NVIDIA-Backed Upscale AI Unveils Token Fabric to Connect Rival AI Chips

The race to build bigger AI systems is not only about buying more GPUs. It is also about moving data between them without leaving expensive processors waiting. On October 8, 2026, NVIDIA-backed networking startup Upscale AI introduced Token Fabric, a hardware-and-software platform intended to connect different brands and types of AI accelerators inside large data centers.

The pitch is simple: companies building AI infrastructure should not have to assemble a completely different networking stack every time they mix GPUs with other specialized AI chips. Upscale says its system brings those connections and the tools that operate them under one architecture.

This is a launch announcement, not proof that Token Fabric already outperforms competing systems in real-world deployments. The company says early-access and joint-validation programs are underway, with general availability planned for early 2027.

Read the October 8 company announcement and Reuters' independent launch report.

Why AI networking has become a bottleneck

When an AI model is spread across multiple accelerators, those processors must exchange data rapidly. A slow connection can make a powerful chip sit idle while waiting for the next piece of work. That matters during both training and inference, especially as companies build systems with many racks of hardware.

Engineers often describe two types of connections. Scale-up links accelerators closely within a pod or rack-scale group. Scale-out connects those groups across a larger data center. A platform that can handle both may reduce integration work for operators, although actual performance depends on the workloads, switches, protocols and software in use.

Token Fabric is aimed at hyperscale operators, specialist AI cloud providers and enterprises running sizable accelerator fleets. It is not a new graphics card, a home router or a feature that will directly speed up a consumer PC.

What Token Fabric actually includes

Upscale describes four key building blocks. SkyFabriX is its purpose-built scale-up switch silicon, with a claimed 115.2 Tbps switching capability. NVIDIA Spectrum-X-powered systems handle the scale-out side. SkyOS is the network operating system, and SkyCMD is the orchestration and monitoring layer intended to help operators manage large groups of network devices.

The company says the architecture supports open standards and different accelerator types, including GPUs and custom XPUs. That promise matters to cloud operators who do not want every future purchase to depend on one chip vendor. But 'open' is not the same as plug-and-play compatibility with every accelerator in every deployment. Actual interoperability will need to be validated with the specific hardware and network configurations customers use.

Upscale also outlines multiple purchasing paths: a full stack, integrated systems with optional management software, or individual silicon and software components for larger operators building their own designs.

Why NVIDIA is involved, even though rival chips are supported

NVIDIA is both an investor in Upscale and a supplier of the Spectrum-X technology used in Token Fabric's scale-out systems. That might seem contradictory for a product pitched at mixed-vendor computing, but the goal is to sell networking infrastructure to data centers that may contain several accelerator families.

For NVIDIA, it is another way its networking technology can be part of AI infrastructure even when the compute layer is not entirely NVIDIA hardware. For Upscale, the combination offers a route to market alongside its own scale-up silicon and software.

The important distinction is between the company's architectural claims and independently measured results. The launch materials describe high bandwidth and simpler operations, but they do not provide a neutral head-to-head performance comparison covering the wide range of mixed-chip installations customers might build.

When will customers actually be able to use it?

Upscale says early-access and joint-validation programs are active now. It plans general availability in early 2027, so the product is not yet broadly shipping to every data-center buyer. Reuters reports that the first elements will roll out before the full stack, with broader deployment proceeding in stages.

No public per-rack or per-port price list was announced in the sources reviewed. This is enterprise infrastructure purchased through vendor discussions, not a retail product with a Philippine peso sticker price. There is also no announced Philippine-specific release or reseller pricing to report.

Why this matters beyond the data center

The average reader will not buy Token Fabric. Still, the economics of AI services depend partly on how efficiently providers use their hardware. Better accelerator utilization could eventually help cloud companies serve more requests from the same installed capacity, although that does not automatically mean lower API prices or cheaper subscriptions.

For developers, the relevant longer-term question is whether mixed-vendor infrastructure becomes easier to deploy without a maze of separate network-management systems. For hardware enthusiasts, it is another reminder that GPU specifications alone do not determine the speed of a large AI cluster.

The next milestone to watch is third-party validation: which chip combinations are supported in production, what real latency and utilization improvements are measured, and whether Upscale meets its early-2027 availability target. Until those details are available, Token Fabric is best understood as a credible new infrastructure proposal with important claims still to be tested.

Company announcement | Reuters coverage.