How the AI Race Is Changing PC Memory, NPUs and the Price of RAM

The AI PC changes what a minimum specification means

For years, memory recommendations were relatively easy to explain. More browser tabs, larger creative projects, games and professional software required more RAM. AI adds a different workload: a model may need to remain loaded while it processes text, audio, images or video, sometimes alongside the applications the user already has open. Memory is no longer only a buffer for conventional software. It becomes part of the space in which inference happens.

Microsoft’s requirements make this visible. Features such as Recall and Click to Do require a Copilot+ PC with a 40-TOPS NPU, 16 GB of RAM and 256 GB of storage. That is an official product threshold, not proof that 16 GB is ideal for every local model. It tells us that 16 GB has become the entry point for a defined class of Windows AI experiences. Larger independent models, development tools or simultaneous workloads can require substantially more.

An NPU is an accelerator, not a container

A neural processing unit is designed to execute AI operations efficiently, particularly sustained inference that would consume more power on a general-purpose CPU or a large GPU. TOPS measures a form of computational throughput. It does not state how large a model the machine can hold, how much context can remain active or how many other applications can run at the same time.

This distinction matters because marketing can make the NPU appear to be the whole AI system. In practice, local inference involves compute, memory capacity, memory bandwidth, storage, software runtimes and model optimization. A fast accelerator can still wait for data if the memory system cannot supply it efficiently. It can also be unable to run a model at all when the model and its working data do not fit within available memory.

Capacity answers the first question: does the model fit?

Model parameters require memory. Quantization can reduce the number of bits used to represent each parameter, allowing a model to occupy less space, but it does not make memory irrelevant. The runtime also needs space for context, temporary calculations and the operating system. A model whose file appears to fit may still fail or perform poorly once the complete workload is considered.

This is why local-AI demonstrations increasingly specify memory alongside compute. AMD says its Ryzen AI Halo platform can be configured with 128 GB of unified memory, while its 2026 Ryzen AI Max PRO 400 announcement describes systems with up to 192 GB. These are specialist developer and workstation configurations, not a new requirement for a typical laptop. They demonstrate the relationship between capacity and ambition: more memory makes larger models and more concurrent agents technically possible.

Bandwidth answers the second question: can data move quickly enough?

Capacity describes how much data can be held; bandwidth describes how quickly it can be moved. AI inference repeatedly transfers model weights and other data to processing units. When the workload is memory-bound, adding theoretical compute does not produce a proportional improvement because the processor spends time waiting for information.

Modern AI PCs therefore combine NPUs and integrated GPUs with high-speed LPDDR or DDR memory. Intel describes its AI PC architecture as cooperation among CPU, GPU and NPU, supported by high-bandwidth memory and cache. Samsung has presented faster LPDDR5X as an on-device AI product because performance per watt and data movement matter in thin devices. The important specification is the complete path between storage, memory and compute, not one headline number.

Unified memory changes how capacity can be used

Traditional PCs often separate system RAM from dedicated graphics memory. A discrete GPU can process AI workloads rapidly, but the model must fit within its VRAM or be divided across slower paths. Unified-memory designs allow CPU, integrated GPU and sometimes the NPU to access a shared pool. This can make a larger portion of installed memory useful to local inference and reduce unnecessary copying.

Unified does not mean unlimited or automatically faster. The operating system, applications and AI workloads compete for the same pool, and vendors may reserve only part of it for graphics or AI use. Buyers should therefore distinguish installed memory from memory actually assignable to a model. Architecture, firmware and software support can be as important as the total printed on the box.

Context and agents can consume memory quietly

A model’s parameter count is not the only source of demand. Long context windows require a cache that grows as the conversation, document set or task history expands. An agent may also operate several models or tools, retain intermediate results and run a browser, coding environment or office application at the same time. The workload can become larger without the user ever downloading a visibly larger model.

This is one reason agentic computers may create a stronger case for 32 GB or more than a simple chatbot does. One small assistant performing a short task can be efficient. Several agents working concurrently over long documents may compete for memory with ordinary applications. The correct requirement depends on concurrency, context and model design, not on the word “AI” alone.

Local AI is not the same as cloud AI

Cloud services can give a modest PC access to very large models because inference runs in a data center. Local execution trades some of that scale for privacy, lower network dependence, predictable latency and the absence of per-token cloud charges. Microsoft’s Foundry Local and Windows ML reflect an effort to make local inference a standard application capability across GPUs, CPUs and NPUs.

The two approaches will coexist. A hybrid application may use a small local model for immediate or private work and call a larger cloud model for tasks requiring more capability. In that environment, not every user needs workstation-scale RAM. The useful question is which work must remain local and how large those local models need to be.

Why 16 GB is becoming a baseline rather than a universal answer

Sixteen gigabytes can support mainstream Copilot+ features and ordinary productivity, which is why it appears as an official minimum. It should not be interpreted as a guarantee for every future AI application. Once memory is soldered, the purchase decision becomes more consequential because the system may never be upgraded.

