Upscale AI has raised $500 million in total funding, doubled its valuation to $2 billion, and grown past 300 employees — and this week it revealed first hard specs for its "SkyHammer" scale-up switch ASIC: 115.2 Tb/sec of aggregate bandwidth, enough to link up to 576 accelerators in a single networking tier.

The company is simultaneously partnering with Nvidia and competing against it. Upscale AI will buy Nvidia's Spectrum-X Ethernet ASICs and build its own scale-out switches around them, while its SkyHammer scale-up fabric positions itself as an alternative to Nvidia's NVSwitch — which Nvidia does not support as a substitute for its own rack-scale glue.

What did Upscale AI actually disclose?

Details came via a webinar presentation from Arvind Srikumar, senior vice president of product and marketing. The confirmed figures:

• SkyHammer delivers 115.2 Tb/sec aggregate bandwidth, which the company says stands toe-to-toe with the best scale-out switch ASICs from Cisco, Broadcom and Nvidia.
• The SkyFabriX architecture will eventually span switch gear across multiple petabits per second of aggregate bandwidth.
• A single SkyHammer tier supports up to 576 accelerators; multiple tiers can scale to "thousands of accelerators."
• Upscale AI's Spectrum-X-based scale-out switches will support 400 Gb/sec, 800 Gb/sec and 1.6 Tb/sec ports.

Upscale AI gave no roadmap for future SkyHammer ASICs and no latency numbers. Co-founder and executive chairman Rajiv Khemani and CEO Barun Kar bring long experience in high-performance systems and networking, which suggests further twists to come.

How does the math compare with NVSwitch?

The bandwidth gap is stark. At 576 devices per tier, each accelerator gets 200 Gb/sec over SkyHammer. With 72 XPUs, SkyHammer delivers 1.6 Tb/sec per port — 200 GB/sec of memory bandwidth, roughly one-sixth of what Nvidia's NVSwitch 5/NVLink 5 stack offers in the Grace-Blackwell NVL72 systems at 1.8 TB/sec. The Vera-Rubin NVL72 generation will push NVSwitch 6/NVLink 6 to 3.6 TB/sec.

Scale comparisons are also instructive. Nvidia supports 72 GPUs in a single domain today, with 576 GPUs possible across multiple tiers but not intended for production workloads. The Rubin Ultra generation is expected to take NVSwitch to 1,152 GPUs. The UALink consortium promises 1,024 XPUs in a single tier.

Upscale AI argues raw bandwidth is not the whole story, particularly for inference.

"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." He added that the fabric ensures "100 percent lossless" transmission and that SkyFabriX benefits from starting "with a clean sheet of paper rather than taking an Ethernet fabric and retrofitting the protocols that are needed for scale up."

Notably, SkyHammer runs UALink over Ethernet (UALoE) rather than native UALink, plus the ESUN memory protocol — a strong hint that under the covers it is an Ethernet switch ASIC.

Why partner with Nvidia on scale-out?

Upscale AI will acquire Spectrum-X ASICs — not Quantum-X InfiniBand — port its SkyOS network operating system onto them, and build its own switches. SkyOS is an AI-optimized version of SONiC and its Switch Abstraction Interface, which Microsoft created in 2016 and which has become a de facto standard across hyperscaler datacenters. SkyOS will span both scale-up and scale-out networks, with the SkyCMD telemetry and orchestration tool treating both as one fabric, and customers can mix and match any CPU, GPU or XPU front-ended by a DPU or NIC.

The catch: Nvidia does not support SkyHammer and SkyOS as an alternative to NVSwitch for scale-up, so there is no way — yet — to do scale-up networking with Nvidia GPUs on this fabric. AMD, by contrast, is counting on companies like Upscale AI to provide UALink, UALoE or ESUN scale-up options, and could conceivably acquire the startup to own one outright.

What happens next?

The market is about to find out what works as UALoE and ESUN get layered onto fast, scale-up-tuned Ethernet switches. Nvidia could respond by cutting per-port bandwidth to raise NVSwitch radix and flatten its network — or, faster still, port the NVLink protocol to SkyHammer and become the dominant supplier of UALoE or ESUN switchery itself.