LONDON, May 7, 2026 — On 6 May 2026, NVIDIA announced the general availability of Multipath Reliable Connection (MRC) on its Spectrum-X Ethernet fabric, marking the first time an RDMA transport protocol designed for multi-path load balancing has moved from proprietary production deployment to an Open Compute Project specification. The protocol, already operational inside AI factories operated by OpenAI, Microsoft and Oracle Cloud Infrastructure, enables a single RDMA connection to distribute traffic across multiple network paths simultaneously — improving throughput, balancing load, and sustaining GPU utilisation across thousands of accelerators. For enterprises evaluating the capital intensity of frontier AI training, MRC's open release raises a pointed question: does Ethernet now close the performance gap with InfiniBand, and what does that mean for procurement strategies in 2026 and beyond? This analysis, drawing on Business20Channel.tv's ongoing AI infrastructure coverage and our enterprise networking analysis archive, examines MRC's technical architecture, its competitive positioning against InfiniBand and rival Ethernet stacks, and the implications for hyperscaler and sovereign AI deployment budgets.

Executive Summary

• NVIDIA released MRC as an open specification via the Open Compute Project on 6 May 2026, after production validation on Spectrum-X Ethernet hardware at OpenAI, Microsoft and Oracle.
• MRC allows a single RDMA connection to fan out across every available network path, delivering hardware-speed failover in microseconds and dynamic congestion avoidance in real time.
• OpenAI's Sachin Katti confirmed MRC was deployed during Blackwell-generation training, avoiding "much of the typical network-related slowdowns and interruptions."
• Microsoft's Fairwater data centre and Oracle's Abilene facility — two of the largest purpose-built AI factories — rely on MRC for frontier large-language-model training at scale.
• Spectrum-X now supports multiplane network architectures with hardware-accelerated load balancing, a design pattern OpenAI pairs with MRC for resilient gigascale fabrics.

Key Developments

MRC: From Concept to Gigascale Production

Multipath Reliable Connection represents a fundamental departure from traditional single-path RDMA transports such as RoCEv2. In conventional designs, each RDMA queue pair maps to a single network route. When that route becomes congested or fails, recovery depends on software timers that can stall a training job for seconds — an eternity when 10,000-plus GPUs must remain synchronised. MRC eliminates that bottleneck by spreading each connection's traffic across all available paths, dynamically rerouting around overloaded or failed links. According to NVIDIA's official Spectrum-X blog post dated 6 May 2026, the protocol's failure-bypass mechanism can detect a path failure and redirect traffic "in just microseconds" entirely in hardware, with no software intervention required.

Production Validation at Three Hyperscalers

The credibility of any new networking protocol rests on production evidence, not lab benchmarks. NVIDIA's announcement names three tier-one deployments. OpenAI deployed MRC during its Blackwell-generation training runs — a period that, by public accounts from OpenAI's research blog, involved clusters numbering in the tens of thousands of GPUs. Microsoft's Fairwater data centre, described in the announcement as one of the largest AI factories purpose-built for frontier LLM training, integrates MRC into its Spectrum-X Ethernet fabric. Oracle's Abilene data centre, part of OCI's expanding GPU cloud, similarly uses MRC to meet the performance and efficiency thresholds demanded by large-scale model training. That three distinct operators — each with different software stacks and operational cultures — report success with MRC strengthens the case for broader adoption.

"Deploying MRC in the Blackwell generation was very successful and was made possible by a strong collaboration with NVIDIA," said Sachin Katti, head of industrial compute at OpenAI. "MRC's end-to-end approach enabled us to avoid much of the typical network-related slowdowns and interruptions and maintain the efficiency of frontier training runs at scale."

Open Specification via the Open Compute Project

NVIDIA's decision to publish MRC as an OCP specification invites scrutiny. Historically, NVIDIA's networking advantage rested on the proprietary nature of InfiniBand, where the company controls silicon, firmware and protocol stack through its Mellanox acquisition completed in April 2020 for approximately $6.9 billion. By contrast, Ethernet is an open standard governed by the IEEE 802.3 working group. Publishing MRC through OCP positions Spectrum-X as the reference platform for a protocol that any vendor could, in principle, implement — though the deep integration with NVIDIA's ConnectX network adapters and Spectrum-4 switches means a competitive moat persists at the hardware layer.

Market Context & Competitive Landscape

Spectrum-X vs. InfiniBand: An Intra-NVIDIA Tension

NVIDIA occupies the unusual position of competing with itself. Its InfiniBand product line, led by the Quantum-2 platform and the forthcoming Quantum-X800 announced at GTC 2025, remains the de facto fabric for the largest NVL72-based SuperPODs. InfiniBand's advantage lies in its mature adaptive routing, credit-based flow control, and sub-microsecond latency. Spectrum-X with MRC narrows that gap on two fronts: multi-path load balancing — which InfiniBand has long offered natively — and hardware-speed failover. The key differentiator favouring Spectrum-X is ecosystem breadth. Ethernet NICs, switches and cabling dominate enterprise procurement, with the Dell'Oro Group estimating in Q1 2026 that Ethernet accounts for roughly 80% of all data-centre switching port shipments globally.*

Broadcom and Ultra Ethernet Consortium

NVIDIA's fiercest Ethernet rival is Broadcom, whose Memory Fabric and Jericho3-AI switches underpin Google Cloud's Jupiter fabric. Broadcom is also a founding member of the Ultra Ethernet Consortium (UEC), which aims to define a vendor-neutral, AI-optimised Ethernet transport by late 2026. UEC's specification targets similar goals to MRC — multi-path transport, packet spraying and rapid recovery — but has not yet reached production deployment. The timing of NVIDIA's MRC open-specification release appears calculated: by establishing production credibility with three marquee customers before UEC finalises its own standard, NVIDIA positions MRC as the protocol incumbents already trust, making UEC adoption a harder sell to risk-averse infrastructure teams.

