In the same 24-hour window in early June, Nvidia and Microsoft delivered a shared message to the market: AI agents may eventually need hardware designed specifically for them. The move signals a larger shift in how companies may think about devices, security, and where intelligence should run.
For enterprise buyers, the question is no longer limited to what a device can do. It also extends to whether the heaviest AI work belongs on the endpoint, in a managed terminal, or in the cloud.
Nvidia’s local-first wager
Nvidia is positioning RTX Spark as the base of a new Windows PC category built around a personal assistant. The compact system combines an Arm-based Grace CPU with a Blackwell GPU and uses Nvidia’s NVLink interconnect to deliver up to 128GB of unified memory.
The company says the platform can run models with roughly 120 billion parameters and handle context windows of up to 1 million tokens without repeated trips to the data center. That makes RTX Spark a direct bet that serious AI work can happen locally on the desktop or laptop.
Nvidia is also preparing RTX Spark in laptop and small desktop formats from Asus, Dell, HP, Lenovo, Microsoft Surface, and MSI. The message is clear: the PC is being recast as a working partner rather than a simple productivity machine.
Microsoft’s cloud-connected alternative
Microsoft is taking a different path with Project Solara. Instead of a new chip platform, it is presenting a software layer based on a lightweight operating system derived from the Android Open Source Project, called Microsoft Device Ecosystem Platform.
That platform runs on reference hardware using chips from Qualcomm and MediaTek. Microsoft has also shown two enterprise concepts for the program, a desk companion and a wearable badge.
Solara is built around what Microsoft calls just-in-time interfaces. In practice, the agent creates the screen needed for a task, which means developers do not have to redesign applications for every device shape.
A crowded race beyond the data center
The battle over AI infrastructure has already been intense in the data center, where Nvidia, Google with its Ironwood tensor processing unit, and AWS with Trainium are all competing to train and run models behind agentic systems.
RTX Spark and Project Solara move that competition closer to the user, directly onto the devices people touch every day. Microsoft is not alone in chasing the agent platform idea, with Google, Salesforce, and OpenAI also pursuing their own approaches.
Even so, Microsoft stands out for arguing early that agents deserve dedicated hardware that is neither a phone nor a traditional PC. That is what makes Solara intriguing, while also leaving its category still undefined.
Two theories about where the heavy work should happen
The deeper split between the two companies is architectural. Nvidia wants the heavy computation to stay local, which is why so much of RTX Spark’s silicon is devoted to a powerful GPU.
Microsoft’s model is closer to chip-to-cloud. In that design, low-power devices handle input and security, while the harder inference work runs in Azure.
In short, Nvidia sees the device as the main engine. Microsoft sees it as a secure front door to intelligence that lives elsewhere.
The doubts are still substantial
Not everyone is convinced that AI agents require new endpoint hardware. Ben Thompson of Stratechery has argued that the ideal local setup for agents is a strong CPU that still depends on the cloud for inference.
From that view, RTX Spark may dedicate die space to GPU cores that still cannot match the memory and bandwidth of data center systems. Buyers could end up paying a premium for local power that cloud infrastructure still delivers more efficiently.
Microsoft’s own design also leaves room for skepticism. If Solara devices ultimately call a cloud-based agent to perform the main task, the hardware on the desk may look more like a managed terminal than an AI workstation.
Microsoft itself says the program is still early. The effort is currently a pilot, and broad availability is not expected until late 2027 or 2028.
What enterprises should watch
For decision-makers, the practical takeaway is to separate orchestration and security from inference. Both announcements are strongest on the first layer, but much less settled on the second.
The benefits are real enough on paper. Local context, lower latency, data that stays inside the building, and centralized device management all have clear value for enterprise workflows.
Still, running frontier models at a desk remains a hard sell while the cloud keeps the edge for the largest systems. The most useful next step is to watch the pilots, not only the product stages.
If deployments at companies such as AccuWeather, CVS Health, and Target show lower handling time or fewer errors in specific tasks, dedicated agent hardware could become part of budget planning for 2027. Until then, the safer reading is that agents will first change how people use the devices they already have.