For a buyer using cloud assistants, email, documents and normal browsing, 16 GB may remain reasonable. A developer experimenting with local models, a creator using AI video or image tools, or a user expecting several local agents may benefit from 32 GB or more. Specialist 64–192 GB configurations address another category entirely. NTS does not recommend one capacity for everyone because the workload boundaries are too different.

Storage matters, but it cannot replace RAM

Models are stored on an SSD and loaded into memory when used. Fast storage reduces loading time and can support memory mapping or offloading, but it remains much slower than RAM. A system can sometimes run an oversized model by moving data between storage and memory, yet the performance penalty may make the experience impractical.

AI also increases persistent storage demand. Model files, local indexes, embeddings, media assets and snapshots can occupy tens or hundreds of gigabytes. Microsoft’s Recall requirements illustrate this separate pressure: the feature needs available storage and pauses snapshot saving when free space falls below a threshold. A balanced AI PC therefore needs both adequate memory for active work and adequate SSD capacity for models and data.

AI demand is changing the memory industry

Data-center accelerators depend on high-bandwidth memory, while servers, AI PCs and edge devices use other forms of DRAM and NAND. These products are not interchangeable, but they share manufacturers, capital budgets, engineering resources and parts of the production ecosystem. As AI infrastructure absorbs investment, memory companies are deciding how capacity should be allocated across markets.

Micron has described memory as strategically important in the AI era and announced record investment to respond to demand. SK hynix announced major new DRAM and NAND fabrication investments in August 2026, with facilities scheduled later in the decade. Those decisions show confidence in sustained demand. They also show why supply cannot expand instantly: advanced fabrication plants require enormous capital and years to build.

Does that mean consumer RAM will become more expensive?

Not necessarily. Strong AI demand can support higher memory prices when supply is tight, and producers may prioritize more profitable products such as HBM or server memory. But consumer DRAM prices also depend on PC shipments, inventories, manufacturing yields, contract negotiations, exchange rates and the normal boom-and-bust semiconductor cycle. New capacity or weaker device demand can ease pressure.

The responsible conclusion is conditional. AI has become a meaningful source of memory demand and capital allocation, so it can influence consumer pricing indirectly. It is not the only force. A prediction that “AI will make all RAM expensive” ignores product differences and market cycles. The stronger signal to monitor is whether conventional DRAM supply growth falls behind PC and device demand while manufacturers expand AI-focused products.

Memory upgrades are becoming less available

Many thin laptops use soldered LPDDR memory because the design saves space, reduces power consumption and can support high bandwidth. The trade-off is that users cannot add RAM later. This turns memory capacity from an inexpensive future upgrade into a decision fixed at purchase.

That matters more in a period of changing software requirements. A machine bought for cloud-based assistants may later be asked to run local transcription, search, image generation or agents. Buyers should check whether memory is soldered, whether any slot remains available and how much memory the platform can expose to its GPU or AI runtime. The cheapest configuration may have the highest long-term constraint.

What buyers should compare beyond the AI label

An “AI PC” badge cannot replace a specification review. Compare total RAM, upgradeability, memory type and bandwidth, NPU performance, integrated or discrete GPU capability, SSD capacity and the software actually supported by the platform. Ask whether the advertised AI feature runs locally, in the cloud or through a hybrid approach.

TOPS figures from different architectures may not translate directly into identical application performance. Model support and runtime maturity matter. A lower headline number with well-optimized software can be more useful than theoretical performance that applications cannot access. The same caution applies to memory: capacity is important, but bandwidth, allocation and software efficiency determine how effectively it is used.

What to watch through 2027

The clearest signal will be the amount of memory offered in mainstream rather than specialist computers. If 32 GB becomes common in mid-range AI PCs, software developers can design for a larger local baseline. If most systems remain at 16 GB, applications will need smaller models, aggressive quantization and hybrid cloud execution.

Also watch whether NPUs gain broader support beyond selected operating-system features, how much unified memory vendors expose to local models, and whether new memory standards improve bandwidth without damaging battery life. On the supply side, monitor DRAM capacity, not only HBM announcements. Consumer prices will reflect the relationship between production and demand across several markets.

The NTS View

The AI race is not simply adding another processor to the computer. It is changing the balance among compute, memory, storage and software. NPUs make sustained inference more efficient, but they do not eliminate the physical requirement to hold models and their working data. As local AI becomes more capable, memory capacity and bandwidth become part of the intelligence a machine can make available.

For mainstream users, 16 GB is becoming a credible starting point, not a promise of unlimited local AI. For development and heavier inference, 32 GB, 64 GB or unified-memory workstations can support very different ambitions. The market should resist presenting all of these machines as one category.

AI demand can influence RAM prices, particularly when manufacturers direct resources toward high-value memory and supply expansion takes years. But price is an outcome to observe, not a conclusion to assume. The durable change is clearer: memory has moved from a supporting specification to a strategic component of the AI computer.