# Your PC will have local AI thanks to Intel, AMD and NVIDIA, but it will cost you 4,000 euros

> Intel, AMD and NVIDIA bring local AI to the PC, but the shift will cost 4,000 euros. Is it worth paying this price for your privacy?
> By José Maldonado · AI Agents · 2026-06-24

There was a time when opening an artificial intelligence application meant sending a packet of data across the ocean, waiting for a response from a massive server powered by a hydroelectric plant in Oregon and seeing the result on screen. That time is dying. The hardware industry has decided that the future of computing will not be built exclusively in the cloud, but in the silicon that breathes directly beneath our keyboards.

The AI PC era has ceased to be a mere marketing slogan and has become a war of architectural trenches. NVIDIA, AMD and Intel are redesigning integrated System-on-Chips (SoCs), not only chasing the brute force of traditional hertz, but seeking to democratize local AI for the mass consumer. The ultimate goal is ambitious: **a computer that not only executes commands, but anticipates intentions without losing a single millisecond of latency or compromising user privacy, and at the same time, with a price premium that will empty many pockets.** 

## Local AI on your PC

The transition toward processors equipped with high-performance Neural Processing Units (NPUs) and unified memory architectures is redefining the limits of personal computing. And in this race, each technology titan has deployed its best cards in this long-distance race.

**[Why Copilot is turning Windows into an unbearable maze](https://observatorioblockchain.com/en/technology/copilot-esta-transformando-windows-en-un-laberinto-insufrible/)**

## NVIDIA: bringing AI to the desktop with RTX Spark

NVIDIA has decided to make a statement in the consumer sector with the presentation of its [**RTX Spark superchip**](https://nvidianews.nvidia.com/news/nvidia-microsoft-windows-pcs-agents-rtx-spark). Breaking with the traditional dependence on the x86 ecosystem in consumer laptops, it is a system based on the ARM architecture developed in collaboration with MediaTek. This SoC combines a high-efficiency microprocessor and integrated graphics that reach a staggering 1 PFlop of AI performance (in FP4 sparse format) and up to 128 GB of unified memory.

NVIDIA's intention is clear: to allow thin laptops and compact desktop computers to run Large Language Models (LLMs) of up to 120 billion parameters locally with contexts of up to one million tokens. This is complemented by previous professional offerings such as DGX Spark, based on the Blackwell data center architecture (GB10), which works as an "AI lab in a box" focused on local creation and prototyping.

But here we need to tone down the hype NVIDIA has used to sell itself as "revolutionary." First, its chip, powerful without a doubt, is far from the most powerful SoC created for a PC; that place belongs to AMD, with its [**AMD Instinct MI300A APU**](https://www.amd.com/en/products/accelerators/instinct/mi300/mi300a.html), presented in 2023. This little beast has a total power of 4 PFlop in FP4, albeit with a consumption of 700 Watts, almost double the power of RTX Spark, due to the use of x86 for its cores.

## AMD: The strength of Strix Halo and the maturation of ROCm

For its part, AMD has bet heavily on its Ryzen AI Max+ Pro 395 processor line, known internally under the codename "Strix Halo." This colossal APU integrates 16 CPU cores with the Zen 5 architecture (x86), combined with Radeon 8060S graphics and a dedicated XDNA 2 NPU capable of delivering 50 TOPS natively, and which becomes direct competition for RTX Spark on the desktop.

Like NVIDIA, AMD's great secret lies in the inclusion of up to 128 GB of unified LPDDR5X memory with a 256-bit bus and a bandwidth of 256 GB/s. AMD's competitive advantage lies in its software ecosystem: Strix Halo inherits the same ROCm and HIP code stack used in its industrial accelerators, facilitating a seamless transition for developers going from the home desktop to supercomputing clusters.

## Performance for everything

Thanks to this, AMD has created a product capable of running Qwen 3.5 122B in FP8, a SOTA (best in the industry) AI at almost 10 tokens/second, with significant room for improvement thanks to the development of new kernels more optimized for ROCm, something the well-known hacker GeoHot, for example, is working on with his project [**tinygrad**](https://tinygrad.org/).

But the most relevant thing about this is that AMD has managed to reach such a level of performance in a chip that consumes 120 Watts at full power, which makes it possible to put it in a real laptop. This is unlike RTX Spark, where we still don't have real consumption data, but due to its structure and power, it is clear that its most powerful version will be above those levels by at least a factor of 2 (about 240 Watts at full power).

**[Open hardware: who controls your computer's chip?](https://observatorioblockchain.com/en/technology/hardware-libre-quien-controla-el-chip-de-tu-ordenador/)**

## Intel: Standardization and Mass Adoption

Finally, Intel, the historic giant of personal computing, focuses its efforts on standardization and volume. With its latest generations of processors, the Santa Clara firm has integrated dedicated NPUs comfortably exceeding the 40 TOPS threshold required by the industry for native execution of AI tasks.

But Intel's strategy is not only about competing at the pinnacle of ultra-enthusiast performance, but about ensuring that every conventional laptop sold on the market has the capacity to process local cognitive models without requiring a dedicated graphics card that drives up energy consumption.

