Upscale AI has raised $500 million in total funding and doubled its valuation to $2 billion since announcing a $200 million Series A in January, and this week the company disclosed specifications for its "SkyHammer" scale-up switch ASIC — while naming Nvidia a supplier for its scale-out Ethernet fabrics.

The startup, now employing more than 300 people, is partnering with the GPU giant even as it builds technology that competes with Nvidia's NVSwitch. It is a positioning that tells you a great deal about where the AI networking market is heading.

What does SkyHammer deliver?

According to a webinar slide presented by Arvind Srikumar, senior vice president of product and marketing, SkyHammer will deliver:

• 115.2 Tb/sec of aggregate bandwidth, which the company says stands alongside the best scale-out ASICs from Cisco, Broadcom, and Nvidia.
• 576 accelerators in a single networking tier, with multi-tier scaling to "thousands of accelerators."
• Eventual switch gear spanning multiple petabits per second of aggregate bandwidth, most likely by ganging multiple SkyHammer ASICs inside a single switch.

The 576-device figure works out to 200 Gb/sec of bandwidth per device. With 72 XPUs, each port gets 1.6 Tb/sec, or 200 GB/sec of memory bandwidth across the scale-up network — one-sixth of the 1.2 TB/sec-class NVSwitch 5/NVLink 5 stack in Grace-Blackwell NVL72 systems. Vera-Rubin NVL72 systems will double that again to 3.6 TB/sec with NVSwitch 6/NVLink 6.

Notably, SkyHammer runs UALink over Ethernet (UALoE) and ESUN, not UALink native — which reveals, as the company's materials indicate, that under the covers it is an Ethernet switch ASIC.

Does bandwidth matter more than latency?

Upscale AI's bet is that predictable latency, not raw bandwidth, wins for AI inference workloads.

"One of the most important metrics when it comes to tokens is latency," Srikumar said. "With AI, the latency has to be super predictable, and we have built into the ASIC predictable jitter. And at the same time, transmission reliability within the domain is also important. We have some protocols which are standards-based that provide link level reliability, but within the fabric we have a mechanism to ensure the fabric is 100 percent lossless."

He added: "We believe SkyFabriX is going to be uniquely positioned in the market because we started with a clean sheet of paper rather than taking an Ethernet fabric and retrofitting the protocols that are needed for scale up."

The company did not release full speeds and feeds, nor a roadmap for future SkyHammer ASICs. Co-founder and executive chairman Rajiv Khemani and chief executive officer Barun Kar both bring long experience in high-performance systems and networking, and further design details are expected.

How does the Nvidia partnership work?

Upscale AI will acquire Nvidia's Spectrum-X Ethernet ASICs and build its own scale-out switches from them. It is also porting its SkyOS network operating system — an AI-optimized version of SONiC and its Switch Abstraction Interface, originally created by Microsoft in 2016 — to run on Spectrum-X silicon.

The Upscale AI versions of Spectrum-X switches will support 400 Gb/sec, 800 Gb/sec, and 1.6 Tb/sec ports.

The choice of Spectrum-X, rather than Nvidia's Quantum-X InfiniBand switches, reflects the founders' conviction that Ethernet will eventually displace InfiniBand. First-generation switches compatible with the Ultra Ethernet Consortium's specifications may accomplish that goal after nearly three decades of InfiniBand setting the high-performance standard. InfiniBand will still beat Ethernet on port-to-port hop latency, but Ethernet can scale out further for a given network topology.

SkyOS will span both scale-up and scale-out networks, and the SkyCMD telemetry and orchestration tool will treat both networks as a single whole. Customers will be able to use any CPU, GPU, or XPU front-ended by a DPU or NIC and mix and match devices to build AI clusters.

Can it challenge NVSwitch?

There is a catch. Nvidia does not support SkyHammer and SkyOS as an alternative to NVSwitch for scale-up, so there is currently no way to do scale-up networking with Nvidia GPUs on the new platform.

AMD, by contrast, is counting on companies like Upscale AI to provide scale-up options supporting UALink, UALoE, or ESUN — and, as industry observers note, could lose patience and simply acquire the startup to own one of the options outright.

Nvidia retains alternatives of its own. It could reduce per-port bandwidth on a future NVSwitch ASIC and increase the radix to flatten its NVSwitch network. Or it could port the NVLink protocol to SkyHammer and become the dominant supplier of UALoE or ESUN switch silicon itself. The UALink consortium promises 1,024 XPUs in a single tier, and Nvidia plans to push NVSwitch beyond its current 576-GPU multi-tier limit to 1,152 GPUs with the Rubin Ultra generation.

The market is about to find out what actually works, as UALoE and ESUN memory protocols land on fast Ethernet switches tuned for scale-up workloads.