A 128GB laptop sounds almost ridiculous until you try to run a large language model locally.
NVIDIA’s upcoming RTX Spark platform can put up to 128GB of unified memory beside a Blackwell GPU and Grace Arm CPU. NVIDIA says these Windows machines can run 120B parameter LLMs locally, work with context windows up to one million tokens, and still fit into thin laptops. That is a very different PC from the gaming notebooks Windows has spent years learning how to manage. The hardware is ready for giant shared memory pools. Windows is only now starting to expose controls that make those pools easier to use for AI.
A hidden feature found in Windows 11 build 29648.1000 makes the timing hard to ignore. The feature is called IntelligentCarveout inside the build. It adds strings such as “Reserved memory for accelerators” and “Memory for graphics and AI acceleration.” The setting appears to let Windows hold back part of unified memory for GPU and AI work instead of leaving the whole decision to the normal shared memory system. Microsoft has not announced it, it is disabled by default, and it may change before release. Still, the feature points at a problem Windows laptops did not really have at this scale before: when CPU, GPU, NPU, Windows, your browser, and a 100GB AI model all want the same memory pool, who gets first claim?
128GB is only useful if the accelerator can really use it
NVIDIA’s RTX Spark page advertises up to 128GB of unified memory, up to 1 petaflop of FP4 AI performance, a 20 core Grace CPU, and a 6,144 core Blackwell RTX GPU. NVIDIA is also selling the platform around local agents, large models, creative work, gaming, and all day battery life. That is a lot of different jobs for one shared memory pool.
Unified memory sounds simple. There is one big bucket instead of separate system RAM and GPU VRAM. The CPU can use it. The GPU can use it. AI accelerators can use it. A model that needs 70GB or 90GB does not have to squeeze itself into a 24GB or 32GB graphics card.
That is the good part.
The messy part is that Windows has historically thought about graphics memory through a mix of dedicated video memory, dedicated system memory, and shared system memory. Microsoft’s own Windows Display Driver Model documentation calculates how much system memory can be made available to graphics and how much can be committed at a time. On ordinary Windows PCs, the “Shared GPU Memory” number you see in Task Manager is not a chunk of RAM permanently stolen by the GPU. It is closer to a ceiling that Windows and the driver can use when needed.
That design made sense for years. Integrated graphics might borrow a few gigabytes. A discrete GPU mostly lived in its own VRAM. Normal applications still expected system RAM to be primarily CPU territory.
RTX Spark changes the scale.
A 120B model at low precision can ask for tens of gigabytes before the runtime, KV cache, context, Windows, browser tabs, developer tools, and background services take their share. Now “shared memory” is not a side feature. It can become the whole reason somebody bought the laptop.
I had to read Microsoft’s memory wording twice because “shared GPU memory” and “reserved memory for accelerators” sound like the same thing. They are not. The new hidden setting appears to be a real carveout. Memory reserved there would no longer be available to normal applications. That is a much stronger decision than simply letting a GPU borrow RAM when pressure rises.
And once you see that difference, NVIDIA’s 128GB number starts looking less like a normal laptop spec and more like an operating system problem.
Windows is testing a knob it never really needed before
Windows Insider build 29648.1000 arrived on August 17, 2026 in Microsoft’s Future Platforms Experimental channel. Microsoft did not mention unified memory controls in the public release notes. People inspecting the build found a new SettingsHandlers_UnifiedMemory.dll file and a hidden feature ID associated with IntelligentCarveout.
The unfinished setting reportedly sits under System and Advanced. It includes language telling users that Windows can reserve additional unified memory for graphics and AI intensive games and applications. It also warns that reserved memory is not available to other apps. Current screenshots from the experimental build show choices around how aggressively Windows should reserve that memory.
That is a very Windows solution to a very new PC problem.
For years, enthusiasts have gone into BIOS menus to change UMA frame buffer sizes on some integrated graphics machines. Others had no control at all. Windows generally managed shared graphics memory itself. Microsoft’s own support answers still tell users there is no normal Windows setting for manually changing that allocation on many current PCs.
Now local AI wants much bigger chunks.
Suppose a laptop has 128GB total. You load a model that wants 85GB once weights, cache, and runtime overhead are counted. Windows still needs memory. Your IDE wants memory. Chrome will happily eat whatever remains, because Chrome has never met free RAM it did not like. The GPU also needs working space. If the system lets normal applications squeeze the AI workload too hard, the model may fail to load, page memory, or run with less context than expected.
Reserve too much for AI and the opposite problem appears. Your expensive laptop can have 128GB installed while Windows and regular apps behave as though far less is available.
That is why a manual control makes sense.
It also shows how unusual these machines are.
The 128GB headline hides a resource fight
Local AI buyers usually think about memory in one direction: more is better.
For model capacity, yes.
For a general purpose Windows laptop, the answer gets annoying. Unified memory is shared by design. That means every extra gigabyte reserved for a model is a gigabyte that cannot simultaneously serve Windows, Photoshop, a game, a browser, or another process. The system can move things around, but moving memory around is not free and hard reservations reduce flexibility.
This becomes even stranger with agents.
NVIDIA is pitching RTX Spark as a PC where agents can write code, generate assets, run tools, work across applications, and keep large models local. A serious agent may have the LLM loaded while also running embeddings, a vector database, a browser, Python, local search, image tools, and normal desktop software. It might keep a long conversation history or process a large repository. The 128GB pool can disappear faster than the marketing picture suggests.
So the useful question is not “Does the laptop have 128GB?”
It is “How much of that 128GB can my actual workload keep available to the accelerator without making the rest of Windows miserable?”
