Write an article about
NVIDIA GPUs dominate today’s AI data centers, but they’re no longer the only processors companies want to use. Cloud providers are developing their own AI chips, while specialized accelerators are entering the market. Getting these different processors to work together efficiently is creating a new challenge for data center operators.
That’s the problem Upscale hopes to solve with its new networking platform, Token Fabric. The NVIDIA-backed startup says the technology will make it easier to connect AI chips from different suppliers, so data centers can use different types of hardware without having to manage separate networks.
This could let data centers use AI chips from rival suppliers. And this is becoming more important now as companies are moving beyond the usual GPU clusters. We’ve seen big tech companies developing their own AI chips, while startups are also coming out with different types of processors. Cloud providers also want to give their customers more options when it comes to hardware.
But the problem is getting all these different chips to work together. It’s not as simple as putting them all in one data center. There are a lot of networking issues that come with it.
Modern AI workloads require processors to exchange enormous amounts of data. For example, thousands of accelerators may need to coordinate calculations across a large model during training. These workloads may also demand low latency and congestion management to ensure there are no unnecessary network delays that end up slowing the entire cluster.
Even with all that expensive hardware, a slow network can hold things back. Upscale is trying to fix this by combining two types of networking, scale-up and scale-out.
(Shutterstock/MAUZART)
Scale-up is basically how the chips within the same system or rack communicate with each other. They need fast connections with very little delay. Scale-out is different. It connects the racks and systems across the data center.
Both are essential to large AI installations. However, they have traditionally involved distinct networking environments.
Token Fabric combines Upscale’s SkyFabriX silicon for scale-up connections with scale-out systems powered by Spectrum-X Ethernet from NVIDIA. To help manage both environments together, the company is also introducing software tools.
Based on Upscale’s earlier SkyHammer architecture, SkyFabriX offers a stated switching capacity of 115.2 terabits per second. It supports open networking standards, such as the OCP ESUN and UALoE. The company’s longer-term roadmap includes multi-petabit switching capacity.
On the scale-out side, Token Fabric supports Ethernet connectivity ranging from 400G and 800G to 1.6T.
The software foundation is provided by Upscale’s SkyOS network operating system. SkyCMD handles orchestration, monitoring and management. The software is designed to identify network bottlenecks and help operators determine where equipment upgrades may be needed.
Customers have the flexibility to purchase the complete platform or individual components, including silicon, systems and software.
There is an interesting competitive dimension to the announcement. NVIDIA has invested in Upscale, and its Spectrum-X networking technology is part of Token Fabric. But Upscale also wants its platform to work with AI chips from other companies, including NVIDIA’s competitors.
Upscale has raised around $500 million so far. That includes another $190 million raised in June as an extension of its Series A round. Reuters reported that the company was valued at $2 billion at the time. CEO Barun Kar is expecting Token Fabric to bring in tens of millions of dollars in revenue next year, and possibly reach the low hundreds of millions later on.
(Shutterstock/PHOTOCREO-Michal-Bednarek)
“The industry has spent decades connecting computers. AI demands something different: one computer, built at data-center scale,” said Barun Kar, CEO, Upscale.
“Upscale Token Fabric brings heterogeneous accelerators, scale-up, scale-out and software together as one coherent AI system, built on open standards, so the network never gets in the way. Operators focus on delivering the best AI compute for their workloads; the integrated stack delivers the performance and time to market.”
For now, the platform isn’t fully available. Upscale has started early-access and testing programs, with the first components expected toward the end of 2026. General availability is planned for early 2027.
Of course, there’s still the question of how well all this will work in actual data centers. Just because a network supports different standards doesn’t mean chips from different companies will work together without any problems. There are still differences in software, memory and how the processors communicate.
What this announcement shows is that as more companies start using different types of AI chips, the networking side of things is going to matter a lot more too, and companies who are able to get ahead early may hold a significant competitive advantage.
.Organize the content with appropriate headings and subheadings ( h2, h3, h4, h5, h6). Include conclusion section and FAQs section with Proper questions and answers at the end. do not include the title. it must return only article i dont want any extra information or introductory text with article e.g: ” Here is rewritten article:” or “Here is the rewritten content:”
