longer. Ryzen AI Halo (Ryzen AI Max+ 395, MTP) took about twice as long as either, at 40 minutes. Qwen 3.6 35B A3B Unsloth Q4_K_M with MTP: Surface eight minutes, MacBook Pro ten minutes, Ryzen AI Halo 13 minutes. Platform memory bandwidth stated as a 300 GB/s maximum. Tested Surface listed at $5,899 with 1TB; lead times late December. Gaming rows cited above. Battery, build, and display withheld. Accessed 2026-10-15.
• Windows Central, Zac Bowden, 15 October 2026, RTX Spark review: community-cited. 128GB configuration defaults to an even split, 64GB GPU allocation for most benchmarks, adjustable. Procyon text generation on the 20-core, 6,144-core, 128GB Surface, on power, Best Performance: Phi-3.5-mini 72.21 tok/s, Mistral 7B 53.07 tok/s, Llama 3.1 8B 49.44 tok/s (TTFT 0.40 s in the results table), Llama 2 13B 27.44 tok/s. Quantization not stated. Spec table: up to 300 GB/s, LPDDR5X up to 9,400 Mbps, 256-bit, NVLink-C2C 600 GB/s. The 6,144-core TDP cell prints 45–140W and is not adopted. Battery and thermals withheld. First RTX Spark PCs launch the next day. Accessed 2026-10-15.
• NVIDIA blog, Gerardo Delgado, 7 October 2026: preorders open, laptops available 16 October, compact desktops in November. Up to 6,144 cores, up to 20-core Grace at 600 GB/s, up to 128GB, vendor claim of one petaflop FP4. Qwen 3.8 Flash Next named as a 125B example. OEMs: Acer, ASUS, Dell, HP, Lenovo, Microsoft, MSI, Gigabyte. Davuluri’s “up to 128 gigs” sentence. Accessed 2026-10-09.
• NVIDIA blog, 2 October 2026: 64GB DGX Spark from partners on 23 October 2026, starting at $4,999. Detail and the 128GB comparison live on DGX Spark 64GB vs 128GB.
• The Register, Tobias Mann, 2 October 2026: 128GB DGX Spark list reported at $6,950.
• Apple US store figures for M5 Max 128GB at $5,099 and 614 GB/s are the ones already published on the M5 guide, fetched 2026-10-06. Not re-fetched for this page.
• Strix Halo prices and the 96GB / 112GB graphics-addressable split are the ones already published on the mini PC guide, manufacturer pages fetched 2026-10-06, plus AMD’s Variable Graphics Memory FAQ linked there. Not re-fetched for this page.
• Unsloth file sizes cited above are the ones already published on the linked can-i-run pages (accessed 2026-09-06 and 2026-09-07 on those pages). They are not RTX Spark benchmarks.
• CompSniper’s RTX 3090 median of $1,499 is the 2026-10-06 fetch already written up on the DGX Spark page. Not re-fetched for this page.