Nvidia-Backed Upscale AI Launches Token Fabric, the Networking Platform That Lets Data Centres Mix AI Chips From Rival Suppliers

Abdul Mannan 1 hour ago • 7 min read

Rows of AI server racks in a darkened data centre hall linked by glowing fibre-optic networking cables, blue and green indicator lights, cinematic editorial photograph

Upscale AI, a startup valued at $2 billion in June and backed by Nvidia, Salesforce and the Singapore state investor Temasek, on Thursday launched Token Fabric, a networking platform designed to let data centres connect artificial intelligence chips from competing suppliers inside a single system, according to Reuters.

Called Token Fabric, the platform combines hardware and software to connect AI processors from different vendors across a data centre, removing the need for operators to integrate separate networking systems for each supplier, Reuters reported. Upscale plans to release the first part of the platform — technology that connects groups of AI chips across a data centre — in the fourth quarter of this year, with the full platform rolling out in stages through 2027.

The pitch is aimed squarely at one of the fastest-growing buyers in the market: neoclouds and large cloud providers that are building out enormous data centres and want to use chips from multiple suppliers rather than being tied to a single vendor’s hardware. Chief executive Barun Kar told Reuters he expects Token Fabric to generate revenue in the tens of millions of dollars next year, potentially reaching the low hundreds of millions as adoption grows.

Under the hood, Token Fabric is built from three layers, according to the company’s announcement, reported by the industry publication HPCwire. The “scale-up” layer — connecting chips inside a tightly coupled pod of accelerators — uses Upscale’s own SkyFabriX silicon and switch trays. The “scale-out” layer, which links clusters of chips across the data centre, is delivered through Upscale-engineered systems powered by Nvidia’s Spectrum-X technology. Sitting above both is management software that identifies bandwidth bottlenecks and flags equipment that may need replacing before it causes downtime.

The declared purpose is economic rather than glamorous. As Reuters put it, the aim of the system is to keep expensive AI chips working rather than waiting for data, helping customers get more from their investment. A top-of-the-line AI accelerator can cost tens of thousands of dollars; every hour it sits idle waiting on a network transfer is money burned. Upscale’s argument is that a purpose-built, open networking fabric will push utilisation higher and, as a result, lower the effective cost of each computed token — “optimise tokens per dollar,” in the company’s own phrasing.

The financing behind the startup reflects how seriously investors are taking that argument. In June, Upscale raised a $190 million extension to its early-stage funding round, led by Premji Invest, the investment unit of Indian billionaire Azim Premji, taking its total funding to $500 million, Reuters reported. Nvidia, Salesforce and Temasek all joined the round — a backer list that gives Upscale both capital and strategic cover, since Nvidia’s own Spectrum-X networking technology is part of the product.

The company describes Token Fabric as an “industry-first open architecture” for so-called AI factories — the term, now widely used across the industry, for data centres in which compute, networking, storage and software must operate as one system. According to the announcement, the platform is designed for hyperscalers, neoclouds, enterprises and chip partners alike, and is built to connect any accelerator, scale accelerators inside a pod, and scale them across the data centre.

The timing is significant. AI networking has become one of the most contested layers of the AI infrastructure stack. Networking vendors today offer either discrete components that customers must integrate themselves or systems assembled from legacy networking gear not purpose-built for AI, Upscale says — a gap it intends to exploit. Market forecasts underline the scale of the prize: Dell’Oro Group projects nearly $1 trillion in AI back-end switch spending through 2030, with Ethernet emerging as the leading scale-up technology, while 650 Group projects AI networking will surpass $200 billion by 2030, according to the company’s announcement.

Nvidia’s networking chief weighed in on the same theme. “The network is the backbone of the AI factory, keeping accelerators connected and productive as systems scale,” the company’s senior vice-president of networking said, in remarks quoted by trade-press coverage of the launch. The comment, from the very company whose Spectrum-X technology Upscale is building on, suggests Token Fabric is arriving with at least the industry’s attention, if not its blessing.

Analysis: Why It Matters

The most important sentence in this launch is not about speeds or feeds — it is the quiet proposition that a data centre no longer has to belong to one chip company. Token Fabric’s central claim is interoperability: buy Nvidia chips for one cluster, AMD accelerators for another, and perhaps Google’s TPUs or Amazon’s Trainium for a third, and the fabric stitches them into a single managed system. That is a direct challenge to the oldest trick in enterprise technology — proprietary lock-in — applied to the hottest hardware market on earth.

There are three reasons this matters now. First, the economics. The AI buildout has made GPU idle time one of the most expensive problems in technology. Training a frontier model can involve tens of thousands of accelerators running for weeks; if the networking layer keeps even a few per cent of them waiting on data, the waste runs into millions of dollars per run. Upscale’s entire pitch — keep chips working, not waiting — is a bet that the next frontier of AI efficiency is not a better chip but a better network. It is an unfashionable bet, because networking does not inspire keynote speeches, but it is where the money leaks.

Second, supply risk. Cloud providers and neoclouds have learned, repeatedly, that dependence on a single supplier is a strategic vulnerability — price leverage, allocation priority in shortages, and roadmap risk all concentrate in one vendor’s hands. A credible multi-vendor fabric gives operators a genuine alternative: mix and match accelerators by workload, cost and availability, and let the network absorb the complexity. That hedging logic is exactly why Nvidia itself is an investor here — even as its Spectrum-X technology sits inside the product, Nvidia would rather own a slice of the interoperability layer than watch it emerge without one.

Third, standards. The industry is converging on Ethernet as the common language for AI scale-up networking — a shift codified by the Ultra Ethernet Consortium, the open-standards effort backed by AMD, Broadcom, Cisco, Intel, Meta and Microsoft. Token Fabric’s open-standards positioning places Upscale inside that current rather than against it. The risk, and it is a real one, is that openness becomes a marketing slogan rather than a technical fact: if the platform works best with Nvidia’s Spectrum-X and Upscale’s own silicon, “connect any accelerator” could prove more aspirational than real. Customers will test that claim ruthlessly.

There are reasons for scepticism. Kar’s revenue guidance — tens of millions of dollars next year, possibly reaching the low hundreds of millions — is modest against $500 million raised and a $2 billion valuation. The rollout stretches through 2027, which in AI-infrastructure years is an eternity; the competitive landscape could look very different by then. And Upscale is entering a market guarded by giants: Broadcom and Arista dominate data-centre networking, Cisco is regrouping with its Splunk-era software strategy, and Nvidia itself sells the networking gear Token Fabric is designed to work with or around. A startup selling openness to hyperscalers — customers who famously build their own networking — faces a long, expensive sales cycle.

But the launch also captures a genuine inflection point. The AI infrastructure market is shifting from a chip-supply scramble to a systems-optimisation game. The first phase of the boom was about getting any accelerators at all; the second phase, now under way, is about making the ones you have compute rather than wait. Networking companies that can prove they raise utilisation will command a premium, because their product effectively manufactures more chips out of the ones already bought. That is the most bullish reading of Token Fabric: not a box, but a way to print compute.

What to watch: first, whether the fourth-quarter release lands on schedule and with which launch customers — early wins with named neoclouds would convert a launch announcement into a business. Second, whether independent benchmarks confirm the utilisation claims; the industry has heard “better networking” pitches before. Third, how the incumbents respond — an Arista or Broadcom acquisition offer, or a competing open-fabric announcement, would be the market’s way of validating Upscale’s thesis. And fourth, whether Nvidia’s dual role as investor, technology licensor and dominant GPU vendor stays benign. A backer that is also the ecosystem’s gravitational centre is useful right up until it is not.