AI infrastructure startup Cornelis raises $205 million, challenging NVIDIA's CUDA ecosystem hegemony with an open interconnect architecture—the second battleground beyond chips.

Cornelis, a company specializing in AI chip interconnect networks, announced on September 14th the completion of a $205 million funding round led by IAG Capital Partners. On the same day, it unveiled a network product called Active Compute Fabric. The company's specific focus is addressing the issue in large model training and inference where significant GPU time is wasted "waiting for data to arrive," leading to idle chips and suboptimal utilization. Cornelis aims to use a smarter network fabric to enable chips to process and send data simultaneously, turning waiting time into parallel processing.

Cornelis is not a newcomer without a foundation—it spun off from Intel in 2020, carrying with it the accumulated expertise in interconnects from a major player. Its approach to countering NVIDIA is not by creating a more powerful GPU but by providing an open architecture: customers can mix and match GPUs and accelerator cards from various vendors into Cornelis's network fabric. Technically, NVIDIA's chips can run on other networks, but they are most optimized for their own software, making a full-stack solution the most hassle-free for customers. Cornelis aims to extract the "network" from NVIDIA's full stack, creating a replaceable and negotiable "second battleground."

NVIDIA's Moat Lies in Software, Not Chips

When the world looks at NVIDIA, it focuses on the stacks of GPUs filling servers. However, what is truly difficult to dislodge is the CUDA ecosystem's two-decade accumulation of frameworks, libraries, and toolchains—almost all AI software is built on this software stack. Chip strength is temporary; ecosystem stickiness is enduring. The bet of companies like Cornelis is clear: since catching up in single-chip performance is unattainable, they will cut into (make an incision) at the "collaboration layer," using open architectures to reduce customer lock-in to a single vendor. This is not about overthrowing NVIDIA overnight but about prying open cracks line by line with open code—the capital is willing to bet on this niche path of "de-NVIDIA-ization."

Cornelis's Strategy: Turning "Waiting" into "Parallelism"

Image Source: TechCrunch

Active Compute Fabric addresses the chronic issue of GPU idleness. In traditional approaches, chips send results to the next stage only after completing a segment, making the network an invisible bottleneck; the larger the Chiplet and cluster scale, the higher the proportion of time spent moving data. Cornelis enables "processing" and "sending" to overlap, squeezing higher effective throughput from the same hardware. The company has already shipped its first-generation product, with a new generation expected later this year. For cloud providers, this directly impacts the cost and energy efficiency of AI training and inference—in an era where "computational power is authority," interconnect efficiency is itself a competitive edge.

Implications for China's Computational Autonomy

Shifting the lens back to China, the practical implications of this business are clear. China's computational buildup relies on pursuits like Ascend, Cambrian, and Hygon, all grappling with the same challenges of "strong chips, weak ecosystems, and fragmented interconnects." While U.S. export controls target advanced manufacturing processes, they cannot block "soft breakthroughs" like open interconnect architectures. For domestic computational power to truly succeed, it must not only chase peak single-chip performance but also integrate network fabrics, software stacks, and heterogeneous mixed training. Cornelis's valuation logic reminds us: the value of AI infrastructure is shifting from "selling chips" to "enabling efficient chip collaboration." Whoever controls the collaboration layer holds the master switch for computational cost and scale.

$205 million may not be a huge sum in the AI infrastructure space, but the signal is clear—competition is spreading from individual chips to the entire network. NVIDIA's hegemony will not collapse overnight, but cracks are being pried open line by line with open code. For China, this is both a reminder and an opportunity: the next phase of computational autonomy lies not in isolated chips but in interconnects. Welding thousands of chips into an efficiently collaborative network is the true challenge and turning point of this battle. When "waiting for data" time is minimized, what is saved is not just electricity but the pace of the entire intelligent industry.

Today's Golden Quote

"NVIDIA's hegemony lies not in chips but in welding them into a software stack. The next phase of AI infrastructure competition is about who can make thousands of chips collaborate more tightly—the key to computational autonomy lies in interconnects."

Follow [Degaoxing Zhiqinglang] for three daily hardcore tech analyses, understanding the industry and national fortune behind the technology.