The Model and Agent Innovation Boom Forces AI infrastructure to Evolve, Creating a New Wave of entrepreneurial Opportunities
Over the past six months, several AI Infrastructure companies have secured single rounds of financing exceeding $100 million. Within this space, AI network communications have become a particularly hot sector. On one hand, Silicon Valley startups focusing on AI networking are closing massive funding rounds at an unprecedented frequency. On the other, publicly traded companies in this domain—especially those in optical communications—have experienced rapid share price growth.
Why is the spotlight intensifying on AI network communications? The answer lies fundamentally in demand. Models are becoming exponentially larger, Token Consumption is surging, and compute power is increasingly scarce. To extrACT more computational capability at a lower cost from the hardware layer, innovation must hAPPen at the foundational Technology level. A proven path forward is accelerating chip-to-chip and node-to-node communication to boost the overall efficiency of computing infrastructure.
Upscale AI is a prime example of a company riding this wave. The startup secured a 200 million in a Series A round led by Tiger Global, Premji Invest, and Xora Innovation, with further backing from Maverick Silicon, StepStone Group, Mayfield, Prosperity7 Ventures, Intel Capital, and Qualcomm Ventures. It is now reportedly in discussions to raise another 200 million in fresh funding.
Founded less than a year ago, Upscale AI's ability to attract massive Investment is closely tied to its founding team. The company was spun out of Auradine—itself a rising AI infrastructure firm now renamed Velaura AI, which focuses on delivering verified, ultra-low-power compute solutions for cloud, edge, and Physical AI applications. Upscale AI’s co-founder and CEO, Barun Kar, previously served as COO of Auradine, while co-founder and Executive Chairman Rajiv K. was the former CEO and currently leads Velaura AI. CTO Puneet Agarwal spent a decade at Broadcom and later served as CTO of the data center division at Marvell. Both Kar and Rajiv K. also held senior roles at major corporations before their previous venture, forming a deeply experienced, industry-veteran team.
Large Parameters, MoE, and Long Context: Model Innovation Drives AI Networking Overhaul
To grasp the importance of AI networking, one must underStand the underlying technology. AI workloads are characterized by high Synchronicity. Modern workloads such as large-scale model training, Mixture of Experts (MoE) architectures, and distributed inference place extreme synchronization stress on the network. During training, model parameter gradients must be transferred among tens of thousands of GPUs in highly synchronized bursts. Inference generates massive fan-out traffic while demanding ultra-low latency. If the network cannot keep pace, GPUs stall and wait, latency spikes, and the efficiency of the entire compute cluster collapses. This is an architectural mismatch, not a problem that mere optimization can resolve.
Traditional networks prioritize versatility, but the complexity introduced by accommodating diverse workloads has become a hindrance in AI scenarios. The need for deterministic communication and the strict synchronization required by GPU collective communication are exceeding the design limits of conventional networking. AI Compute clusters demand networks that can support deterministic, synchronized, high-throughput communication at massive scale. This means AI networking must be fundamentally rebuilt from the ground up, designed around the real requirements of Scale-Up and Scale-Out connections.
A deeper look points directly to the models themselves. Two characteristics place extreme pressure on AI cluster networks: the exponential growth in model parameters and the continuous evolution of long-context Windows and Chain-of-Thought (CoT) reasoning. Take the recently released DeepSeek V4 Pro as an example: it features 1.6 Trillion Parameters and a context window of 1 million tokens. A 1.6T model requires 1.6TB of mEMOry, far exceeding the capacity of a single GPU, forcing the model to be sharded across multiple accelerators and making chip-to-chip communication an immediate bottleneck. The ultra-long context window causes a Dramatic expansion in the KV cache volume, which also surpasses the HBM memory capacity of a single GPU. These factors create a dual squeeze on memory capacity and communication bandwidth.
A Full-Stack Revolution, Not Just Chip-Level Innovation
The true solution for training and smooth inference of models with massive parameters and ultra-long context windows lies in reDeFining the "compute boundary." This requires connecting more GPUs into what functions as a "super GPU," utilizing ultra-high-speed networking with sub-microsecond latency and high-throughput collective communication capabilities. This gives rise to the rack-scale architecture. Take NVIDIA's NVL72 as an example: it treats 72 GPUs not as independent devices but as a unified machine operating with memory semantics, featuring an internal NVLink bandwidth of 130TB/s. This introduces two AI infrastructure connection layers: rack-scale GPU interconnect (Scale-Up) and cluster-scale fabric interconnect (Scale-Out). These two layers must operate in concert to make thousands of GPUs work efficiently as a single, distributed computing engine.
