The artificial intelligence revolution has moved beyond the era of massive models and into the era of massive infrastructure. As the industry pushes toward artificial general intelligence (AGI), the bottleneck is no longer just the number of parameters in a neural network or the raw FLOPS of a single GPU. The primary constraint has become the network—specifically, the ability to move staggering amounts of data between hundreds of thousands of accelerators with near-zero latency.
To solve this, a powerful consortium led by OpenAI and supported by the “titans of silicon”—AMD, Broadcom, Intel, Microsoft, and NVIDIA—has unveiled MRC (Multipath Reliable Connection). This protocol represents the most significant shift in data center networking in over a decade, specifically engineered to support the “gigascale” clusters required for the next generation of AI.
The Networking Wall: Why Standard Ethernet Failed AI
To understand why MRC is necessary, one must understand the unique demands of distributed AI training. Training a model like GPT-5 or its successors involves a process called “All-Reduce” synchronization. In this model, thousands of GPUs work on a slice of a problem, then must stop and share their results with every other GPU before moving to the next step.
In a traditional network, if one packet of data is delayed or lost (a phenomenon known as “tail latency”), the entire training cluster—worth billions of dollars—sits idle. Standard Ethernet was designed for the internet, where if a webpage takes 50 milliseconds longer to load, the user barely notices. In AI training, that 50ms delay, multiplied across millions of synchronization steps, results in weeks of lost time and millions of dollars in wasted electricity.
Before MRC, the industry relied on two main paths:
- InfiniBand: Highly efficient but expensive, difficult to scale to hundreds of thousands of nodes, and largely controlled by a single vendor.
- RoCEv2 (RDMA over Converged Ethernet): An attempt to bring InfiniBand-like speeds to Ethernet, but prone to “incast congestion” and complex to manage at scale.
MRC is the evolution that takes the best of these worlds and optimizes them for the “gigascale” era.
What is MRC? A Technical Deep Dive
Multipath Reliable Connection (MRC) is an open-standard networking protocol designed to sit on top of standard Ethernet physical layers but replace the traditional ways data is routed and recovered. It is being contributed to the Open Compute Project (OCP), ensuring it becomes the industry standard rather than a proprietary tool.
1. True Multipathing (Packet Spraying)
Traditional networks typically use “flow-based” routing. If GPU A talks to GPU B, all packets for that conversation follow a single path. If that path becomes congested, the connection slows down even if other paths in the data center are empty.
MRC introduces Packet-Level Multipathing. Instead of sending a stream of data down one lane, it “sprays” individual packets across every available path in the network simultaneously. This ensures that the total bandwidth of the data center fabric is utilized at nearly 100% efficiency. The hardware at the receiving end is then responsible for reassembling these packets in the correct order.
2. Microsecond-Scale Recovery
In a cluster with 100,000 GPUs, hardware failures are not a rare occurrence—they are a mathematical certainty. Cables fail, transceivers overheat, and line cards glitch.
Standard protocols often take milliseconds or even seconds to “time out” and realize a path is dead. MRC features Hardware-Based Fast Failover. Because the protocol is aware of all paths simultaneously, if one path drops, the network hardware detects the failure in microseconds and immediately reroutes the remaining packets to healthy paths. The training run never sees a “hiccup.”
3. Congestion Control and Incast Management
One of the biggest killers of AI performance is “incast,” where multiple GPUs try to send data to a single receiver at the same time, overwhelming the receiver’s buffers.
MRC utilizes advanced, telemetry-driven congestion control. It uses real-time data from the network switches to “throttle” or “steer” traffic before a buffer overflow occurs. Unlike previous iterations of Ethernet that used “Pause Frames” (which could cause network deadlocks), MRC uses a more surgical approach, slowing down only the specific traffic causing the issue without stopping the entire fabric.
The Strategic Alliance: Why These Six Companies?
The composition of the MRC group is as important as the technology itself. Each member brings a critical piece of the puzzle:
- OpenAI: The “Customer Zero.” OpenAI defines the requirements. They know exactly how their models break traditional networks and provided the performance targets MRC had to hit.
- Microsoft: The “Global Builder.” Through Azure, Microsoft provides the massive real-world laboratory to deploy and test MRC at a scale few others can match.
- NVIDIA: While NVIDIA has its own proprietary InfiniBand and Spectrum-X Ethernet, their involvement in MRC ensures that their H-Series and B-Series GPUs remain the gold standard for connectivity, regardless of the fabric choice.
- Broadcom: The “Silicon Architect.” As the leader in high-end Ethernet switching silicon (Tomahawk and Jericho series), Broadcom’s implementation of MRC into their Thor Ultra NICs is what makes the protocol a physical reality.
- AMD & Intel: The “Alternative Engines.” For the AI market to stay healthy, there must be competition. AMD’s Instinct accelerators and Intel’s Gaudi/Falcon Shores lines need a standardized, high-performance network to compete effectively against NVIDIA’s vertically integrated stacks.
Comparison: MRC vs. The Competition
| Feature | Standard RoCEv2 | InfiniBand | MRC (Multipath Reliable Connection) |
| Routing | Single Path (Flow-based) | Adaptive (Proprietary) | True Multipath (Packet Spraying) |
| Scalability | High, but complex | Moderate (Limited Radix) | Extreme (Gigascale) |
| Failure Recovery | Software-driven (Slow) | Hardware-driven | Hardware-driven (Microseconds) |
| Interoperability | Universal | Proprietary | Open Standard (OCP) |
| Tail Latency | High during congestion | Low | Ultra-Low / Deterministic |
The “Gigascale” Impact on AI Development
Why does this matter to the average person? Because the speed of AI progress is currently tied to the efficiency of the “Compute Loop.”
Currently, if an AI company wants to double their training power, they can’t just buy twice as many GPUs—they often find that the network overhead eats 30% of that new power. This is known as the “Scaling Tax.”
MRC aims to reduce the Scaling Tax to near zero. By allowing clusters to scale to 500,000 or even 1,000,000 GPUs working as a single cohesive unit, MRC enables:
- Massive Context Windows: Models that can remember and process entire libraries of books or hours of high-definition video in a single prompt.
- Faster Iteration: Reducing a 6-month training run to 2 months, allowing for faster safety testing and deployment.
- Lower Costs: Higher efficiency means less electricity wasted and lower costs for API users and consumers.
The Future: Toward an Open Fabric
The contribution of MRC to the Open Compute Project (OCP) marks a turning point. It signals that the industry has realized that while they will compete on GPU architecture and model weights, the “plumbing”—the network fabric—must be a shared, open infrastructure.
In the coming years, we expect to see “MRC-Ready” labeling on everything from network interface cards (NICs) to fiber optic transceivers. As OpenAI and Microsoft move toward clusters that resemble small cities in terms of power consumption and physical footprint, MRC will be the nervous system that keeps those millions of artificial neurons firing in perfect sync.
The networking wall has been breached. With MRC, the path to AGI is no longer limited by the speed of the connection, but only by the scale of our ambition.