Upscale AI Challenges Nvidia’s Scale-Up Dominance with SkyHammer Fabric and $2 Billion Valuation

Achieving these petabit scales will likely involve ganging multiple SkyHammer ASICs together within a single switch. This is a proven strategy used by companies like Broadcom and Meta, where a "baby two-tier" network can be constructed from six independent ASICs to double the bandwidth of ports. While this approach can increase latency due to additional port hops, it offers a more cost-effective path to massive bandwidth than developing a single, monolithic 2X-bandwidth device. Upscale AI appears to be betting on the economic efficiency of this architecture to provide a competitive alternative to proprietary stacks.

The scalability of the SkyHammer memory domain is equally ambitious. The architecture reportedly supports up to 576 accelerators in a single networking tier. By utilizing multiple tiers, the company claims it can scale to "thousands of accelerators." This is a direct shot at the UALink consortium, which has promised support for 1,024 XPUs in a single tier, and Nvidia, which is aiming to scale its NVSwitch fabric to 1,152 GPUs in the upcoming "Rubin Ultra" generation.

However, a raw comparison of bandwidth reveals a significant gap. At 576 devices, the SkyHammer chip provides approximately 200 Gb/sec of bandwidth per device. This is a fraction of the 1.8 TB/sec offered by the NVLink 5 generation used in Nvidia’s Grace-Blackwell systems, or the 3.6 TB/sec expected in the Vera-Rubin generation. Even at a smaller scale of 72 XPUs, SkyHammer delivers roughly 1.6 Tb/sec per port, which is still about one-sixth the memory bandwidth of the NVSwitch 5 stack.

Upscale AI’s defense of this disparity rests on the belief that bandwidth is not the only metric that matters, especially for AI inference. Arvind Srikumar, Senior Vice President of Product and Marketing at Upscale AI, argues that predictable latency and jitter are the true bottlenecks for token generation in AI models. Srikumar notes that while traditional Ethernet fabrics often rely on a fixed, less-parallel pipeline, SkyHammer was designed from a "clean sheet of paper" to minimize latency and ensure 100 percent losslessness within the fabric.

"One of the most important metrics when it comes to tokens is latency," Srikumar explained. "We have to produce, but this is where it’s a little bit different. With AI, the latency has to be super predictable, and we have built into the ASIC predictable jitter. 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."

This focus on "reliable and predictable latency" suggests that Upscale AI is targeting a specific segment of the market—those running massive inference workloads where the overhead of proprietary, high-bandwidth training fabrics may be overkill. The company is leaning into standards like UALink over Ethernet (UALoE) and the ESUN memory protocols. The fact that SkyHammer utilizes UALoE indicates that, despite its optimizations, it remains an Ethernet-based ASIC at its core.