Upscale’s Token Fabric targets the network bottleneck in mixed-chip AI data centers

The story

Upscale has introduced Token Fabric, a proposed networking stack for artificial-intelligence data centers that want to use accelerators from more than one supplier. The Santa Clara startup says the architecture will connect processors inside tightly coupled groups and then link those groups across a facility through one operating and management layer. Reuters independently reported the launch on October 8 and said the first component is expected in the fourth quarter, with the broader platform rolling out in stages through 2027.

The product is not a single switch. Token Fabric combines four layers: Upscale's SkyFabriX scale-up silicon and switch trays; scale-out systems built with Nvidia Spectrum-X Ethernet technology; the SkyOS network operating system; and SkyCMD, software intended to orchestrate and observe thousands of network elements. Customers may buy the silicon, integrated systems, software or a fuller stack, according to the company announcement.

Scale-up and scale-out solve related but different problems. Scale-up links accelerators within a pod so they can behave more like one large computing system, where latency and predictable data movement are critical. Scale-out connects pods and racks across a data center, letting jobs use more machines and reach storage or other services. A weakness in either layer can leave expensive processors waiting for data instead of producing model-training or inference results.

Upscale says SkyFabriX will deliver 115.2 terabits per second of switching capacity and support standards including Ethernet/IP, OCP ESUN and UALoE, with support for evolving protocols such as UALink. Its scale-out equipment is planned in 400-gigabit and 800-gigabit configurations moving toward 1.6 terabits per second. These figures are company specifications and roadmap targets, not independently verified performance results. The release does not provide workload benchmarks, end-to-end latency, power consumption, pricing or a list of production customers.

That distinction is important because Upscale describes the design as open and suitable for any GPU, XPU or network interface card. The ambition is technically meaningful: cloud operators increasingly want specialized processors for different workloads without building a separate network and management system around each supplier. But interoperability is more than a standards checklist. Buyers will need proof that accelerators from different vendors can share the fabric while preserving reliability, collective-communication performance and fault recovery at scale.

There is also a strategic tension inside the pitch. Token Fabric is designed to reduce dependence on isolated, vendor-specific networks, yet its scale-out systems use Nvidia Spectrum-X silicon. That does not make the platform closed, and Nvidia publicly endorsed Upscale's use of the technology. It does mean the strength of the open-architecture claim will be judged by the interfaces around the silicon, the range of third-party accelerators validated and whether customers can substitute components without losing support or operational visibility.

The company plans general availability in early 2027 and says early-access and joint-validation programs are underway. Reuters reported that CEO Barun Kar expects Token Fabric revenue in 2027 to reach tens of millions of dollars and potentially the low hundreds of millions. That forecast is forward-looking. Upscale's own release cautions that development schedules, manufacturing, supply chains and performance may differ from current expectations.

Upscale has the financing to attempt a full-stack launch. In June it raised $190 million in a Series A-1 extension led by Premji Invest, taking disclosed funding to $500 million and the company's valuation to $2 billion. Nvidia, Salesforce Ventures, Temasek and Seligman Ventures joined that round. Funding and ecosystem endorsements indicate strategic interest, but they do not establish that Token Fabric can meet its performance or adoption targets.

INNOVOX analysis: the announcement reflects a broader shift in the AI infrastructure contest. The scarce resource is no longer simply the fastest accelerator; it is useful compute delivered by an entire system. Networking determines how much of the installed silicon remains productive, how quickly failures are isolated and whether operators can combine hardware generations without redesigning the facility. A credible open fabric could widen customer choice and reduce integration work. A stack that is open mainly in specification, but narrow in validated deployments, would deliver far less leverage.

What to watch now is execution rather than another specification. The strongest evidence would be named customers running mixed accelerators, independent comparisons against established Ethernet and proprietary interconnects, and published measurements for latency, throughput, congestion behavior and energy use. Delivery of the first scale-out component in the fourth quarter will test the schedule. Broader validation through 2027 will show whether Token Fabric becomes a practical multi-vendor foundation or remains an ambitious roadmap.

INNOVOX analysis

Token Fabric is a bet that the next point of control in AI infrastructure will be the network connecting accelerators, not only the accelerators themselves. Its practical value will depend on whether customers can mix hardware without losing the performance and operational simplicity that vertically integrated systems provide.

What to watch

Watch for Q4 hardware deliveries, verified interoperability with non-Nvidia accelerators, independent latency and throughput tests, power-efficiency data, named deployments and evidence that the advertised open standards work across vendors in production.