OpenAI’s MRC Protocol: Why Data-Center Networking Is the New AI Training Battleground

Why GPUs Sit Idle, and Why That’s Now a Network Problem

Large-scale AI training clusters are among the most expensive pieces of computing infrastructure on Earth. Yet a significant fraction of their GPUs is often not doing useful work. The reason is not always a lack of model code or training data; it is the network. When a single link between switches, NICs, or pods becomes congested or fails, the standard recovery path can stall the flow of gradients, activations, and checkpoints. A stalled pipeline leaves expensive accelerators waiting, and waiting accelerators burn capital, power, and time.

That is the problem OpenAI is now trying to fix at the protocol layer. On May 7, the company joined AMD, Broadcom, Intel, Microsoft, and NVIDIA to announce the Multipath Reliable Connection (MRC) open network protocol. The stated goal is to make large-scale AI training clusters faster and more reliable by reducing the amount of GPU idle time caused by network problems.

MRC: Multiple Paths, One Connection

MRC is short for Multipath Reliable Connection. In a conventional cluster, a reliable connection is tied to one path. If that path is slow or broken, the connection waits or retries. MRC changes that assumption: a single logical connection can use multiple physical paths at once, and traffic can be moved to an alternate path when the primary one degrades.

The mechanism is simple to describe but hard to implement at scale. Training workloads are synchronized across many GPUs; a delay in one slice of the network can ripple outward. By allowing data to switch automatically when a link is congested or fails, MRC aims to keep the aggregate pipeline full. The open question is not whether multipath networking is useful, but whether it can be made performant and interoperable across the heterogeneous hardware stack that now fills modern data centers.

The Vendor List Is the Message

The most striking detail about the announcement is not a technical spec, but the roster of co-signers. AMD, Broadcom, Intel, Microsoft, and NVIDIA are not companies that normally agree on much. They compete for sockets, for cloud share, and for AI training market position. When they all show up on the same protocol, it is a signal that the network layer has become a common bottleneck too big for any single vendor to solve.

• AMD, Intel, and NVIDIA design the accelerators, NICs, and switch silicon that move packets.
• Broadcom supplies the switching and interconnect ecosystem that knits racks together.
• Microsoft operates one of the largest AI training clouds and is a major backer of OpenAI.

Their participation means MRC is not a niche research proposal. It is an attempt to make multipath resilience a baseline expectation of AI data-center networking, much like RDMA became a baseline for high-performance computing. If the protocol succeeds, the entire stack, from NIC firmware to switch ASICs to cloud orchestration, will have to support it.

From Throughput to Antifragility

The MRC announcement landed on the same day as another infrastructure move: Anthropic’s compute partnership with SpaceX, which allowed Claude Code and API quotas to be raised. The SpaceX deal is about raw capacity; MRC is about efficiency. Together, they illustrate the same shift in how AI companies think about infrastructure.

The early phase of the AI buildout was about peak throughput: more FLOPS, more bandwidth, more GPUs. The next phase is about reliability under pressure. Capacity is no longer the only scarce resource; so is predictability. Training runs can last long enough that a single network incident can force a checkpoint restart, wasting days of work. The cost of such failures includes not just the lost compute, but also the research schedule and the model release window. MRC is an attempt to buy what Nassim Taleb would call “antifragility”: a system that absorbs disruptions rather than being broken by them.

Networking Is the New Battleground

If MRC is widely adopted, the competitive landscape of AI infrastructure will shift. For the next few years, model architecture and training algorithms will still matter, but the moat will increasingly be built on the ability to keep large clusters running at high utilization. That is a problem of data-center networking, scheduling, and physical logistics.

The same day’s news also included a smaller but related point: Cursor 3.3 introduced context-usage diagnostics for agents, and Cloudflare now lets agents create accounts and deploy services. These are reminders that AI is moving from demos to operational workloads. When a failure costs money or time, the quality of the underlying infrastructure becomes the product. MRC is the clearest sign yet that the networking layer is where that quality will be won or lost.