And in that task there is a key player: Elon Musk. [**Musk and his Terafab project**](https://www.reuters.com/business/autos-transportation/intel-join-musks-terafab-mega-ai-chip-project-2026-04-07/), has become the biggest driver of development and investment for Intel to finally make its A18 and A14 chip manufacturing nodes successful. If achieved, Intel would finally be at the forefront of chip development, a place that Taiwanese TSMC snatched from it more than 10 years ago.

All this seems to be going well, since Intel and Musk have announced that the A18 node is already ready for production, and the A14, A10 and A7 nodes are already in development and [**will be ready for production in 2028, 2030 and 2032**](https://www.techpowerup.com/349245/intel-14a-node-enters-risk-production-in-2028-10a-and-7a-nodes-on-the-roadmap), respectively. And that only means one thing: Intel will be able to produce more powerful AI chips with lower consumption, something desktop and laptop users interested in this technology will especially appreciate.

## Microsoft and the Linux community react

Of course, a revolutionary architecture built for AI is nothing without an operating system capable of managing it. On this front, the responses reflect two deeply opposing philosophies.

## Microsoft and its bet on Windows on ARM

Microsoft has understood that the paradigm of local AI agents requires a structural change in the core of its operating system. Its close alliance with NVIDIA to provide native support for chips like RTX Spark under Windows on ARM is proof of this.

To give real utility to this power, the company is deploying NVIDIA OpenShell and new security and containment primitives in Windows. The goal is to allow personal AI agents to run in closed and isolated environments within the main machine, ensuring that the assistant accesses the user's sensitive data to anticipate everyday tasks without telemetry escaping to the cloud. Likewise, the professional software ecosystem is already responding to this change: Adobe is redesigning iconic tools such as Photoshop and Premiere from scratch to squeeze the performance of these unified memories.

**[The Linux revolution in cinema: from Titanic to the oceans of Avatar](https://observatorioblockchain.com/en/technology/la-revolucion-de-linux-en-el-cine-de-titanic-al-oceano-de-avatar/)**

## Linux: between maturation and fragmentation

Meanwhile, the penguin ecosystem has a unique opportunity, which comes with the traditional growing pains of open source software. On the one hand, the progress of projects such as AMD's ROCm stack on Linux has been a massive relief, allowing Linux users and engineers to compete face to face with the historical hegemony of NVIDIA CUDA in a fully open-source environment.

However, the community's real challenge lies in the fragmentation of NPU drivers. While Microsoft works hand in hand with manufacturers to unify APIs, development on Linux depends on the maturation of kernel patches that implement native and efficient support for Intel's integrated accelerators and optimized ARM cores. The system's developers prioritize the implementation of strict privilege separation, seeking to ensure that the execution of open source models (such as Llama or Mistral) is done directly on local hardware with absolute technical auditability and without depending on external corporate services.

## The Apple mirror

But why is all this happening? That is a very accurate question, and it has an easy answer: **Apple and its new Macs.** What the PC industry is experiencing is not an invention that came out of nowhere; it is a massive response, five years late, to the masterstroke initiated by Apple.

When the Cupertino firm presented the Mac M1 series (a transition that today extends with maturity to the current M5 generation), it broke the rules of personal computing by massively introducing the concept of Unified Memory Architecture (UMA).

In a traditional x86 computer, the CPU and the dedicated GPU have separate memories (RAM and VRAM) communicated through the PCIe bus. This physical bridge generates a massive bottleneck: for a graphics card to process an AI parameter, the data must be loaded into RAM, travel over the PCIe bus and be stored in VRAM.

## A unique solution

Apple solved this by placing the memory directly next to the SoC's silicon. CPU, GPU and the Neural Engine share the same block of high-speed memory. If an AI model takes up 30 GB, Apple's GPU can access it instantly, eliminating transfer latency.

Both NVIDIA's RTX Spark and AMD's Strix Halo, as well as Intel's future offerings, replicate this engineering philosophy almost identically. By integrating massive 256-bit buses and equipping machines with up to 128 GB of shared-access memory, the PC ecosystem can finally look Apple's computers in the eye. The race is no longer about how many extra frames per second can be squeezed out of a latest-generation video game, but about how many tokens per second the computer of a writer, a programmer or a student can process completely disconnected from the network.

**[Why an AI token has nothing to do with a Blockchain one](https://observatorioblockchain.com/en/technology/diferencia-token-ia-token-blockchain/)**

## A computer that becomes personal again

With all this, the PC is experiencing its era of greatest technical introspection in the last two decades. By adopting the philosophy of high-efficiency integrated SoCs and massive unified memories, the ecosystem is preparing for an inevitable future.

A tomorrow where private data does not constantly travel over the network, where digital privacy is once again protected by the physical walls of our homes and offices, and where machines finally live up to their original name: being truly personal computers.

Of course, all of this comes at a cost. NVIDIA's RTX Spark in its most powerful version will not be below $5000, AMD's Strix Halo will be in the $4000 range and Intel, for now, does not have similar hardware, but its future developments will possibly be in a similar range.

Yes, your computer will be more personal, you will have monstrous power to play, you will have local AI, but the price will initially make reaching that prohibitive. It will take some time and for the market to regulate the price so that we can access this technology, something that will take a couple of years, years that will be the most transformative in the sector since the arrival and mass adoption of the Internet.
