The Ultra Ethernet Rebellion: Big Tech's War on Nvidia's NVLink

Everyone looks at the GPU. When analysts discuss Nvidia’s trillion-dollar valuation, they point to the H100 or the new Blackwell chips. But the silicon is only half the story. Nvidia’s most impenetrable moat isn’t the processor; it is the proprietary plumbing connecting those processors together. Now, a coalition of tech giants is building a battering ram called the Ultra Ethernet Consortium to tear that moat down.

Before you read: We previously explored how physical bandwidth limits are forcing a pivot to Silicon Photonics and how memory pooling is breaking the server chassis with CXL. This article tackles the proprietary networking protocol strangling those exact clusters.

Table of contents

• The $6.9 Billion Heist
• The Ultra Ethernet Consortium (UEC)
• Packet Spraying and the Math of Congestion
• Fun facts
• Will it actually work?
• Sources

The $6.9 Billion Heist

To understand the current war, you have to look at 2019. Nvidia bought Mellanox, a networking company specializing in InfiniBand, for $6.9 billion. At the time, some analysts thought Jensen Huang overpaid. In retrospect, it was the heist of the century.

AI training doesn’t happen on a single GPU. It happens across tens of thousands of them, acting as one massive supercomputer. If GPU A is waiting for a calculation from GPU B, it sits idle. In AI infrastructure, idle compute is literally burning millions of dollars.

Standard Ethernet was built for the internet—it drops packets, re-routes traffic inefficiently, and introduces latency spikes. Standard Ethernet dropping a packet during a massive AI training run is like pausing a Formula 1 race because someone dropped their car keys on the track.

InfiniBand, and by extension Nvidia’s proprietary NVLink, solved this. They provide lossless, ultra-low-latency communication. But the catch is brutal: if you want that performance, you buy Nvidia GPUs, Nvidia NICs (Network Interface Cards), Nvidia switches, and Nvidia cables. You are locked into their ecosystem from the silicon to the wall socket.

The Ultra Ethernet Consortium (UEC)

Frankly, it is embarrassing it took the industry this long to realize that relying on a single vendor for the entire backbone of the AI revolution was a suicidal business strategy.

In 2023, AMD, Intel, Broadcom, Meta, and Microsoft finally panicked enough to form the Ultra Ethernet Consortium (UEC). Their goal is simple: rebuild Ethernet from the ground up so it can match or beat InfiniBand, without the proprietary lock-in.

Historically, data centers tried to fix Ethernet for high-performance computing using RDMA over Converged Ethernet (RoCE). But configuring RoCEv2 in a massive cluster requires so much manual tuning that it is less computer science and more dark magic. You don’t configure it; you negotiate with it.

Packet Spraying and the Math of Congestion

The UEC is rewriting the transport layer to handle what network engineers call “elephant flows”—massive, sustained bursts of AI data.

The core innovation is “packet spraying.” Traditional Ethernet sends a file down a single path. If that path hits a traffic jam, the whole transfer slows down. UEC breaks the data into pieces and sprays them across every available physical path in the network simultaneously, reassembling them at the destination.

It completely eliminates network bottlenecks. It also introduces out-of-order packet delivery, which standard Ethernet hates, but UEC handles natively in hardware. It is the equivalent of sending 100 delivery trucks down 100 different side streets to avoid highway traffic, knowing exactly how to park them perfectly at the destination.

Fun facts

• Did you know Nvidia’s networking division alone generates more revenue today than the entirety of AMD’s gaming segment? The cables are literally more profitable than PlayStations.
• InfiniBand cables are so thick and heavy that data center architects literally have to reinforce the raised floors just to prevent the networking racks from crashing into the basement.

Will it actually work?

The hyperscalers are motivated by pure financial terror, which is usually a great catalyst for engineering breakthroughs. Microsoft and Meta do not want to hand 60% of their capital expenditure budgets to Jensen Huang for the rest of the decade.

But Nvidia has a massive head start. NVLink and InfiniBand are battle-tested at the 100,000-GPU scale today. The UEC 1.0 specification was only recently finalized, and actual silicon supporting it is just starting to tape out.

Hardware takes time. You cannot software-patch a switch to support massive multi-path packet spraying; it requires new ASICs. The open-source rebellion is finally armed and organized, but they are fighting an enemy that currently owns the battlefield.

Sources

• The Ultra Ethernet Consortium: UEC 1.0 Specification Architecture
• SemiAnalysis: Nvidia’s Networking Moat and the InfiniBand Monopoly
• ServeTheHome: Broadcom’s Tomahawk 5 and the Future of AI Networking
• IEEE Spectrum: Why Ethernet is Failing AI Data Centers