In recent months, several AI infrastructure companies have secured single rounds of funding exceeding $100 million. Within this space, AI network communications have become an especially hot sector. On one hand, Silicon Valley startups focused on AI networking are frequently announcing massive funding rounds; on the other, publicly traded AI Networking companies, particularly those in optical communications, are seeing their stock prices surge.
Why is AI network communication gaining so much trACTion? The driving force is fundamental: demand. Model sizes are ballooning, Token Consumption is skyrocketing, and computing power is becoming a scarce resource. To extract more computing power at a lower cost from the hardware side, innovation must hAPPen at the foundational Technology layer. Accelerating chip-to-chip and node-to-node communication to boost the efficiency of the entire computing infrastructure is a path that is rapidly being validated.
One company riding this wave is Upscale AI. It raised a 200 million Series A round led by Tiger Global, Premji Invest, and Xora Innovation, with participation from Maverick Silicon, StepStone Group, Mayfield, Prosperity7 Ventures, Intel Capital, and Qualcomm Ventures. Most recently, reports indicate the company is in talks to raise an additional 200 million.
Massive Parameters, MoE, and Long Context Windows: Model Innovation Demands AI Networking Innovation
How can a company less than a year old attract such massive funding rounds in quick succession? Its founding team holds a significant part of the answer. Upscale AI was incubated out of Auradine, an emerging AI Infrastructure company now known as Velaura AI, which focuses on delivering proven, ultra-low-power computing solutions for cloud, edge, and Physical AI applications. Upscale AI's Co-founder and CEO, Barun Kar, was previously the COO of Auradine. Co-founder and Executive Chairman Rajiv K was previously Auradine's CEO and now serves as CEO of Velaura AI. CTO Puneet Agarwal spent a decade at Broadcom and also served as CTO of the data center division at Marvell. Both Barun Kar and Rajiv K also have experience at large enterprises prior to their previous venture. In short, this is a deeply experienced team with decades of industry knowledge.
Why is AI networking so important? The answer lies in the technical fundamentals. AI computing workloads are highly Synchronous. Modern workloads such as large-scale model training, MoE architectures, and distributed inference place immense synchronous pressure on the network. During training, parameter gradients must be transferred among thousands of GPUs in highly synchronized bursts. Inference generates massive fan-out traffic while demanding extremely strict latency requirements. If the network cannot keep pace, GPUs stall and wait, latency soars, and the efficiency of the entire computing cluster collapses. This is an architectural mismatch, not a problem that can be fixed with simple tuning.
Traditional networks that prioritize versatility introduce complexities that become liabilities in AI scenarios. Deterministic communication and the strong synchronization required by GPU collective Operations are exceeding the design limits of traditional networks. The network needed for an AI computing cluster must support deterministic, synchronous, and high-throughput communication at a massive scale. AI networking must be rebuilt from the ground up, designed around the real requirements of Scale-Up and Scale-Out connections.
Breaking this down further leads us to the models themselves. Two characteristics of modern models place immense pressure on cluster networks: the exponential growth of parameter sizes and the continued evolution of long context windows and Chain-of-Thought reasoning. For example, the recently released DeepSeek V4 Pro has 1.6 Trillion Parameters and a 1-million-token context window. A model with 1.6T parameters requires 1.6TB of mEMOry, which far exceeds the capacity of a single GPU. It must be sharded across a large number of accelerators, making chip-to-chip communication an immediate bottleneck. The ultra-long context window causes the KV cache to balloon, also exceeding the HBM capacity of a single GPU. Both of these factors create intense dual pressure on memory capacity and communication bandwidth.
A Full-Stack Revolution, Not Just Chip-Level Innovation
To enable smooth training and inference for these large-parameter, long-context models, the real solution is to reDeFine the "computing boundary." This means connecting more GPUs via an ultra-high-speed network with sub-microsecond latency and high-throughput collective communication capabilities, allowing them to function as a single "super GPU." The rack form factor emerged from this need. Take NVIDIA's NVL72 as an example: it no longer treats the 72 GPUs as independent devices but operates them as a single, coherent machine with memory-semantic capability, boasting an internal NVLink bandwidth of 130TB/s.