Upscale AI has developed a network architecture tailored specifically for these two connection layers. For rack-scale AI interconnect (Scale-Up), it offers the SkyHammer chip architecture. For cluster-scale AI networking (Scale-Out), it provides Open Ethernet. SkyHammer is a chip architecture engineered to break through Scale-Up AI networking bottlenecks. Built on open standards, it targets deterministic latency, extreme bandwidth, and predictable performance at hyperscale, enabling GPUs and XPUs to function as a highly synchronized compute engine. One of its haLLMark features is deterministic latency, meaning the time required for data transmission between components within the rack can be predicted and controlled with high certainty.
SkyHammer is constructed from the ASIC level upward, using a holistic co-design approach across the chip, system, and rack layers to ensure seamless orchestration. Every link has been re-engineered—from how data flows inside the chip, to how the fabric adapts under load pressure, to how superclusters maintain predictability under heavy synchronization stress. It supports emerging standards like ESUN, UEC, and UALink, while also reserving headroom for future innovations. With its flexible architecture, SkyHammer can smoothly adapt to new standards without reconfiguration or compromise, achieving interoperability in diverse, open environments without sacrificing performance. Products based on the SkyHammer architecture are planned for release in 2026.
Open Ethernet targets cluster-scale AI fabric (Scale-Out). At the cluster level, AI systems require openness, interoperability, and massive bandwidth. Upscale AI is building an AI-optimized Open Ethernet fabric. This system will be built on Nvidia Spectrum-X Ethernet switch silicon and the SONiC network operating system, providing end-to-end support. By integrating ASIC-native telemetry capabilities, deterministic lossless Ethernet behavior, and industry-standard network workflows, the system delivers predictable performance, simplified Operations, and high reliability at scale. In short, it connects thousands of GPUs into a unified, high-performance network to support distributed training and large-scale inference. For this initiative, Upscale AI has joined the NVIDIA Partner Network and is working closely with NVIDIA and its ecosystem partners on Reference Architectures and validated designs to accelerate the deployment of hyperscale AI Data Center networks.
Upscale AI's work clearly extends beyond building a faster network chip; it pursues tight coupling between chip, system, and software. Operating large-scale AI compute clusters requires continuous insight into congestion status, synchronization behavior, and GPU utilization across the entire network fabric. This includes high-performance RDMA networking, adaptive congestion management, GPU-oriented telemetry and observability, and real-time operational visibility covering the complete fabric. Upscale AI is optimizing across all these areas to construct the deterministic networking foundation essential for running modern AI compute clusters.
A Mismatch Between Model Requirements and AI Infrastructure Creates Diverse Entrepreneurial Opportunities
AI computing infrastructure holds enormous growth potential. In fact, it may remain in a state of alternating innovation with AI software, particularly models, for the long term. Each time model architecture evolves and creates a structural mismatch with the hardware or software of AI infrastructure, new opportunities emerge. The current landscape reflects precisely this dynamic: the combined forces of MoE architectures, massive parameters, ultra-long context windows, and the agent-driven thirst for tokens are creating a scenario where AI compute supply is falling short of demand—while simultaneously opening doors for infrastructure innovation.
In the compute chip segment alone, we have tracked Unconventional AI (which raised 500 million) in the past six months. In the AI-driven chip design space, we’ve seen Ricursive (raising 60 million). And in AI data center networking, alongside Upscale AI (which has already raised 200 million), there are players like Eridu (raising 90 million).
China has already achieved global leadership in open-source AI models, notably with the recent release of DeepSeek V4. In the realm of AI infrastructure, China is currently in a catching-up phase, but this very fact signals immense room for innovation. Observing the Chinese venture capital market, a large number of innovative companies have begun to emerge, and a portion of them have already seen initial success.
Comments & Questions (0)
No comments yet
Be the first to comment!