For nearly 40 years, using a computer meant doing the same basic thing. You clicked something, the computer responded, and then it waited for you to click again. You opened an app, typed a document, closed it, and opened another app. The machine was always reactive, always waiting for your next command. It never thought ahead. It never worked without you. That era is ending.
At ‘Computex 2026’, the world’s largest annual technology trade show held in Taipei, Nvidia and Microsoft jointly announced what many analysts are calling the biggest shift in personal computing since the invention of the graphical user interface (GUI).
The centrepiece of this announcement is a new chip called the ‘Nvidia RTX Spark’, a piece of silicon so powerful that it can run on your laptop AI systems that, until recently, required a room full of servers. The promise being made to consumers is simple but radical. Your next computer will not just respond to you. It will work for you.
The problem with the old design
To understand why the RTX Spark matters, it helps to understand what has been holding computers back.
A traditional laptop or desktop is built like a relay race. The central processor (the CPU, think of it as the computer’s brain), and the graphics chip (the GPU, responsible for rendering images and, more recently, running AI tasks) are separate components sitting on a circuit board, passing information back and forth along a channel called a PCI Express bus.
It works, but it is slow by the standards of what modern AI demands. Every time the CPU needs to hand something to the GPU, or vice versa, there is a delay, like a relay runner pausing mid-stride.
Worse still, graphics chips come with their own dedicated memory, typically eight to 16 gigabytes on a consumer laptop, and that memory is separate from the main system memory. If you try to run an AI model that is larger than the GPU’s memory allowance, the system slows to a crawl as it struggles to shuffle data between the two pools.
These bottlenecks are the reason that running a sophisticated AI model locally, on your own device without sending your data to a remote server, has historically been impractical for ordinary users.
What the RTX Spark actually does
The RTX Spark scraps this inherited design and starts fresh. Rather than placing a CPU and GPU on separate chips connected by a slow bridge, it fuses them together into a single package using a technology called NVLink-C2C, a high-speed internal connection that allows the two processors to communicate roughly as fast as if they were a single unit.
More importantly, the RTX Spark introduces a unified memory pool of up to 128 gigabytes that both the CPU and GPU can access simultaneously and at equal speed. There is no relay race. Both processors reach into the same shared space and pull what they need instantly.
The memory operates at 300 gigabytes per second, meaning the entire contents of a standard hard drive can be moved through it in under a second. This is what makes it possible for a three-pound laptop to locally run an AI language model with 120 billion parameters and a context window of one million tokens. To put those numbers in human terms, you could feed the chip the entire text of roughly 750 novels, and have its reason across all of them at once, on battery, on a plane.
The chip itself is built on TSMC’s three-nanometre manufacturing process, among the most advanced in the world, and features a 20-core Arm-based CPU paired with a Blackwell-architecture GPU containing 6,144 processing units.
The AI performance figure Nvidia quotes is one petaflop, one thousand trillion mathematical operations per second. It is a number that belonged to national supercomputing facilities not very long ago.
Why this requires a new operating system
Powerful hardware only matters if software can use it properly. This is where Microsoft’s role in the partnership becomes essential, and where the history of Windows on Arm becomes relevant.
Arm-based chips, the architecture used by Apple Silicon and most smartphones, have been available in Windows laptops for years, and they have largely underperformed. The reason was software compatibility.
Most Windows applications were written for a different chip architecture (x86, used by Intel and AMD), and running them on Arm required a translation layer that introduced slowdowns and compatibility failures. Many apps simply did not work.
Microsoft has spent three years, in close collaboration with Nvidia, rebuilding the relevant parts of Windows 11 to function natively on the RTX Spark’s architecture. Two changes stand out.
The first is a new scheduling system called Workload Profile Scheduling (WPS). Modern chips have different types of processing cores, some optimised for efficiency, some for raw power. WPS, built into the Windows kernel, watches what each application is doing in real time and routes tasks to the appropriate cores automatically. Background tasks go to the efficient cores, while intensive AI work gets the powerful ones. The user does not have to configure anything.
The second is a thoroughly overhauled translation layer called Prism, which now uses specialised instruction sets to run legacy x86 applications with near-native speed. Software you already own should work without modification.
Microsoft also worked with gaming companies, including Riot Games (which makes League of Legends and Valorant) and Krafton (maker of PUBG), to bring their titles natively to the Arm platform.
Critically, the anti-cheat software that these games rely on has also been ported, removing what was previously the single largest barrier to competitive gaming on Arm-based Windows devices.
What an ‘Agent’ actually means
Much of the announcement centred on a word that is currently doing heavy lifting in the technology industry. ‘Agent’. It is worth being precise about what this means.
Current AI tools are what technologists call transactional. You send a message, the AI responds, and then it waits. Each exchange is essentially independent. The AI does not take initiative, does not monitor your files, does not complete a multi-step task over several hours while you do something else.
An agentic AI behaves more like an autonomous assistant with genuine initiative. You might ask it to go through your emails from the last month, find every invoice, organise them by supplier, flag anything overdue, and prepare a summary report.
