Surface Laptop Ultra Costs $2,599 to $5,900, and Its 120B-Model Claim Only Fits in the 128GB Tier

The headline figure on Microsoft's Surface Laptop Ultra is one petaflop, and it is the least useful number on the spec sheet. Microsoft's own page and every outlet that has covered the October 7 pricing attach that figure to FP4, a four-bit format, and none of the four pages read for this piece states how fast the memory behind it runs.

The number that matters here is not compute but the price of each gigabyte of unified memory, because that is what decides which local models the machine can hold at all.

What Microsoft is selling and when it ships

Microsoft opened pre-orders on October 7 at $2,599.99 on its product page, with a promotion running to November 23. Gizmodo reports a launch date of October 16; the Microsoft page gives no ship date.

The chip is Nvidia's RTX Spark, a 20-core Arm CPU co-developed with MediaTek paired with a Blackwell GPU of up to 6,144 CUDA cores, according to Windows Latest. The chassis is 15 inches, under 18 mm thick and under 4.5 lb, with two fans.

Microsoft announced the machine on May 31, when Brett Ostrum, corporate vice president for Surface, described up to 128GB of unified memory shared between CPU and GPU. The same post says one petaflop is enough to run models of up to 120B parameters locally.

Three price tiers, and where the memory sits

Gizmodo's breakdown of Microsoft's event is the only place the tiers appear side by side. Its own base-RAM figures conflict, so treat the low end as approximate.

Tier	CPU	Memory options	Starting price
Base	18-core	up to 32GB	$2,599
Mid	20-core	32, 48, 64 or 128GB	$3,700
Top	20-core	128GB	$5,900

Read the table as a memory ladder. The 20-core tier at $3,700 starts at 32GB; the $5,900 configuration is the 128GB one. If storage is equal across the two, which Gizmodo does not say, 96GB of extra memory costs $2,200, or about $23 a gigabyte at the margin.

1. Base, 18-core2599 USD
2. Mid, 20-core3700 USD
3. Top, 128GB5900 USD

Source: Microsoft product page (base) and Gizmodo (mid and top tiers), accessed 2026-10-08

Why 120 billion parameters is a memory claim

A 120B-parameter model stored at four bits per weight needs 60GB for the weights alone (120 billion parameters at half a byte each). That is my arithmetic, not Microsoft's, and it ignores the key-value cache that grows with context length. It still shows why only the 128GB tier can honour the claim, and Microsoft's product page adds that the GPU can address less than the full pool.

Speed is the second variable. Generating one token from a dense model reads every weight once, so the ceiling is bandwidth divided by model size. At 60GB, each 100 GB/s of bandwidth buys at most 1.67 tokens per second; a mixture-of-experts model that reads fewer weights per token would do better. The product page does not list bandwidth, and Windows Latest says the same.

The comparison Microsoft chose

Engadget reports that Microsoft compared the Ultra with a MacBook Pro using an M5 Pro chip, claiming 6.2 times the speed at video generation, 4.3 times at image creation and 2.1 times faster to first token. These are Microsoft's figures, and no independent party is named as having run them.

The pairing is informative. First-token latency is set mainly by compute, which suits a GPU with FP4 tensor cores. Sustained generation after the first token is set by bandwidth, and that is where no figure was offered. Gizmodo adds that Apple sells a similar split, with its M5 Max tier at $5,000 or more with 128GB.

This is not the first Arm laptop whose headline claims outran its disclosures; the Adreno X2 benchmark claims followed a similar pattern, and the lesson from Apple's neural engine bandwidth work is that the bus, not the multiplier array, usually sets the ceiling.

What the cheaper tiers can and cannot hold

The same arithmetic runs downward. At four bits per weight, 32GB of memory could in theory hold weights for a 64B-parameter model, but the operating system, the cache and the GPU's addressable share all come out of that pool, so the practical ceiling is lower. Microsoft's product page says only that the GPU can address less than the total.

Engadget also lists a Surface RTX Dev Box from $5,999, a desktop carrying the same chip family. A laptop at $5,900 and a desktop at $5,999 sit a hundred dollars apart; Engadget does not give the Dev Box's memory, so the pairing shows the price band, not a memory price. Engadget adds Microsoft's claim that the Ultra delivers 99 percent of its maximum performance on battery, which is vendor-supplied and untested.

What would change my read

Three things. A published memory bandwidth at or above the level of a discrete laptop GPU would make the 128GB tier a bargain for local inference. An independent test of tokens per second on a 60GB model would settle the question Microsoft's comparison skips. And a clear statement of how much of the pool the GPU can address would tell buyers whether 128GB is a ceiling or a headline.

Until then the sensible reading is narrow: $2,599 buys a thin Arm laptop with a capable GPU, and $5,900 buys the only configuration that matches the 120B claim.