Key Takeaways

  • Upscale AI raised $190 million in Series A-1 funding, reaching $500 million total and a $2 billion valuation.
  • The company is developing open-standard networking to improve communication across heterogeneous AI chips.
  • Growing enterprise AI workloads and competition from Nvidia, Broadcom, Arista Networks, and Nexthop AI are shaping the high-stakes market.

When PCs connected to the internet in the 1990s, companies like Cisco and Juniper became the connective tissue of that era. When cloud computing took off, Arista and Broadcom rose to prominence. Now, Upscale AI is building networking infrastructure for the next shift, driven by AI workloads and the demands of modern GPU clusters.

The Santa Clara startup closed a $190 million Series A-1 round, bringing total investment to $500 million and pushing its valuation to $2 billion. That milestone arrived in under 18 months. Premji Invest led the round, and new investors included Nvidia, Salesforce Ventures, Temasek, and Seligman Ventures. Existing backers like Mayfield, Tiger Global, StepStone, Maverick Silicon, and Prosperity7 also participated.

Analysts project strong growth in the AI networking infrastructure market, with spending on Ethernet switches for AI and high-performance computing rising more than 25% year-over-year, according to IDC 2024. Additionally, enterprise network traffic driven by AI workloads is projected to increase more than 30% by 2028, based on findings from Gartner. This pressure forces companies to rethink how they build data center fabrics to support the scale and synchronization required for training large AI models.

Dell’Oro Group forecasts spending on AI data center switches could surpass $100 billion annually by 2030. Meanwhile, Microsoft, Google, Meta, and Amazon are expected to heavily increase their infrastructure spending in 2026, almost doubling their 2025 expenditures. That creates a giant opening for vendors who can deliver specialized solutions for GPU-to-GPU communication.

The company is pursuing one central objective: reduce the friction that exists when GPUs and accelerators from different manufacturers operate within the same cluster. Today, Nvidia’s NVLink and InfiniBand technologies dominate the landscape and often require customers to use Nvidia hardware end-to-end. The startup's cofounder and executive chairman stated that future systems will be more heterogeneous, noting that customers seek flexibility and open architectures instead of single-vendor lock-in.

The cofounder and CEO added that GPUs from different vendors need interoperability to communicate seamlessly. This open-standard networking fabric aims to make that a reality, enabling chips from various manufacturers to communicate at full speed. The approach aligns with a broader move toward open standards in networking, including IEEE 802.3 Ethernet specifications and RDMA over Converged Ethernet, both of which underpin low latency and high throughput in AI clusters. Several companies across the ecosystem, including AMD, Intel, Google, Meta, and Microsoft, back the open networking standard that the organization is building upon.

An investor from Premji Invest noted that the compute layer was not originally designed for generative AI workloads, and by extension, networking, storage, and caching layers were also not built for these requirements. Every layer of the modern data center stack is being rethought. This mirrors findings from the Cloud Native Computing Foundation, which observed that over 60% of surveyed organizations are modernizing network and infrastructure tooling to better support containerized and AI workloads.

Custom chip development is expensive, often costing hundreds of millions of dollars at advanced nodes before any product makes it to market. Suppliers also require multi-year commitments, including paid reservations for manufacturing capacity. The firm has expanded to several hundred employees in less than a year, requiring substantial financial stability to support this rapid growth.

Nexthop AI, a direct competitor focused on custom AI networking for hyperscalers and neocloud customers, raised $500 million in Series B funding in March at a $4.2 billion valuation. Meanwhile, Nvidia and Broadcom continue to invest heavily. Both have massive research and development budgets, and Nvidia remains in a powerful position because its proprietary networking technologies are widely deployed in AI training clusters.

Whether open standards can match or exceed the performance of Nvidia’s proprietary technologies in production environments remains to be seen. Yet the appetite for alternatives is clearly present, especially among hyperscalers that run vast AI workloads. This funding round indicates a shift toward a highly competitive race to build the connective infrastructure for the next generation of computing.