Featured image: Network cabling illustrates data-centre interconnects. Contextual photograph; not Upscale Token Fabric equipment or a verified heterogeneous deployment. Photo: Taylor Vick / Unsplash. Unsplash licence.
An AI cluster is not only a collection of fast chips. The chips have to exchange data quickly enough that expensive accelerators do not spend their time waiting for the network. That problem becomes harder when a site wants equipment from more than one supplier.
Upscale introduced Token Fabric on 8 October 2026, presenting it as an open, full-stack network for heterogeneous AI computing. It combines short-range scale-up connections, Ethernet-based scale-out systems and network software. This is a commercial product announcement supported by company specifications, not an independent benchmark of a completed customer deployment.
Scale-up and scale-out solve different bottlenecks
Scale-up networking connects accelerators within a tightly coordinated system so they can share data and memory at very high speed. Scale-out networking connects many servers or racks, allowing a cluster to grow horizontally. Large AI workloads need both levels to behave predictably.
Proprietary links can deliver strong performance inside a vendor’s own platform, but they can also make mixing accelerators difficult. Upscale’s pitch is that standards-based interfaces can connect GPUs, other accelerators, memory and storage without forcing the entire cluster into one supplier’s architecture.
“Open” requires precision. A published or industry-governed interface can make components easier to combine. It does not guarantee that every device works together without tuning, that management tools expose the same features or that performance matches a vertically integrated design.
The portfolio joins silicon, switches and software
For scale-up, Upscale describes SkyFabriX silicon with 115.2 terabits per second of switching capacity and support for standards being developed through the Open Compute Project, including Ethernet Scale-Up Networking and Ultra Accelerator Link over Ethernet. Terabits per second is aggregate data-moving capacity, not the speed one model or one accelerator will automatically receive.
For scale-out, the company is building on NVIDIA Spectrum-X Ethernet switch silicon. Its March 2026 partnership announcement describes an AI-optimised version of SONiC, the open-source network operating system widely used by large data-centre operators.
Token Fabric adds management, telemetry and automation intended to treat the two network layers as one operational system. Telemetry is data about how the network is behaving—congestion, errors, delay and device health—which software can use to diagnose or adjust the fabric. The software layer may prove as consequential as the switch silicon because operators need one view of failures and congestion across both domains.
Heterogeneous compute is the hard promise
Heterogeneous means the cluster contains different types or brands of computing devices. The attraction is flexibility: operators could choose accelerators for particular workloads, avoid a single-vendor dependency or adopt new hardware without rebuilding every surrounding layer.
The cost is integration. Collective communications used in AI training must behave consistently across devices. Memory semantics, drivers, security, congestion control and failure recovery all have to align. Standards can define a common language while implementations still differ in maturity and efficiency.
Reuters reported that the launch is aimed at connecting chips from rival suppliers in AI data centres. The 8 October report provides independent context for the multi-vendor ambition, while the detailed performance figures remain vendor claims.
The decisive evidence will come from mixed clusters
Useful independent testing would identify the exact accelerators, workload, network topology and software stack. It would measure application throughput, latency under congestion, accelerator utilisation, power use and recovery after a link or switch fails.
A headline switching capacity cannot answer those questions. Nor does support for an emerging standard prove interoperability with every future product carrying the same label. Certification tests, customer deployments and reproducible benchmarks will be more informative than the architecture diagram alone.
Token Fabric is significant because it turns the open-networking argument into a portfolio spanning the close connections inside an AI system and the Ethernet fabric between systems. If it works as described, operators gain more freedom to mix hardware. The launch establishes the product direction; real heterogeneous clusters still have to establish the result.
Reporting and source checks completed on 12 October 2026. By FutureTechDose Editor.
Leave a Reply