A company called Upscale is building what it calls a "Token Fabric" — a networking layer designed to let AI accelerators from different vendors work together over a better network, as reported by Kingy AI.

That single sentence carries the whole pitch. Upscale's Token Fabric wants different AI chips to work on a better network, and the company is betting that the next bottleneck in AI infrastructure isn't the accelerator itself. It's the wiring between them.

The idea targets a real tension in the datacenter. Large AI clusters today lean heavily on single-vendor stacks, and operators mixing GPUs from one maker with accelerators from another pay a integration tax in software, drivers and network plumbing. Upscale's pitch is to abstract that mess into a fabric that treats compute as fungible.

The reported framing — "different AI chips" and "a better network" — points at two problems at once. Heterogeneous compute is one. Interconnect bandwidth and latency, the classic wall that cluster operators hit as they scale training and inference jobs, is the other.

The source material does not specify process nodes, funding amounts, named customers or shipping timelines. No capacity figures were given. No revenue or valuation numbers appear in the report.

What the report does establish is the ambition: Upscale is positioning the Token Fabric as middleware for multi-vendor AI compute, sitting between the model workload and a mixed fleet of silicon.

Whether that lands depends on adoption. Interconnect and scheduling layers live or die by who integrates them, and the report names no partners yet. Upscale will need silicon vendors or hyperscalers to commit to the fabric before the pitch becomes a product.