AI continues to fuel demand for new networking technology and the vendors that build it. As hyperscalers and neocloud providers race to deploy ever-larger clusters of accelerators, the backend networks that interconnect GPUs and XPUs have become a critical bottleneck. Upscale AI, a startup that emerged from stealth in September 2025 with a $100 million seed round, is positioning itself at the center of this trend. The company has now raised an additional $190 million in a Series A-1 extension, bringing its total funding to $500 million and its valuation to $2 billion. The round was led by Premji Invest and includes new investors Nvidia, Salesforce Ventures, Seligman Ventures, and Temasek. Existing backers—Maverick Silicon, Mayfield, Prosperity7 Ventures, StepStone Group, and Tiger Global—also participated.
The Santa Clara-based company describes itself as a pure-play AI networking company focused on the backend networks that connect GPUs and XPUs for hyperscalers and neoclouds. Upscale builds AI networking hardware across two tracks. Scale-up networking connects GPUs and XPUs within a cluster, enabling high-speed communication between accelerators in the same server rack or across adjacent racks. Scale-out networking extends connectivity across nodes, linking multiple clusters together to form larger training or inference systems. Both tracks are essential for modern AI workloads, which require massive parallel processing and low-latency data exchange.
Since the company emerged from stealth, it has expanded its efforts significantly. In a major milestone, Upscale inked a partnership with Nvidia on SpectrumX Ethernet, aligning its scale-out roadmap with Nvidia's own AI-optimized networking platform. Additionally, Upscale announced that it is developing its own custom scale-up switch ASIC, code-named Skyhammer. This silicon is designed from the ground up for AI scale-up use cases, setting it apart from commodity data center chips that are often repurposed for AI workloads. Rajiv Khemani, co-founder and executive chairman of Upscale AI, explained that the world is moving toward heterogeneous AI compute, and Upscale wants to be a player by providing a network that supports Nvidia-based infrastructure as well as other GPU and XPU vendors.
Skyhammer: Upscale’s purpose-built scale-up silicon
Skyhammer is Upscale AI’s flagship effort on the scale-up side of AI networking: a purpose-built switch silicon designed to connect GPUs and XPUs at very high speed inside AI systems. Khemani noted that unlike commodity data center chips that are often reused for AI, Skyhammer is being developed specifically for AI scale‑up use cases and is tightly coupled to Upscale’s broader full‑stack strategy, which spans silicon, systems, and software. The ASIC is engineered to handle the unique traffic patterns of distributed AI training, where collective communication operations like all-reduce and all-gather dominate. By optimizing the silicon for these patterns, Upscale claims it can reduce tail latency and improve overall job completion times.
Khemani declined to share detailed timelines, but he said Upscale expects to reveal product details on Skyhammer later this year, with actual deployment synced to when GPU and XPU vendors are ready. “The Skyhammer product doesn’t work by itself,” he explained. “It works in conjunction with XPUs and GPUs, and so for us to be deployed, the XPUs and GPUs need to incorporate scale‑up capabilities to interoperate with us.” This means that Upscale is not only developing its own silicon but also working closely with accelerator vendors to ensure compatibility. The company is likely targeting the upcoming generation of GPUs from Nvidia, AMD, and Intel, as well as custom XPUs from cloud providers like Google’s TPUs and Amazon’s Trainium.
Nvidia, Spectrum X, and strategic capital
Nvidia sits at the center of Upscale AI’s story, both as a technology partner and now as a strategic investor. On the product side, Upscale has aligned its scale-out roadmap with Nvidia’s Spectrum X Ethernet platform to build AI‑optimized network fabrics that connect GPU and XPU nodes across the data center. The Spectrum X–based systems are expected to reach the market later this year, with Khemani noting that products are already in customer labs and are booked for early deployments. This collaboration gives Upscale credibility in the crowded AI networking space, where Nvidia has historically dominated with its InfiniBand-based offerings. However, the industry is increasingly shifting toward Ethernet for AI due to its lower cost, broader ecosystem, and improved performance in scale-out scenarios. Spectrum X represents Nvidia's own bet on Ethernet, and Upscale is riding that wave.
That technical collaboration expanded into a capital relationship, with Nvidia joining Upscale’s latest funding round. Nvidia’s investment is both strategic and financial; it signals that Nvidia sees Upscale as a key partner in building the next-generation AI networking stack. For Upscale, having Nvidia as a backer provides not only capital but also access to technical resources, customer relationships, and market validation. The relationship also helps Nvidia expand its reach beyond its own networking hardware, allowing it to influence the development of complementary technologies.
Software, token efficiency and what comes next
The goal for many organizations today is to optimize AI token usage. That’s a topic often referred to as “tokenmaxxing,” and it’s an area where Khemani noted that networking has a role to play. Efficient token usage directly impacts the cost and speed of AI training and inference. As models grow larger and more complex, the amount of data that must be moved between accelerators increases dramatically. A suboptimal network can lead to idle GPUs waiting for data, wasting both time and money. Upscale is building a full-stack offering that includes software for both large hyperscalers and emerging neocloud providers.
Khemani emphasized that there is not a single right answer to the challenge of token efficiency. He sees it as workload-dependent, and changing as the industry moves from simple prompts to agents onto more complex loops and multi-step applications. “I don’t think we have a choice other than to make sure that AI infrastructure is used in as optimal a fashion as it can,” Khemani said. “So, you have to build an infrastructure that is somewhat flexible to adapt to whatever AI applications require.” This means that Upscale's software stack must be capable of dynamically adjusting network policies based on the workload, whether it's real-time inference for a chatbot or long-running training for a foundation model.
The broader AI networking landscape is becoming increasingly competitive. Established players like Cisco, Juniper, and Arista are all vying for a piece of the pie, while startups like Enfabrica, Groq, and Cerebras are developing their own networking technologies. However, Upscale's focus on both scale-up and scale-out, combined with its tight integration with Nvidia's Spectrum X platform, gives it a unique position. The company is also benefiting from the trend toward disaggregated and composable infrastructure, where accelerators and memory are pooled across the data center. Skyhammer could become a key enabler of that vision, allowing organizations to build flexible, high-performance AI systems that can adapt to changing demands.
Looking ahead, Upscale plans to continue expanding its engineering team and product portfolio. The new funding will be used to accelerate the development of Skyhammer, ramp up production of Spectrum X-based systems, and expand the company's go-to-market efforts. With $500 million in total funding and a $2 billion valuation, Upscale is well-positioned to become a major player in the AI networking market. The company's success will depend on its ability to execute on its product roadmap and win over hyperscale customers, many of whom have historically preferred to build their own networking solutions. But with Nvidia as both a partner and investor, Upscale has a powerful ally in its corner.
Source:Network World News