Many of today's most capable open-weight AI models won't fit into the memory of a typical PC. Usually, running them means shrinking them, offloading the work to the cloud, or building a server cluster with specialised networking gear. Microsoft offers a different route: connect two or more Windows 11 PCs with USB4 cables and let their memory and GPU power work together on a single model.

Psst!Add AlexTech.ai asPreferred Source on Google

The feature is called RDMA over USB4 and is currently in developer preview. RDMA, remote direct memory access, lets machines in data centres exchange data quickly. Microsoft brings it to regular computers through the same programming interface Windows uses with datacenter network adapters, so compatible software can treat the linked PCs as one system.

official image of Windows USB4 feature

Image: Microsoft Learn — https://learn.microsoft.com/en-us/windows-hardware/design/component-guidelines/usb4-settings-enablement

No special network cards, switches or third-party drivers are needed, just USB4-capable PCs and the right cables. According to figures Microsoft shared, transferring 16 KiB of data takes about 50 microseconds, compared with 194 microseconds for Ethernet over USB4: nearly four times faster.

The system targets inference, meaning running an already trained model, and requires a compatible inference framework. Specialist outlets point to demanding setup requirements in the preview: test signing and Secure Boot turned off, so it is aimed at users comfortable with the system.

In practice, the feature makes it possible to run locally models that today would need a dedicated server or the cloud, keeping data on the user's own machines. The cost is a more complicated setup, for now within reach mainly of developers.

Microsoft has not announced a general availability date or pricing. The preview will show how well the approach holds up with models of real-world size.