The Santa Clara, California-based company said the Series A-1 extension brings its total funding to $500 million, less than a year after it emerged with a seed round of more than $100 million. The latest round was led by Premji Invest and drew new backing from Nvidia, Salesforce Ventures, Seligman Ventures and Temasek. Existing investors including Maverick Silicon, Mayfield, Prosperity7 Ventures, StepStone Group and Tiger Global also participated.
The fundraising places Upscale AI among a small but fast-growing group of infrastructure start-ups aiming to reshape the data-centre stack beyond graphics processing units. While AI investment has been dominated by demand for accelerators, the company is targeting the fabric that links chips, memory and storage across vast clusters. That layer has become critical as model training and inference workloads stretch across tens of thousands of processors and require predictable, low-latency data movement.
Upscale AI’s pitch rests on open-standard networking designed for AI-native workloads rather than conventional enterprise traffic. The company is building silicon, systems and software that can connect accelerators, memory pools and storage into a high-performance network fabric. Its approach seeks to reduce congestion, packet loss and coordination delays that can leave costly AI chips idle despite heavy capital spending on compute.
The company is led by chief executive Barun Kar and executive chairman Rajiv Khemani. Khemani previously founded Innovium, a data-centre networking chip company acquired by Marvell Technology in a deal valued at about $1.1 billion. That track record has helped Upscale AI attract early confidence from investors familiar with the economics of high-end switching silicon and hyperscale infrastructure procurement.
The latest financing follows a $200 million Series A announced in January, led by Tiger Global, Premji Invest and Xora Innovation, with participation from investors including Intel Capital and Qualcomm Ventures. That round took the company’s total funding above $300 million and established it as a unicorn within months of launch. The new extension doubles its valuation from that level and gives it a larger balance sheet to move products from evaluation to deployment.
Demand for AI networking is accelerating as hyperscalers, cloud providers and specialist “neocloud” operators build larger clusters for training and serving models. The pressure is not limited to compute supply. Data-centre developers are also competing for power, optical components, high-bandwidth memory, switches and interconnect systems. North American data-centre power demand is projected to more than double from 31 gigawatts in 2025 to 66 gigawatts in 2027, underlining the scale of the build-out.
Networking has become a defining constraint because AI workloads behave differently from traditional cloud applications. Large training runs require synchronised communication across many accelerators, with performance hurt sharply when the network becomes uneven or congested. Inference at scale adds another layer of complexity as companies attempt to serve users quickly while controlling energy, memory and routing costs.
That shift is creating openings for suppliers offering Ethernet-based or open networking alternatives to proprietary systems. Spending on data-centre switches used in AI back-end networks is forecast to exceed $100 billion by 2030. Ethernet is gaining ground across both scale-out and scale-up designs because customers want interoperability across accelerators, software stacks and server vendors, although proprietary fabrics still hold an important position in parts of the market.
Upscale AI faces powerful incumbents. Nvidia is expanding its Spectrum-X Ethernet platform alongside its dominant accelerator business. Broadcom remains deeply embedded in switching silicon. Cisco and Arista Networks are strengthening data-centre networking portfolios for AI clusters, while Marvell is active across custom silicon, optical connectivity and infrastructure chips. Start-ups in adjacent areas are also trying to capture value as customers rethink how compute, networking and storage are assembled.
The investor interest reflects a broader recalibration of AI infrastructure spending. The largest technology companies are committing hundreds of billions of dollars to data centres, power contracts and specialised hardware, but investors are increasingly focused on whether that spending can translate into useful capacity and revenue. A cluster filled with accelerators can underperform if the network cannot move data fast enough, making infrastructure efficiency as important as raw chip supply.
Follow Arabian Post
Select Arabian Post as your preferred source on Google and MSN News for trusted business news and Arab politics and updates.
