Nvidia-Backed Upscale AI Launches Token Fabric to Connect Rival AI Chips
Nvidia-backed artificial intelligence networking startup Upscale AI has launched a new platform designed to help data-center operators connect AI chips from different suppliers, addressing one of the biggest infrastructure challenges emerging as companies diversify beyond a single accelerator vendor.
The platform, called Token Fabric, combines networking hardware and software to connect AI processors across a data center without requiring customers to build and manage separate networking systems for different types of chips. Upscale, which was valued at $2 billion in June 2026, is targeting hyperscalers, neoclouds and other operators building increasingly complex AI infrastructure.
The launch is significant because the AI infrastructure market is gradually moving toward heterogeneous computing — data centers using accelerators from multiple suppliers rather than relying entirely on one chip ecosystem.
What is Upscale AI's Token Fabric?
Token Fabric is essentially a networking architecture designed to make different AI accelerators work together more efficiently.
Upscale says the platform can connect accelerators inside a system and across the wider data center. Its architecture combines the company's own networking technology with software and systems based on Nvidia Spectrum-X Ethernet for scale-out networking.
The company says its software layer is designed to provide a unified operating and management environment across GPUs and other AI accelerators, which it refers to as XPUs.
That matters because AI data centers are becoming much more complicated. Operators may use GPUs, custom AI accelerators, networking processors and other specialized hardware depending on the workload, cost and availability.
Without a common networking layer, integrating these systems can require multiple technologies and software stacks.
Upscale's pitch is that Token Fabric can simplify that process.
Why connecting rival AI chips matters
The AI chip market has historically been dominated by Nvidia's integrated ecosystem.
Nvidia provides not only GPUs but also networking products, software and interconnect technologies designed to work closely together. That integration has been a major competitive advantage as companies build large AI clusters.
But demand for AI computing has become so large that hyperscalers and neocloud providers increasingly have incentives to consider alternatives.
AMD, custom accelerators developed by major cloud companies and specialized AI chips are all part of the broader shift toward a more diverse computing environment.
The problem is interoperability.
A data center operator does not simply need individual chips to perform well. Thousands of accelerators must communicate rapidly enough to keep expensive hardware fully utilized.
If processors spend too much time waiting for data to move between systems, the economic value of the computing hardware falls.
Token Fabric is designed to address precisely that bottleneck by providing a common networking framework for heterogeneous AI infrastructure. Reuters reported that Upscale's goal is to help keep expensive AI chips working rather than waiting for data.
Nvidia is both an investor and a technology partner
Nvidia's involvement makes the launch particularly interesting.
Upscale is backed by Nvidia, but the startup is also building a platform that can connect hardware from different suppliers. That may appear contradictory at first, but it reflects a broader opportunity around AI infrastructure.
Nvidia can benefit from the continued expansion of AI data centers even when customers deploy a more diverse mix of hardware, particularly when Nvidia networking technology remains part of those systems.
Under the Token Fabric architecture, Upscale's scale-out systems use Nvidia Spectrum-X silicon, while its own SkyFabriX technology is used for scale-up networking. The company says customers can choose individual silicon, systems, software or the complete stack.
This gives Nvidia exposure to an open and heterogeneous architecture without requiring every component of the AI factory to be Nvidia-made.
For Nvidia investors, however, the longer-term question is whether more heterogeneous data centers strengthen Nvidia's networking opportunity or gradually reduce the advantages of its tightly integrated ecosystem.
Upscale has raised $500 million so far
Upscale is no longer a small early-stage project with only a concept.
The company raised a $190 million extension to its funding round in June, taking total funding to approximately $500 million. The financing was led by Premji Invest, the investment arm of Indian technology billionaire Azim Premji. Nvidia, Salesforce and Singapore state investor Temasek are also among its backers.
The funding gives Upscale the financial resources to develop networking silicon, systems and software while competing in a market populated by much larger companies.
That competition is substantial.
The AI networking market includes Nvidia, Broadcom, Cisco and Arista, while other startups and semiconductor companies are also developing alternatives aimed at AI clusters.
Upscale is therefore attempting to establish itself at a critical point in the AI infrastructure stack before the market becomes even more entrenched.
The business opportunity could be substantial
Upscale believes the demand for AI networking will grow rapidly as AI clusters become larger and more diverse.