This introduces two critical connectivity layers in AI infrastructure: rack-level GPU interconnects (Scale-Up) and cluster-level network fabric interconnects (Scale-Out). These two layers must operate in concert to enable thousands of GPUs to work efficiently as a unified, distributed computing engine.
To address these two layers, Upscale AI has developed a network architecture purpose-built for AI. For rack-level interconnects (Scale-Up), it has the SkyHammer chip architecture; for cluster-level fabric (Scale-Out), it has Open Ethernet.
SkyHammer is a chip architecture designed to break through Scale-Up networking bottlenecks. Based on open standards, its goal is to deliver deterministic latency, extreme bandwidth, and predictable performance at hyperscale, enabling GPUs and XPUs to work in concert as a highly synchronous computing engine. One of its key features, deterministic latency, means the time for data to travel between components within a rack is highly predictable. SkyHammer is built from the ASIC level up, with coordinated design across the chip, system, and rack levels to ensure every layer works in hArmony. Every link is re-engineered: from how data flows within the chip, to how the fabric adapts under load, to how a super-cluster maintains predictability under immense synchronization pressure. It supports emerging Standards like ESUN, UEC, and UALink, while also leaving room for future innovations. With its flexible architecture, SkyHammer can smoothly adapt to new standard definitions without reconstruction or compromise. Products based on the SkyHammer architecture are planned for release in 2026.
Open Ethernet is designed for cluster-level AI fabrics (Scale-Out). At the cluster level, AI systems demand openness, interoperability, and massive bandwidth. Upscale AI has built an AI-optimized Open Ethernet fabric that will be constructed on Nvidia Spectrum-X Ethernet switch chips and the SONiC network operating system, providing end-to-end support. By integrating ASIC-native telemetry, deterministic lossless Ethernet behavior, and industry-standard network workflows, the system delivers predictable performance, simplified operations, and high reliability at scale. In essence, it connects thousands of GPUs into a unified, high-performance fabric to support distributed training and large-scale inference. For this project, Upscale AI has joined the NVIDIA Partner Network and is working closely with NVIDIA and its ecosystem to accelerate the deployment of large-scale AI Data Center networks.
Notably, Upscale AI hasn't stopped at building a faster network chip. It achieves tight coupling among its chips, systems, and software. running a large-scale AI computing cluster requires continuous insight into the congestion state, synchronization behavior, and GPU utilization across the entire network fabric. This encompasses high-performance RDMA networking, adaptive congestion management, GPU-directed telemetry and observability, and real-time operational visibility across the entire fabric. Upscale AI optimizes all these links to build the deterministic networking foundation necessary for modern AI clusters.
The Mismatch Between Model Demands and Infrastructure Creates Opportunities
The AI infrastructure sector still has immense growth potential. In fact, it is likely in a state of long-term, alternating innovation with AI software, especially models. When model architecture evolves and creates a structural mismatch with hardware or software, new opportunities arise.
The current landscape reflects this perfectly. The combination of MoE Architectures, massive parameters, long context windows, and the insatiable need for tokens from AI Agents is creating a supply-demand crunch for computing power, while simultaneously sparking infrastructure innovation. In the computing chip space alone, over the last six months, we have tracked Unconventional AI (raising 500 million). In AI-driven chip design, there is Ricursive (raising 60 million). And, of course, the data center network interconnect space featuring companies like Upscale AI (having raised 200 million), Eridu (raising 90 million).
While China's open-source AI models, notably the recently released DeepSeek V4, have achieved global leadership, the country's AI infrastructure layer is still in catch-up mode. This gap, however, also signals enormous room for innovation. Observing China's venture capital market, a large number of innovative companies are emerging, and some have already achieved preliminary success.
Comments & Questions (0)
No comments yet
Be the first to comment!