Token Fabric, Explained: The Open Network That Wants to Connect Rival AI Chips

The universal remote for your accelerator shelf

Imagine buying a new television and discovering it needs its own brand of HDMI cable, its own brand of remote control, and a different wall outlet than every other TV you own. Absurd — but that is roughly what buying AI accelerators looks like today. An NVIDIA GPU speaks NVLink inside the rack and InfiniBand across the cluster; rival chips speak their own dialects; every chip family drags its own networking stack into your data center. Mix suppliers and you end up building, powering, and babysitting parallel networks.

On October 8, 2026, a startup called Upscale AI said it wants to sell the industry a universal remote. Its new platform, Token Fabric, is an open networking stack — one scale-up network, one scale-out network, and one software layer — designed to connect GPUs and other AI chips from multiple suppliers without a separate network per chip family. Reuters covered the launch the same day: Upscale, valued at $2 billion in June and backed by NVIDIA, Salesforce, and Singapore's Temasek, is betting that neoclouds and large cloud providers will pay for the freedom to mix accelerators.

What Upscale actually launched

Token Fabric is three things in one box, according to Upscale's launch announcement:

• Scale-up: SkyFabriX silicon and switch trays. Upscale's own purpose-built scale-up switch chip (SkyFabriX), based on its SkyHammer architecture, plus tray systems (custom, ORV3 — OCP's Open Rack v3 — standard rack) that pool GPUs and other XPUs into large compute domains. It supports open standards — ESUN, UALoE, SUE-T, Ethernet/IP — with room for UALink's evolution: deterministic, ultra-low-latency, memory-semantic communication (chips reading each other's memory directly, like NVLink's shared memory pool), with a multi-petabit-per-second roadmap and copper plus optical variants (pluggable, near-packaged, and co-packaged optics).
• Scale-out: NVIDIA Spectrum-X Ethernet. Upscale-engineered switch systems on NVIDIA's Spectrum-X Ethernet silicon — 400G and 800G today, a path to 1.6T and future generations.
• Operate as one: SkyOS and SkyCMD. An AI-native network operating system (SkyOS) plus an operations layer (SkyCMD) — one place to configure, monitor, and fix both fabrics, with software that spots bottlenecks and flags gear that may need replacing.

The schedule is ambitious: the first piece — connecting groups of AI chips across a data center — lands in Q4 2026, with the full platform staging through 2027. CEO Barun Kar told Reuters he expects revenue next year in the tens of millions of dollars, potentially the low hundreds of millions; a $190 million June extension, led by Premji Invest, brought total funding to $500 million.

Why AI clusters need two networks

Token Fabric is two networks under one roof because AI networking has two different jobs. Scale-up links accelerators inside a tightly coupled domain: a server, a rack, or a pod, over distances of centimeters to a few meters, with latency under a microsecond and several terabits per second of bandwidth per chip. This is NVLink and NVSwitch territory (NVLink dates to 2014), and increasingly the open standard UALink, whose first specification landed in the first half of 2025 and supports up to 1,024 accelerators connected through UALink switches. NVIDIA's NVL72 rack shows the shape: 72 Blackwell GPUs in one rack, 576 across eight racks.

Scale-out connects those domains across the data center: hundreds of meters to kilometers, latency in the single-digit milliseconds, 400G or 800G per chip today. This is InfiniBand, Ethernet with RoCE (RDMA over Converged Ethernet — chips reading each other's memory, bypassing the CPU), the Ultra Ethernet Consortium's work, and NVIDIA's Spectrum-X.

One network cannot do both jobs. Scale-up's direct-link, memory-semantic designs cannot stretch far beyond their domain — they are capped by power, cooling, and cable reach. Scale-out's switched fabrics can reach the whole floor but cannot deliver memory-level latency and determinism at a sane cost. Every large AI cluster is therefore two layers: a tightly coupled scale-up domain feeding into a scale-out fabric. That boundary is exactly where Token Fabric is trying to plant its flag — one vendor, both layers, one operations story. (Our two-networks guide walks the split in more detail.)

Why "open" is the whole bet

The real story is not the hardware — it is the word "open." NVIDIA's NVLink is proprietary: buy NVIDIA GPUs, get NVIDIA's scale-up network. The industry's answer is UALink, an open standard backed by a consortium of chip and system companies, with its first specification out in the first half of 2025. Upscale is planting SkyFabriX squarely in that camp: open protocols, any accelerator, Ethernet where possible.

The tailwinds are enormous: Dell'Oro forecasts nearly $1 trillion in AI back-end switch spending through 2030, with Ethernet becoming the leading scale-up technology, and the 650 Group projects AI networking past $200 billion by 2030. That is what the $2 billion valuation is pricing in.

There is a power story too. A conventional 1.6-terabit pluggable module burns around 25 watts and roughly 15–20 picojoules per transmitted bit — the signal travels a long way across the circuit board and needs power-hungry processing to survive. Co-packaged optics (CPO) puts the optical engine millimeters from the switch chip, cutting energy to around 3–5 picojoules per bit and module power to about 5 watts. Token Fabric's scale-up roadmap explicitly includes CPO variants — the power-per-bit answer to ever-denser domains.

What it means for network engineers

Keep two things in your head at once. First, the direction is real: open scale-up standards, Ethernet everywhere, co-packaged optics for power, and one software layer instead of three vendor consoles are where AI networking is heading — and hyperscalers and neoclouds genuinely want to mix accelerators without rebuilding networks. UALink, the Ultra Ethernet Consortium, and CPO belong on your reading list now.

Second, stay skeptical in the healthy way. Token Fabric is an announced architecture with a staged rollout, not a deployed product — first piece in Q4 2026, full platform staging through 2027. The open-standards claims live or die on real interoperability testing, and the operations story (SkyOS/SkyCMD unifying both fabrics) is likely the hardest part; public detail there is still thin. Watch for evidence at the OCP Global Summit, October 12–15 in San Jose, where open AI infrastructure is the headline act.

Three questions for any open-fabric pitch: which accelerators have actually been tested together on it (not just listed as partners), who owns the software when something breaks at 3 a.m., and how the scale-up domain grows when you outgrow one rack.

Go deeper hands-on: our two-networks guide covers the scale-up/scale-out split, the SONiC guide covers what an open network OS really has to do, the ECMP hash visualizer shows why load-balancing math matters on AI scale-out fabrics, and the optic power-budget tool lets you price the CPO trade-off yourself.