- Upscale AI's recent $190M Series A-1 extension lifts total funding to $500M and its valuation to $2B, with Nvidia joining as an investor
- The pure-play startup is building open-standard silicon, systems and software for AI scale-up and scale-out, but is not yet shipping
- It must unseat Cisco, Arista, Broadcom and Nvidia while leaning on standards like UALink and Ultra Ethernet
Upscale AI now has $500 million and a wildly ambitious mandate: build a new networking company from silicon up, aimed squarely at the AI data center, and do it against the largest names in the business.
The Santa Clara startup added $190 million in a Series A-1 extension late last month, led by Premji Invest, bringing total funding to $500 million at a $2 billion valuation, less than a year after leaving stealth. Nvidia is among the investors. Founded by Barun Kar and Rajiv Khemani and spun out of hardware firm Auradine, the company raised a $200 million Series A in January.
"Purpose-built for AI is a strong statement, and it's not just in one dimension, it's really creating that value through and through the stack, starting from the silicon to the system to the software," said Deepti Chandra, Upscale AI's VP of product management, strategy and marketing. Its scale-up silicon, SkyHammer, uses the open Switch Abstraction Interface and SONiC; on scale-out, Upscale is building systems on Nvidia's Spectrum-X silicon.
Chandra argued that Upscale's focus is its strength. Unlike bigger rivals such as Cisco — of which company Chandra is a veteran — Upscale is chasing only two problems: scale-up, which links accelerators, memory and storage inside a rack, and scale-out, which ties racks and data centers together. It is skipping the broader enterprise, campus and WAN markets that weigh down incumbent silicon.
Upscale is betting that the network, not the GPU, is becoming the limiting factor in large AI clusters — a view now widely shared, which is exactly why the space is crowded.
The open-standards pitch
Upscale is betting on open standards — SONiC, Ultra Ethernet, UALink and ESUN — to win buyers wary of lock-in. "The best technology is always driven by sourcing. There will always be a multi-vendor ecosystem," Chandra said, adding that open standards let vendors "normalize on an open knowledge base."
The catch: several of those standards are still in specification or sampling stages, with volume production widely expected in 2027 and 2028.
Crowded field and no product yet
The competition is formidable and already shipping. Cisco is pushing its Silicon One G300 switching chip and projects more than $3 billion in AI revenue this fiscal year. Arista, whose CEO Jayshree Ullal calls Ethernet the "eventual winner" for AI networking, lifted its 2026 AI target to $3.25 billion. Broadcom's merchant Tomahawk silicon and Scale-Up Ethernet already sit in hyperscaler racks, and HPE's newly acquired Juniper and Dell round out a field of far larger, entrenched vendors.
Upscale's biggest weakness is simple: it isn't selling anything yet. Asked directly whether it is shipping, Chandra deferred: "Please wait for the product announcements that will be following through shortly." The company says evaluations are underway with hyperscalers and neoclouds, with products due late this year or early next. Until then, its $2 billion valuation rests on pedigree, a war chest and a bet that focus beats scale.