Upscale AI

Market position

Sources position Upscale AI as an early-stage challenger seeking to become the networking incumbent of the AI era, with Fortune drawing an analogy to Cisco and Juniper in the PC-internet era and Arista and Broadcom in the cloud era. As of April 2026 the company had not yet released a product, and both flagship products were still marked 'coming soon' on its site as of the sources fetched. The addressable market is framed by a Dell'Oro Group forecast that spending on AI data center switches will surpass $100 billion annually by 2030, with the five largest tech companies expected to spend roughly $660-690 billion on infrastructure in 2026.

Upscale AI differentiates on open standards and vendor neutrality, contrasting its fabric with proprietary, vendor-locked interconnects such as NVIDIA's NVLink and InfiniBand, and on delivering a vertically integrated stack from custom ASIC through systems and network OS rather than a single layer. It also emphasizes support for heterogeneous compute, allowing accelerators from multiple manufacturers in one cluster. An industry analyst quoted at launch (Alan Weckel, 650 Group) characterized the company as building a level playing field versus closed vendor-locked systems.

Technology

The stack spans custom switch ASICs, rack-scale and data-center systems, and network software. SkyHammer targets scale-up rack networking with deterministic, low-latency connectivity between accelerators, memory and storage, built from the ASIC up. The scale-out line uses NVIDIA Spectrum-X Ethernet switch silicon with an AI-optimized implementation of SONiC, the open-source network operating system used in hyperscale environments. The company's technology is grounded in open standards: SONiC, Ultra Accelerator Link (UALink), Ultra Ethernet, and the Switch Abstraction Interface (SAI), with announced capabilities including a hardened SAI/SONiC-based NOS with in-service upgrades, ultra-low-latency high-bandwidth AI network fabrics, and standards-based rack platforms supporting multivendor deployments.

Go-to-market

The company sells to large infrastructure operators directly, stating it is engaged with several hyperscalers and neocloud infrastructure providers with customer evaluations and deployments underway across both scale-out and scale-up networking environments. Its website offers a demo booking flow. It also builds market presence through participation in open networking standards bodies (Ultra Ethernet Consortium, UALink, Open Compute Project, SONiC Foundation) and through its partnership with NVIDIA on Spectrum-X-based systems.

Hyperscale cloud providers and neocloud AI infrastructure providers operating large-scale AI training and inference clusters, spanning rack-scale training deployments to cloud-scale distributed clusters.

Geography

Headquartered in Santa Clara, California, United States; press releases are datelined Santa Clara. Sources do not describe other locations.

History

Upscale AI was incubated at Auradine, a blockchain and AI chip startup where cofounders Barun Kar and Rajiv Khemani worked together. It launched publicly in September 2025 with more than $100 million in seed funding co-led by Mayfield and Maverick Silicon, with participation from StepStone Group, Celesta Capital, Xora, Qualcomm Ventures, Cota Capital, MVP Ventures and Stanford University, and with a team of 100+ technologists drawn from Marvell, Broadcom, Intel, Cisco, AWS, Microsoft, Google, Palo Alto Networks and Juniper Networks. A $200 million Series A followed in January 2026. In March 2026 the company announced its NVIDIA Spectrum-X-based scale-out fabric. In April 2026 it was reported to be in talks for a third round of $180-200 million at roughly a $2 billion valuation; the $190 million Series A-1 was confirmed on June 22, 2026, bringing total funding to $500 million and valuation to $2 billion in under 18 months.

Risks & controversies

Fortune notes several risks: custom chip design at leading manufacturing nodes costs hundreds of millions of dollars before a product ships, and suppliers increasingly require companies to commit to and pay for production slots up to two years in advance, meaning the $500 million raised will not last indefinitely. Competition is intense — Nexthop AI, a direct rival targeting the same hyperscaler and neocloud customers, raised a $500 million Series B in March 2026 at a $4.2 billion valuation, while Broadcom and NVIDIA have far larger R&D budgets. The open networking standard Upscale AI builds on, backed by AMD, Intel, Google, Meta, Microsoft and 80+ other companies, still needs to prove it can match NVIDIA's proprietary performance in real-world AI training. TechCrunch also observed that the company had raised three rounds in seven months without shipping a product, in a market where valuations can outpace fundamentals.