Upscale AI
by Upscale AI
Pure-play AI networking: scale-up switch silicon and open Ethernet systems for GPU clusters
Upscale AI builds networking silicon, systems and software designed only for AI clusters — a scale-up interconnect called SkyHammer aimed squarely at NVIDIA's NVSwitch, plus open Ethernet scale-out switches running an AI-tuned SONiC. It sells to hyperscalers, neoclouds and large enterprises building GPU and XPU fleets.
Upscale AI is a networking company founded in 2025 that builds silicon, systems and software exclusively for AI infrastructure, on the argument that general-purpose networking carries legacy baggage AI clusters cannot afford. Its scale-up product, SkyHammer, is a rack-scale ASIC interconnect using a memory-semantic load-store architecture with in-network acceleration for collective communication, competing directly with NVIDIA's NVSwitch inside NVL72-class racks and supporting the UALink and ESUN open standards rather than a proprietary fabric. It will be sold both as a standalone ASIC and as integrated switch blades and racks. The scale-out line is open Ethernet systems built on NVIDIA Spectrum-X switch silicon running an AI-optimised build of SONiC, the open network operating system originally developed by Microsoft and now used across hyperscale environments; Upscale AI is a Premier Member of the SONiC Foundation and participates in UltraEthernet and OCP's SAI and ESUN work. The company was co-founded by CEO Barun Kar and Executive Chairman Rajiv Khemani — Kar a Juniper Networks and Motorola veteran, Khemani formerly in C-suite roles at Innovium, which Marvell acquired, and at Cavium, Intel, NetApp and Sun — and spun out of hardware firm Auradine. It reached roughly 200 staff by March 2026. Funding has been unusually fast: a $100M seed, a $200M Series A in January 2026 led by Tiger Global, Premji Invest and Xora Innovation, and a $190M extension in June 2026 led by Premji Invest with NVIDIA, Salesforce Ventures, Temasek and Seligman Ventures joining, taking the total to roughly $500M at a $2B valuation.
Infrastructure and network architects at hyperscalers, neocloud GPU providers and very large enterprises building multi-rack training or inference clusters, who want a scale-up fabric on open UALink/ESUN standards instead of committing the interconnect layer to a single accelerator vendor.
A second source for the rack-scale interconnect — the layer where a GPU fleet is otherwise locked to its accelerator vendor's proprietary fabric and pricing.
At a Glance
- Category
- Infrastructure & Cloud
- Pricing
- Contact for pricing
- Target Market
- CTOs, VP Infrastructure, Network Architects, Data Center Engineering Leads, Cloud Providers
- Deployment
- Self-hosted
- Founded
- 2025
- Team Size
- 201-500
- Customers
- None disclosed. Hardware reported as under evaluation by multiple hyperscalers and neocloud operators as of June 2026
Key Features
- ✓SkyHammer scale-up interconnect
A rack-scale ASIC using a memory-semantic load-store architecture to synchronise accelerators, memory and storage inside a rack with deterministic GPU-to-GPU latency.
- ✓In-network collective acceleration
Offloads collective communication operations into the switch, comparable in intent to NVIDIA's SHARP, which is where large training runs lose time.
- ✓UALink and ESUN standards support
Builds to open scale-up standards rather than a proprietary fabric, so the interconnect layer is not tied to one accelerator vendor's roadmap.
- ✓Open Ethernet scale-out systems
Switches built on NVIDIA Spectrum-X silicon for distributed training and large-scale inference across heterogeneous AI infrastructure between racks.
- ✓AI-optimised SONiC network OS
Ships a modified build of the open-source SONiC operating system used at hyperscale, avoiding a proprietary NOS and the operational retraining it forces.
- ✓Silicon, systems and software in one stack
SkyHammer is offered both as a standalone ASIC and as integrated switch blades and racks, letting buyers choose their integration depth.
Use Cases
- •Rack-scale GPU cluster interconnect
Replace the accelerator vendor's proprietary scale-up fabric inside a rack with a standards-based alternative supporting UALink and ESUN.
- •Multi-rack distributed training fabric
Connect racks with open Ethernet scale-out switching so large training jobs span the cluster without a proprietary inter-rack fabric.
- •Neocloud GPU-as-a-service buildouts
Give GPU cloud providers a second source for networking, which is the component with the least competitive supply in an AI rack.
- •Heterogeneous accelerator estates
Run one scale-out Ethernet fabric across mixed GPU and XPU hardware instead of maintaining a separate fabric per accelerator vendor.
- •SONiC-standardised network operations
Extend an existing SONiC-based operational model into AI clusters rather than adopting a separate proprietary network operating system.
