Upscale AI, a start-up based in Santa Clara, California, introduced SkyHammer, a ground-up interconnect architecture designed specifically for large-scale AI systems. Unlike conventional front-end switch or PCIe-based fabrics, SkyHammer aims to eliminate performance bottlenecks by treating the entire AI cluster as a single coherent system, fusing CPUs, GPUs, and memory through nanosecond-latency load/store access. The company positions SkyHammer as a response to the growing limits of traditional data center networking under trillion-parameter AI workloads.
The architecture is built from the ASIC level up, with deterministic flow control, real-time telemetry, adaptive load handling, and built-in resiliency integrated into the core design. Upscale AI says every design element was intentional—removing unnecessary features that introduce latency and complexity. SkyHammer supports multiple open standards and emerging protocols such as ESUN, Ultra Ethernet Consortium (UEC), and UALink, allowing future interoperability with evolving GPU, XPU, and accelerator ecosystems. Products based on the architecture are slated for release in 2026.
SkyHammer’s open, modular design enables predictable performance from rack to supercluster scale. It emphasizes deterministic behavior, operational resilience, and adaptability across heterogeneous environments, aligning with industry efforts to make AI infrastructure more efficient, open, and sustainable.
• AI-native architecture purpose-built for collective AI workloads, not legacy server or storage systems
• Deterministic latency and guaranteed bandwidth for synchronized large-scale clusters
• Deep telemetry and observability built in from inception, not added post-deployment
• Compatibility with open interconnect standards including ESUN, UEC, and UALink
• Products based on SkyHammer planned for 2026 availability
🌐 Analysis: Upscale AI’s SkyHammer emerges amid a wave of innovation focused on purpose-built AI fabrics. Its emphasis on open standards positions it alongside recent ecosystem developments such as UALink and Ultra Ethernet, both seeking to define interoperable interconnects for AI workloads. With backing from industry analysts and participation by hyperscalers and GPU vendors, SkyHammer could play a role in shaping the next generation of scale-up AI infrastructure where deterministic performance and openness are paramount.
Upscale AI was co-founded by Barun Kar (CEO) and Rajiv Khemani (Executive Chairman), both of whom are serial entrepreneurs in networking, security, and silicon businesses. Their leadership team draws heavily from experience at firms such as Marvell, Broadcom, Intel, Cisco, AWS, Microsoft, Palo Alto Networks, and Juniper Networks.
In September 2025, Upscale emerged from stealth with a headline $100 million seed round, co-led by Mayfield and Maverick Silicon, with participation from Qualcomm Ventures, StepStone Group, Celesta Capital, Stanford University, and others. The startup had been incubated by Auradine, a blockchain and AI compute infrastructure company.
🌐 We’re tracking the latest developments in AI infrastructure and networking architecture. Follow our ongoing coverage at: https://convergedigest.com/category/ai-infrastructure/
