Nvidia 128-gigabyte AI portable under review: bandwidth gap plus purchase guide

I picture memory as an ink bottle in front of a printer: a huge bottle with a narrow neck still prints slowly. The RTX laptop sits in the same spot; Nvidia raised the shared pool to 128 GB but never published the width of the pipe, the bandwidth. Two figures float around, 273 and 300 GB per second, and neither is official. My math is simple division: gigabytes flowing per second divided by gigabytes read for each text piece gives the ceiling of that machine. That ceiling lands below 100 pieces, and I read it as a cap, not a forecast. When I place two laptops side by side I use the same ruler: twice the pipe means roughly twice the ceiling. I throw out speed claims with no source using that ruler.

The component ships in two laptop setups: the full one carries a 20-core processor with a 6144-core GPU , the cut-down one an 18-core processor with a 5120-core graphics part. Memory is the real split: entry setups come with 24 and 32 GB, where versions up to about 14B settle comfortably while 32B barely squeezes in. At 64 GB a 32B version breathes. The 120B class wants the 128 GB setup; weights alone take 60 to 65 GB, plus the share the system keeps for itself. Guidance from Microsoft confirms it: on the 128 GB laptop the pool visible to graphics lands near 110 GB. A unified pool means no separate video memory; everyone drinks from one cup. My rule runs like this: buy twice the memory of the largest weights you plan to run.

Price per gigabyte: same money, different cup

Dividing each price by the memory inside gives my most honest table. The 128 GB Surface laptop at 5899 dollars lands near 46 dollars per gigabyte; with a 15-inch screen under 4.5 pounds it is the only laptop in this lineup. Pre-orders opened October 7, first units arrive October 16. The RTX dev box at 5999 dollars writes 47 dollars per gigabyte and ships in November with developer tools ready. Nvidia own desktop DGX Spark now costs more with equal memory: a 3000 dollar launch price became 3999, then 4699, then 6950 on October 2. The Framework side cuts the 128 GB setup to 3449 dollars, near 27 dollars per gigabyte, the cheapest memory here. On the Apple side a 128 GB Mac Studio near 5399 dollars lands at 42 dollars per gigabyte. Records compiled by Wccftech show Asus and Dell setups arriving with the same laptop component.

On speed I use measured figures only. On GPT OSS 120B the GB10 desktop and the Strix Halo Framework box write at a similar pace: DGX lands between 35 and 60 pieces per second, Halo between 34 and 55, and the two ranges overlap almost fully. Long context costs speed; a box showing 53 pieces with an empty chat drops near 39 at 32K input, while newer software builds report above 60 empty. Reading input is where Nvidia leads: DGX reads up to 1956 pieces per second while Halo stays in the 835 to 1230 band. That gap is the wait before the first word when you paste a long document. A single independent test shows the M5 Max setup at 75.6 pieces; different engine, so no clean hardware verdict, but it lines up with a 614 GB per second pipe. GPU memory read speed sets writing speed, not raw compute.

Marketing claims against measured results

The 1 petaflop headline on every slide reads as decoration to me, a lab record in sparse FP4 form. A fair comparison wants dense BF16 power, which stays unpublished. Two claims from Microsoft deserve the same caution: the box runs weights above 120B locally and holds a 1 million piece context at interactive pace, yet version, compression, and cache storage stay unstated. A September comparison gave the RTX laptop up to 2.1 times faster first-word times; the test used a 27B version, Microsoft ordered the language test, Nvidia ran the image and video runs. Worse, the rival was the M5 Pro setup at 307 GB per second, not the M5 Max at 614. Writing speed, the memory-bound part, never appears in that table. My rule is short: a claim with missing conditions is not a claim.

The sharpest practical gap between the boxes is the software home: the RTX laptop runs Windows 11 on ARM while DGX runs Linux. AI tools install natively on the laptop, older x86 programs pass through the Prism translation layer. Nvidia opened native ARM development with toolkit version 13.4; PyTorch installs from the company package index and the Windows ML path uses TensorRT. The dev box arrives with VS Code and helper tools ready, plus GPU pass-through inside WSL. No incompatibility list exists yet; kernel drivers, cheat protection, and some VPN clients are the areas to watch. Gaming stays open: notes from July about trial drivers on a pre-production Surface unit mention rough edges, with no final-driver results. A single report says no add-in graphics card is supported. If your device must serve VPN and games, test both before paying.

Where this laptop actually wins

Still, this laptop owns a real slot: no other portable box pairs CUDA support with 128 GB of memory. That pair used to demand a desktop; now it fits in a bag. Price favors it too: the 128 GB laptop and the dev box both undercut the equal-memory DGX box. You also keep the Windows programs you already use instead of managing a separate Linux machine. Time to market impresses as well: about 4.5 months from announcement to first delivery, against 9 months for DGX. Input reading is the likely strength: GB10 already leads Halo there, and Microsoft own tests point the same way for RTX . If it behaves like GB10, long documents and large code bases should open fast. Every line stays conditional until independent tests land; the picture sharpens once first units reach hands on October 16.

The roadmap plus memory prices answer the wait-or-buy question together. Nvidia points to Vera CPU with Rubin graphics and LPDDR6 memory in 2027 and 2028, then the Rosa Feynman wave in 2029 and 2030, with no pipe-width figure on those slides either. Reports from TrendForce put memory prices up 13 to 18 percent in the third quarter of 2026 and 10 to 15 in the fourth; the maker side warns the shortage can run into 2028. DGX already rose twice with identical hardware inside. My read: the next component lands stronger but never cheaper. Waiting buys knowledge, not discounts. Power figures carried by StorageReview complete the table: laptop setups run 45 to 80W, the desktop box is rated 140W.

My buying map runs like this. If CUDA support, a Windows home, and 128 GB of memory must come together, this laptop has no rival; I would eye the 128 GB Surface or Asus setup but wait 2 or 3 weeks, never locking 5899 dollars before October 16 tests. For a CUDA desktop build I pick the Dev Box over DGX: equal memory for less money with tools ready. Work living on Linux, clusters of a few boxes, or night-long background jobs stays with DGX; the 863 total pieces across 256 parallel jobs earns its keep there. A private home chat box skips RTX and looks at the 3449 dollar Framework setup; same writing class at half the price. If writing speed leads and the Apple system fits, the 614 GB per second M5 Max Mac Studio is the fastest pen. Past 128 GB the 192 GB Gorgon Halo setup is the only option, with shipments just starting. I would skip the 64 GB DGX and the 24-32 GB entry setups for large weights: the first is priciest per gigabyte, the second cannot hold 120B.

Do monthly subscriptions die here? My answer is no. Open versions up to the 120B class cover private chat and documents you refuse to upload; the OpenAI gpt-oss 120B version at 5.1B active parameters sets that bar. Frontier models plus top image and video work still want the cloud. These devices take over part of the stack, never the whole of it. My takeaway runs like this: a genuine option for people who learn, test, and keep data on the machine, not for anyone wanting one box to do everything. Prices move monthly, so I recheck every figure on decision day.