A traditional AI would help you think through how to do that. An agent would simply do it, opening applications, reading files, processing data, and returning with a finished result, without needing you to supervise each step.
The leading open-source framework enabling this kind of behaviour is called OpenClaw, developed by Austrian engineer Peter Steinberger. It allows agents to access local file systems, run applications, interact with databases, and execute complex workflows automatically.
Enterprise software companies are moving quickly in the same direction. SAP is embedding Nvidia’s agent execution software into its business planning tools, and ServiceNow is building agents that can automate corporate workflows directly on employees’ devices.
Jensen Huang, Nvidia’s chief executive, summarised the shift with characteristic bluntness at his Computex address.
“For 40 years, you launched apps. Click. Type. With RTX Spark and Microsoft Windows, you ask, and the PC does the work,” the tech boss remarked.
Security problem that comes with autonomy
Giving software broad access to your files, applications, and network connections is, of course, a significant security risk. Cisco’s AI security research team demonstrated earlier 2026 that malicious instructions hidden inside an innocent-looking document could trick a locally running agent into exposing private user credentials, a technique called prompt injection.
Separately, a computer science student named Jack Luo discovered that his local OpenClaw agent had autonomously created a dating profile on a service called MoltMatch, and begun screening prospective partners without his knowledge.
These cases represent genuine vulnerabilities that become more serious as agents acquire more autonomy and broader system access.
Nvidia’s response is a sandboxed runtime environment called OpenShell. When an agent runs inside OpenShell, every action it attempts, whether accessing a folder, sending a network request, or running a piece of code, is intercepted and checked against a policy file that the user controls.
If the action is not explicitly permitted, it is blocked. If a cloud service needs to be queried, a built-in privacy router automatically strips out personal identifying information before the request leaves the device.
When an agent is blocked by a policy, it explains why to the user and suggests what permission change would allow it to proceed, requiring manual approval before continuing. Security without losing usability is the stated goal, though the practical robustness of this system will only become clear once it is in the hands of a large user base.
The economics of not paying per word
There is a straightforward financial argument for local AI that often gets obscured by the technical discussion. Running AI through cloud services costs money, and for heavy users, it costs a lot of money. Cloud providers charge per token, roughly per word, for AI inference.
A corporate customer-service system handling a thousand conversations a day can generate upwards of 60 million tokens a month. Scaled across a large organisation, cloud AI fees can run to tens of thousands of dollars monthly.
On a local device with the RTX Spark, inference costs nothing beyond the electricity used. Once the hardware is purchased, every query is free. For individuals spending more than a hundred dollars a month on cloud AI subscriptions, a high-performance local workstation could pay for itself within a year.
Local AI also has practical advantages beyond cost. Responses come back in under ten milliseconds compared to 100 to 500 milliseconds for cloud responses. It works without an internet connection, and sensitive data never leaves the device.
The trade-off is the upfront cost. The premium on high-density unified memory has been driven up further by a global memory shortage that the industry has taken to calling ‘RAMageddon’, making the top-specification devices more expensive than a typical consumer laptop.
The competitive landscape
The RTX Spark does not arrive in an empty field. Apple has been shipping unified-memory Arm-based chips in its Macs since 2020, and its M5 family, refreshed in early 2026, remains formidable, with the M5 Max offering memory configurations up to 128 gigabytes and bandwidth exceeding 400 gigabytes per second. Apple also introduced a new budget entry, the MacBook Neo, powered by an A18 Pro chip, and priced at $599.
Qualcomm has responded with its Snapdragon C platform, targeting budget laptops starting at $300, aimed at students and small businesses that want some AI capability without a premium price. AMD’s Ryzen AI Max 400 chips bring up to 60 TOPS of dedicated AI processing and large unified memory support to the Windows x86 ecosystem, directly targeting the MacBook Pro’s architecture. Intel’s Panther Lake Core Ultra X9 processors, featuring a powerful integrated Arc GPU, deliver up to 180 platform TOPS.
The honest summary is that 2026 is the year every major chipmaker decided, simultaneously, that AI processing belongs in the device you are holding rather than in a data centre you will never see.
What ships this autumn
Six premium laptops will launch with the RTX Spark this autumn, including Microsoft’s own Surface Laptop Ultra, Dell’s XPS 16 Creator Edition, Asus’s ProArt creative workstations, and machines from HP and Lenovo aimed at enterprise and developer users.
A further 30 laptop models and 10 compact desktop machines are planned to follow. Nvidia is also launching the DGX Station for Windows, a desktop workstation designed for AI researchers and enterprise developers who need performance beyond what a laptop thermal envelope can sustain.
The laptops arriving in 2027 will be faster, which is always true. But the fundamental nature of what a personal computer does, and who initiates the doing, is being renegotiated.
Whether the security architecture is robust enough, whether agents will be genuinely useful or merely gimmicky, and whether the cost of the hardware justifies the economics of local inference are all questions that will be answered in practice rather than in keynote presentations. But the direction of travel is unmistakable, and the silicon to make it real is now, finally, small enough to fit in a bag.