Its own announcement cited industry estimates projecting hundreds of billions of dollars in AI networking opportunities over the coming years. Dell'Oro forecasts nearly $1 trillion in AI back-end switch spending through 2030, while 650 Group expects AI networking to exceed $200 billion by 2030. These are industry forecasts, not Upscale revenue projections.
The underlying trend is easier to understand.
AI models are becoming larger, inference workloads are expanding and data centers need to connect increasingly large numbers of accelerators. Networking therefore becomes more important as computing capacity scales.
In traditional computing, networking was often viewed as supporting infrastructure.
In AI factories, it can become a direct determinant of how effectively expensive processors are used.
That creates an opportunity for companies that can reduce latency, increase bandwidth and simplify operations.
Upscale expects meaningful revenue from Token Fabric
Upscale CEO Barun Kar told Reuters that the company expects Token Fabric revenue next year to reach tens of millions of dollars, with the potential to grow into the low hundreds of millions.
Those numbers are expectations rather than established financial results, and the product is only now entering the market.
Upscale plans to release the first part of Token Fabric in the fourth quarter of 2026, initially connecting groups of AI chips across a data center. The complete platform is expected to roll out in stages through 2027. Its broader announcement says general availability is planned for early 2027, with early-access and joint-validation programs already underway.
Execution will therefore be critical.
The technology needs to work reliably at very large scale, and customers will have to be convinced that an open architecture can deliver performance comparable with more tightly integrated proprietary systems.
What this means for Nvidia
For Nvidia, the development is unlikely to be about Token Fabric becoming a major direct revenue contributor immediately.
The more important issue is the changing structure of AI infrastructure.
Nvidia has built a powerful competitive position by combining GPUs, networking, interconnects and software. But as customers look for more flexibility, open networking could become increasingly important.
Token Fabric potentially gives Nvidia another way to participate in heterogeneous AI data centers through its Spectrum-X networking technology, even when the compute layer contains chips from multiple suppliers. Nvidia itself highlighted Upscale's use of Spectrum-X as a way to support open and heterogeneous AI infrastructure.
At the same time, the rise of open standards could eventually make it easier for customers to mix competing accelerators.
That is a strategic trade-off worth watching.
Risks for Upscale AI
The opportunity is large, but so are the execution risks.
First, Upscale is competing against companies with enormous engineering resources and established customer relationships.
Second, open standards do not automatically guarantee better performance. AI customers care about latency, reliability, software compatibility and total cost of ownership, not simply interoperability.
Third, the company still has to move from funding and development into large-scale commercial deployments.
Finally, AI networking technology is evolving quickly. Standards such as UALink and other open approaches are still developing, meaning today's architecture may face new technical requirements as AI clusters change.
The company itself warns that its product roadmap and availability are subject to risks including development delays, manufacturing issues, supply-chain constraints and changing market conditions.
What investors should watch next
Because Upscale is privately held, the immediate investor impact is more relevant to Nvidia and the broader AI infrastructure ecosystem than to Upscale's own stock.
The most important developments to monitor are:
Customer deployments of Token Fabric
Adoption by hyperscalers and neocloud providers
Performance compared with proprietary networking architectures
Growth of multi-vendor AI clusters
Expansion of Nvidia Spectrum-X
Adoption of open AI networking standards
Upscale's progress toward commercial revenue
Competitive responses from Broadcom, Cisco, Arista and other networking companies
If heterogeneous AI infrastructure becomes the norm, networking could become an increasingly important layer of the AI investment cycle.
Bottom line
Upscale AI's launch of Token Fabric is more than the introduction of another networking product. It reflects a potentially important shift in the AI data-center market toward systems that can combine chips from different suppliers.
The Nvidia-backed startup is betting that hyperscalers and neocloud operators will want the flexibility to mix GPUs and specialized AI accelerators without creating a separate networking architecture for every platform. Its use of Nvidia Spectrum-X also gives Nvidia a role in that more open infrastructure model.
The key test now is commercial execution. Token Fabric is scheduled to roll out in stages through 2027, and Upscale needs to prove that heterogeneous AI networking can deliver the performance, reliability and simplicity customers require.
For investors, the broader takeaway is that the next phase of the AI infrastructure race may not be determined by chips alone. The companies that connect, manage and optimize those chips could become equally important to the economics of AI data centers.
Follow the blog for more AI, semiconductor, Nvidia and technology-market updates.
This article is for informational and educational purposes only and should not be considered investment advice.