en.Wedoany.com Reported - U.S. startup Upscale AI has completed a $190 million Series A-1 extension round, bringing total funding to $500 million at a $2 billion valuation. The company plans to build a new network company for chips starting from AI data centers.
Founded by Barun Kar and Rajiv Khemani, this Santa Clara-based enterprise was spun off from hardware company Auradine and completed a $200 million Series A round in January this year. The current round is led by Premji Invest, with Nvidia also among the investors. Deepti Chandra, Vice President of Product Management, Strategy, and Marketing, stated that being "built for AI" is not a single-dimensional optimization but a full-stack value creation from silicon to systems to software. Its expansion chip, SkyHammer, adopts an open Switch Abstraction Interface and SONiC; in scale-out scenarios, Upscale is building systems on Nvidia's Spectrum-X chips.
Chandra believes focus is Upscale's competitive advantage. Compared to rivals like Cisco, the company tracks only two issues: scale-up (connecting accelerators, memory, and storage within racks) and scale-out (connecting racks and data centers), skipping broader markets such as enterprise, campus, and wide area networks. Upscale's core philosophy is that networks, rather than GPUs, are becoming the limiting factor in large AI clusters—a view widely recognized in the industry.
On open standards, Upscale is betting on SONiC, Ultra Ethernet, UALink, and ESUN to attract buyers wary of vendor lock-in. Chandra said, "The best technology is always driven by procurement, and there will always be a multi-vendor ecosystem," adding that open standards allow vendors to "standardize on an open knowledge base." However, several of these standards are still in the specification or sample stage, with mass production expected in 2027 and 2028.
The current market is highly competitive with products already shipping. Cisco is promoting its Silicon One G300 switching chip, with projected AI revenue exceeding $3 billion this fiscal year. Arista CEO Jayshree Ullal called Ethernet the "ultimate winner" for AI networks and raised Arista's 2026 AI target to $3.25 billion. Broadcom's commercial Tomahawk chips and Scale-Up Ethernet are deployed in hyperscale data center racks. Hewlett Packard Enterprise's newly acquired Juniper Networks, along with Dell, also form a deeply entrenched vendor base.
Upscale's biggest weakness is that it currently has no products for sale. When asked if it has shipped, Chandra responded, "Please wait for upcoming product announcements." The company stated it is conducting evaluations with hyperscale cloud and new cloud service providers, with products expected to launch by the end of this year or early next year. Until then, its $2 billion valuation relies on its background, cash reserves, and the strategy that focus outweighs scale.
