Upscale AI
by Upscale AI
The pure-play AI networking company building the full-stack fabric for GPU and XPU clusters
Upscale AI is a full-stack AI networking company building silicon, systems, and software—including its SkyHammer scale-up switch ASIC and open Ethernet scale-out fabrics—to connect GPUs, XPUs, memory, and storage inside the largest AI training and inference clusters. It is aimed at hyperscalers and neocloud providers who need deterministic, low-latency interconnects for AI infrastructure.
At a Glance
- Category
- Infrastructure & Cloud
- Pricing
- Contact for pricing
- Target Market
- CTOs, CIOs, Infrastructure Architects, Cloud/Data Center Operators
- Founded
- 2025
- Headquarters
- Santa Clara, California, United States
Key Features
- ✓SkyHammer scale-up ASIC
A clean-slate switch silicon purpose-built to interconnect accelerators, memory, and storage within a rack at high bandwidth with deterministic, low latency.
- ✓Open Ethernet scale-out fabrics
Scale-out networking built on NVIDIA Spectrum-X switch silicon with a SONiC-based NOS to connect heterogeneous AI clusters.
- ✓Full-stack silicon, systems, and software
An end-to-end platform spanning custom silicon through systems and software optimized for AI workload economics and token efficiency.
- ✓Open-standards architecture
Supports heterogeneous compute on open standards including UltraEthernet, UALink, and OCP SAI/ESUN to avoid proprietary lock-in.
Use Cases
- •Frontier-model training clusters
Provide deterministic, high-bandwidth interconnect for synchronized multi-accelerator training at scale.
- •Large-scale inference serving
Reduce communication bottlenecks between accelerators to improve token efficiency and predictable performance for inference.
- •Neocloud AI data centers
Give emerging GPU-cloud providers an open, purpose-built AI fabric spanning single racks to large distributed clusters.
Ideal For
Best For
- ✓Building deterministic scale-up networks inside rack-scale AI clusters
- ✓Connecting GPUs, XPUs, memory, and storage for large-scale training and inference
- ✓Hyperscalers and neoclouds standardizing on open AI networking fabrics
Deployment
Market Analysis
Pros
- ✓Purpose-built for AI scale-up/scale-out interconnect bottlenecks
- ✓Strong strategic backing from NVIDIA and Salesforce Ventures
- ✓Founding team with prior networking-silicon exit (Innovium)
Cons
- ✗Scale-up SkyHammer products not yet generally available (2026 release)
- ✗Competing against entrenched networking incumbents
Pricing
Enterprise
Contact for pricing
- ✓SkyHammer scale-up switching (coming 2026)
- ✓Open Ethernet scale-out fabrics
- ✓Full-stack silicon, systems, and software
Pricing is not publicly listed; scale-up products based on the SkyHammer architecture are announced for release in 2026.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Prime Intellect
The open superintelligence stack — own your intelligence with custom compute, RL training, and inference in one platform.
Union.ai
The commercial platform for Flyte — build durable, dynamic AI workflows and agents in pure Python
TrueFoundry
Enterprise AI gateway and agentic deployment platform — secure, scalable, governed
Sail Research
Max-efficiency inference infrastructure for long-horizon AI agents