Upscale AI launches Token Fabric to link rival AI chips

Key Takeaways

Token Fabric is designed to connect GPUs, custom accelerators and network cards from multiple suppliers.
Upscale AI combines its own scale-up networking chips with Nvidia’s Spectrum-X Ethernet technology for scale-out connectivity.
The first component is planned for Q4 2026, followed by a staged expansion through 2027.

Upscale AI has introduced Token Fabric, a networking platform intended to help neoclouds and hyperscalers build AI data centers with processors and network cards from different suppliers. The launch addresses a growing infrastructure challenge: companies increasingly want access to several accelerator ecosystems, but the networking that binds those chips together can create technical and commercial lock-in.

The idea is straightforward, even if the engineering is not. Token Fabric is designed to connect GPUs, custom AI accelerators and network cards across data centers, reducing customers’ dependence on any single chip supplier. Reuters reported that the first component is planned for release in Q4 2026, with additional elements scheduled to arrive in stages through 2027.

AI clusters rely on two broad layers of connectivity. Scale-up networking links accelerators within a tightly coupled computing system, where bandwidth and latency can have a substantial effect on workload performance. Scale-out networking connects larger numbers of systems across racks and data-center environments. Token Fabric uses Upscale AI’s own networking chips for the scale-up layer, while Nvidia’s Spectrum-X Ethernet technology forms part of its scale-out design.

That combination is notable because Nvidia is both an investor in Upscale AI and a major force behind vertically integrated AI infrastructure. Nvidia’s proprietary NVLink is an important alternative for customers building tightly connected GPU systems. As Fortune has observed, this leaves Upscale AI pursuing a broad networking role while operating alongside an investor whose technology is central to the current AI market.

While heterogeneous infrastructure is appealing on a purchasing spreadsheet, making different accelerators behave like one coherent computing resource is much harder. Chips may use different memory architectures, software environments and communication methods. Networking can reduce some barriers, but customers will still need to consider workload scheduling, management software and application compatibility. Can an open fabric offer enough performance without recreating lock-in elsewhere in the stack? That is likely to become a key evaluation point.

Token Fabric is reported to support UALink and Ethernet for Scale-Up Networking, or ESUN. UALink is backed by a consortium that includes AMD, Intel, Google, Meta, Microsoft and more than 80 other companies. The breadth of that membership reflects industry interest in standardized accelerator interconnects that can compete with proprietary approaches. Coverage carried by the Economic Times also positions Upscale AI’s launch within the push to connect chips from rival suppliers.

For infrastructure buyers, the commercial implications could matter almost as much as the technical details. A credible multi-vendor fabric can give operators more flexibility when choosing accelerators for training, inference or specialized workloads. It may also improve procurement leverage and let data-center operators introduce newer chips without redesigning the entire network. That said, adoption will depend on demonstrated performance, reliability and management at production scale.

Upscale AI enters this phase with substantial financial backing. A $190 million funding extension in June 2026 brought total funding to approximately $500 million and valued Upscale AI at $2 billion. The capital gives Upscale AI room to develop silicon, software and a broader ecosystem, all of which tend to require long investment cycles.

The rollout schedule now becomes the practical test. Initial availability in Q4 2026 should offer the first evidence of how Token Fabric performs, while the staged expansion through 2027 will show whether Upscale AI can turn interoperability standards into deployable infrastructure. If customers begin mixing Nvidia, AMD and custom accelerators more aggressively, networking could become one of the most contested control points in the AI data center.

Editorial oversight by Rich Tehrani — CEO of TMC and creator of ITEXPO.

Disclosure: Rich Tehrani is also CEO of RT Advisors and a Registered Representative offering securities through Four Points Capital Partners LLC (Member FINRA/SIPC). Nothing in this article is an endorsement or recommendation to buy or sell any security or sector mentioned. RT Advisors may have or have had business relationships with companies mentioned in TMC Insight articles. Such relationships do not affect editorial judgment or coverage.