That is the number I want to see in reviews.
A laptop could ship with 128GB and still behave badly if memory policy is too conservative. Another system could reserve a huge AI slice and then punish ordinary applications. OEM firmware, NVIDIA’s driver, Windows memory management, and the inference runtime all become part of the experience.
This is why spec sheet comparisons are going to get messy. Two laptops with the same RTX Spark chip and the same 128GB capacity may not feel identical if their firmware defaults, cooling, power limits, or memory policies differ.
The RAM number is becoming only the start of the review.
This is bigger than NVIDIA
It would be easy to make IntelligentCarveout an RTX Spark story because the timing is excellent. NVIDIA and Microsoft announced RTX Spark Windows PCs together in May. The first systems are expected from ASUS, Dell, HP, Lenovo, Microsoft, MSI, and other partners. The platform is exactly the kind of hardware that benefits from a large accelerator reservation.
But AMD already pushed Windows into similar territory.
Ryzen AI Max systems use large shared memory pools and can dedicate a big portion to graphics and local AI. AMD’s Halo machines showed why this matters before RTX Spark arrives. A model that cannot fit into 16GB or 24GB of VRAM can live inside a much larger shared pool.
Apple users will look at this whole discussion and wonder why Windows is only getting here now. Apple Silicon has treated unified memory as a normal part of the machine for years. Mac users do not usually think about a separate GPU VRAM pool because CPU and GPU live inside the same memory design from the start.
Windows had a harder job. It supports decades of PC hardware, discrete GPUs, integrated graphics, different firmware rules, x86 machines, Arm machines, and an enormous driver base. It cannot rewrite memory behavior as though every Windows computer suddenly became one unified design.
RTX Spark makes that old flexibility collide with a new workload.
And Windows has another complication here: RTX Spark uses an Arm based Grace CPU. Windows 11 on Arm can run x86 and x64 programs through Microsoft’s Prism emulation layer. Prism has improved a lot, but Microsoft itself documents that compatibility settings sometimes trade performance for app support. NVIDIA and Microsoft therefore are not just introducing a huge shared memory laptop. They are asking developers to accept a different CPU architecture at the same time.
That is ambitious.
Maybe too ambitious for version one.
A 120B model on a laptop will expose every weak assumption
NVIDIA says RTX Spark can run 120B parameter LLMs with up to a one million token context locally. That is the sort of sentence that makes people imagine a private frontier style assistant running beside Outlook.
The model size is only one piece.
Long context consumes memory through KV cache and other runtime state. Agents may run more than one model. Some applications will still be x64 programs under Prism. Games and creative tools can want GPU memory at the same time. Windows security features, background services, and normal desktop apps keep running because this is still your laptop, not a headless Linux server.
This is where the “personal AI computer” idea becomes more interesting than another TOPS race.
The PC has to decide what matters right now.
Maybe you are compiling code while a 70GB model sits loaded in the background. Maybe Photoshop opens a giant project while an agent is processing a repository. Maybe you close the AI app and expect that memory to become available immediately. Maybe a game wants a large graphics allocation five minutes later.
A fixed BIOS carveout is clunky for this kind of machine.
A Windows level control can be smarter.
That does not mean Microsoft’s hidden feature is finished or that it will solve everything. It is experimental, unannounced, and disabled. Microsoft could rename it, change how it works, restrict it to certain hardware, or remove it completely. Anyone treating the current screenshots as a promised Windows feature is getting ahead of Microsoft.
But the feature exists for a reason.
The old assumption that GPU memory is a smaller side pool is breaking down.
I would not buy RTX Spark for the 128GB number alone
The hardware sounds great on paper. I am interested in it. A Windows laptop with CUDA, Blackwell graphics, 128GB unified memory, and enough capacity for models that normally need workstation class hardware is exactly the kind of strange PC local AI people have been asking for.
Still, I would wait for independent tests.
I want to know how much memory Windows lets AI runtimes reserve by default. I want to know whether the manual carveout survives sleep and power profile changes. I want to know what happens when a 100GB workload is running and a normal x64 desktop application opens under Prism. I want battery tests with a large model actually generating tokens, not a generic web browsing loop. And I want to see how quickly the memory returns to ordinary apps after the AI workload closes.
Those answers matter more than another photo of a thin laptop beside “128GB.”
Cooling matters too. A chip can have enough memory for a 120B model and still throttle during long inference sessions. Unified memory bandwidth matters. Runtime support matters. Arm compatibility matters.
The laptop has to be good at being a laptop.
That sounds obvious, but AI PC marketing is getting very good at making us forget it.
Windows is finally treating memory like an AI setting
The hidden IntelligentCarveout control is small compared with the RTX Spark announcement. There is no giant launch event around it. Microsoft did not even put it in the August 17 release notes.
I think it may tell us more about the next generation of AI PCs than another petaflop number.
A 128GB unified memory laptop forces Windows to treat memory allocation as something the user may care about directly. Local models are large enough that accelerator memory is no longer a background detail. It becomes a choice between AI capacity and everything else the computer is doing.
That is new for mainstream Windows PCs.
NVIDIA’s hardware is arriving at exactly the moment Microsoft is experimenting with a way to expose that choice.
Maybe IntelligentCarveout ships with RTX Spark. Maybe it arrives later. Maybe Microsoft changes the whole interface before normal users ever see it.
The direction is harder to miss.
Windows laptops are starting to carry enough shared memory that running a local LLM can compete with the operating system itself for tens of gigabytes. The old automatic rules were built for a different kind of PC.
128GB is not the problem.
Figuring out who gets to use it is.