AI Infrastructure Is Diversifying Beyond Nvidia's Stack

AI Infrastructure Is Diversifying Beyond Nvidia's Stack

Most of the conversation around AI infrastructure over the last few years has centered on one company: Nvidia. Its GPUs and its tightly integrated networking stack have effectively defined how large AI clusters are built. That's starting to shift, as new players raise real money to offer alternatives — and the reasons why matter for any business planning long-term AI infrastructure spend.

A new interconnect option enters the picture

Cornelis Networks, a company spun out of Intel, recently raised $205 million and introduced what it calls Active Compute Fabric — an open, GPU-agnostic networking layer built specifically for AI clusters. The pitch is straightforward: instead of locking customers into Nvidia's own interconnect technology, Cornelis is offering a networking layer that can sit underneath GPUs from multiple vendors, giving data center operators more flexibility in how they build and scale AI training and inference clusters.

This matters because networking, not just raw GPU compute, is one of the biggest bottlenecks in large-scale AI training. When you're coordinating work across thousands of GPUs, how fast and efficiently they can talk to each other has a direct effect on training time and cost. Historically, getting the best-performing interconnect has meant buying into Nvidia's full stack. An open, vendor-agnostic alternative changes that calculus.

Why this is bigger than one company's funding round

The emergence of credible alternatives to a single dominant vendor's stack is a pattern familiar to anyone who has watched infrastructure markets mature. It tends to happen once demand is large and stable enough to support competition, and once customers start pushing back on the cost and lock-in of a single-vendor approach. AI infrastructure spending has grown enormously over the past few years, and with that growth has come more scrutiny from the enterprises and cloud providers footing the bill.

For large buyers — cloud providers, big enterprises building their own AI infrastructure, and AI labs training frontier models — vendor diversification offers a few concrete benefits:

• Negotiating leverage. A credible second (or third) option for critical infrastructure components gives buyers more room to negotiate on price and terms.
• Reduced lock-in risk. Betting an entire infrastructure strategy on one vendor's roadmap is risky if that vendor changes pricing, prioritization, or product direction.
• Room for specialization. An open networking layer that isn't tied to a single GPU vendor makes it easier to mix hardware from different suppliers as needs change.

What this means for smaller businesses too

Most small and mid-sized businesses aren't building their own AI training clusters — that's the domain of cloud providers and large enterprises. But the health of AI infrastructure markets still affects everyone who consumes AI as a service. More competition among the companies building the underlying infrastructure tends to put downward pressure on the cost of running AI workloads over time, and it reduces the systemic risk of the entire industry depending on a small number of infrastructure vendors.

It's also a useful reminder for any business evaluating AI vendors of its own: whether you're choosing a cloud AI provider, a model API, or a specialized AI tool, the same lock-in questions apply at a smaller scale. Understanding whether a vendor's approach is open and portable, or proprietary and closed, is worth factoring into procurement decisions — not just for infrastructure giants, but for any company building meaningful dependency on an AI vendor.

The bigger trend

AI infrastructure is still a young, fast-moving market, and one company's earlier dominance doesn't guarantee that it stays uncontested. As more capital flows into alternatives — whether in networking, chips, or the software layers running on top of them — buyers across the spectrum, from hyperscalers down to businesses simply choosing which AI tools to adopt, benefit from having genuine options rather than a single default choice.