Ideal For
Best For
- ✓Hyperscalers and neocloud operators standardising rack-scale AI clusters who want an alternative to NVIDIA's proprietary NVSwitch fabric
- ✓Organisations betting on the open UALink and ESUN scale-up standards rather than a single-vendor interconnect
- ✓Teams already running SONiC who want an AI-tuned distribution and a vendor active in the SONiC Foundation and OCP
- ✓Large training and distributed inference deployments where collective-communication latency, not raw GPU count, is the bottleneck
- ✓Buyers assembling heterogeneous accelerator fleets that need one scale-out Ethernet fabric across them
Not Ideal For
- ✗Anyone needing to deploy this year: product launch is targeted around the second half of 2026 and the hardware is described as under evaluation by hyperscalers and neoclouds, with no general availability date and no named customers
- ✗Enterprises without a hardware evaluation function — this is switch silicon and systems requiring lab qualification, not software you trial
- ✗Teams that want published specifications or third-party benchmarks before committing, since detailed SkyHammer specs have not been released
- ✗Smaller GPU estates of a few racks, where the economics of a specialised scale-up fabric do not justify displacing the accelerator vendor's bundled interconnect
- ✗Risk-averse buyers who need a multi-year support track record from an infrastructure vendor founded in 2025
Deployment
Market & Ratings
None disclosed. Hardware reported as under evaluation by multiple hyperscalers and neocloud operators as of June 2026
Market Analysis
Pros
- ✓Attacks the least-contested layer of the AI rack — the scale-up interconnect — where buyers currently have essentially one option
- ✓Standards-first strategy across UALink, ESUN, SONiC and OCP, which matters to buyers whose main objection to the incumbent is lock-in
- ✓Founding team has shipped networking silicon before: Khemani led Innovium through its acquisition by Marvell, and Kar came from Juniper and Motorola
- ✓Roughly $500M raised and a $2B valuation inside a year gives it the capital to complete a silicon programme, which kills most networking startups
- ✓Backing from NVIDIA, Intel Capital, Qualcomm Ventures and Salesforce Ventures signals supply-chain and customer-side interest, not just financial investors
Cons
- ✗Nothing is generally available yet — launch is targeted for the second half of 2026, there are no named customers, and hardware is only described as under evaluation
- ✗Detailed SkyHammer specifications have not been published, so there is no way to compare it against NVSwitch or Broadcom's Tomahawk 6 on radix, bandwidth or power
- ✗The UALink ecosystem it depends on is itself immature, and competing implementations currently tunnel the protocol over Ethernet rather than using purpose-built switches
- ✗No practitioner feedback exists anywhere — Hacker News returns zero stories, and there are no G2, TrustRadius or Reddit deployment reports, which is expected pre-GA but leaves claims unverified
- ✗Displacing an incumbent interconnect means displacing its software ecosystem and operational tooling too, which is where challenger networking silicon historically stalls
- ✗Building scale-out systems on NVIDIA's Spectrum-X silicon while competing with NVIDIA's NVSwitch creates a dependency on a direct competitor
Pricing
Silicon, systems and OEM engagements
Contact for pricing
- ✓SkyHammer available as a standalone ASIC or as integrated switch blades and racks
- ✓Scale-out Ethernet systems on NVIDIA Spectrum-X silicon
- ✓AI-optimised SONiC network operating system
- ✓Direct engagement with the vendor for evaluation
No list pricing exists and none should be expected — this is switch silicon and rack-scale systems sold through direct enterprise and OEM engagement, priced per deployment against volume, integration depth and whether the buyer takes bare ASICs or finished blades and racks. The product had not reached general availability at the time of research, with launch targeted around the second half of 2026 and hardware described as under evaluation by multiple hyperscalers and neocloud operators, so any quote today is an evaluation engagement rather than a purchasable SKU. The commercial argument is second-sourcing the interconnect layer and the pricing leverage that creates, not a lower sticker price.
Security & Compliance
Sources
This page was written from 5 sources, 3 on domains other than upscale.com.
- 1.upscale.com — upscale.comvendor
- 2.upscale.com — from 100m seed to unicorn in months upscale ai closes oversuvendor
- 3.theregister.com — upscale skyhammer nvidia
- 4.siliconangle.com — ai networking provider upscale ai raises 190m 2b valuation
- 5.sdxcentral.com — nvidia backs ai networking switch silicon startup upscale in
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Gimlet Cloud
Multi-silicon inference cloud that splits AI agent workloads across GPUs, CPUs and accelerators
DigitalOcean Managed Agents
Managed agent runtime with microVM sandboxes, 16,000+ governed tools and serverless inference, billed on active CPU
Modular
MAX inference framework and Mojo language for serving AI models on NVIDIA, AMD and other chips
ZML/LLMD
Free, Python-free LLM inference server that runs open models on NVIDIA, AMD, Google TPU, Intel and Apple chips from one binary