Competitive Comparison

Source: NVIDIA blog, 6 May 2026; Broadcom and UEC public disclosures; Dell'Oro Group market estimates Q1 2026. Items marked * are estimated or pre-production targets.

Technical Architecture: How MRC Works Inside Spectrum-X

Multi-Path Load Balancing and Congestion Avoidance

MRC's core mechanism distributes each RDMA connection's packets across every available Equal-Cost Multi-Path (ECMP) route in the fabric. Unlike conventional ECMP hashing — which pins a flow to a single path — MRC treats the entire fabric as a resource pool. According to NVIDIA's 6 May 2026 announcement, this design ensures "every GPU gets the bandwidth it needs throughout a training run" by dynamically avoiding overloaded paths in real time. Intelligent retransmission handles data loss with precision: rather than re-sending entire windows of data, MRC identifies and retransmits only the specific lost packets, reducing recovery overhead to a minimum.

Multiplane Network Architectures

A second innovation disclosed in the announcement is Spectrum-X Multiplane support, which OpenAI uses alongside MRC. In a multiplane topology, the network comprises multiple independent fabrics — each a fully connected plane — offering alternate communication paths between GPUs. NVIDIA's Spectrum-X platform provides hardware-accelerated load balancing across these planes, a capability that at the current generation of NVIDIA networking silicon appears unique among Ethernet switch vendors. For a cluster of 100,000 GPUs, multiplane design combined with MRC means that no single switch failure, no single link outage, and no single congestion event can stall the collective training job — a resilience requirement that moves from desirable to essential as frontier model parameter counts push beyond 1 trillion.

Source: NVIDIA Spectrum-X blog, 6 May 2026. Analysis by Business20Channel.tv editorial team.

Industry Implications

Hyperscale Cloud and Sovereign AI

For cloud providers such as Microsoft Azure, Oracle Cloud and emerging sovereign AI programmes in the EU, UK and Middle East, MRC's availability as an open OCP specification lowers the perceived risk of vendor lock-in. Governments investing in national AI compute — the UK's £1.3 billion AI Opportunities Action Plan published in January 2025 is one example — typically mandate open standards in procurement. MRC on Spectrum-X Ethernet satisfies that requirement in a way InfiniBand cannot, potentially steering sovereign infrastructure budgets toward Ethernet-based designs.

Financial Services and Healthcare

In financial services, where AI-driven risk modelling and inference workloads are expanding, Ethernet compatibility matters because existing data-centre fabrics are overwhelmingly Ethernet-based. Migrating to InfiniBand for AI training means maintaining two parallel networks — a cost and complexity burden that MRC could eliminate. In healthcare, the growing use of federated learning across hospital networks — governed by regulations such as the EU's AI Act (Regulation 2024/1689) — favours Ethernet topologies that integrate with existing IT governance frameworks. MRC's fine-grained telemetry, highlighted in NVIDIA's announcement, gives compliance teams the visibility they need to audit data flows across multi-site training runs.

Defence and Government

Military and intelligence AI programmes, which demand both performance and supply-chain diversity, stand to benefit from an open Ethernet-based AI fabric. The US Department of Defense's Chief Digital and AI Office has publicly favoured open architectures. MRC's OCP status aligns with that policy direction, though actual procurement timelines remain subject to the US government's lengthy certification processes.

Business20Channel.tv Analysis

The Strategic Logic of Opening MRC

NVIDIA's decision to open-source MRC through OCP is not altruism — it is competitive positioning executed with precision. By releasing MRC before the Ultra Ethernet Consortium delivers its own multi-path transport specification, NVIDIA achieves three objectives simultaneously. First, it establishes MRC as the de facto standard by virtue of production provenance: no other AI-optimised Ethernet transport can claim deployment at three of the world's largest AI factories. Second, it undercuts the UEC's central value proposition — that the industry needs a vendor-neutral alternative to NVIDIA-controlled networking — by making MRC itself vendor-neutral at the specification layer. Third, it protects NVIDIA's hardware margin: while the protocol is open, the optimised implementation runs on Spectrum-4 switches and ConnectX-7 adapters, which NVIDIA controls end to end. This mirrors a pattern familiar from NVIDIA's CUDA strategy: open enough to attract developers, closed enough to sell silicon.

What the Consensus Is Missing

Most commentary on AI networking frames the debate as InfiniBand versus Ethernet — a binary that obscures a more nuanced reality. The true competitive axis in 2026 is not protocol family but vertical integration. NVIDIA's advantage with Spectrum-X is not merely that it runs Ethernet; it is that NVIDIA controls the GPU (Blackwell), the NIC (ConnectX), the switch ASIC (Spectrum-4), the host software (NCCL), and now the transport protocol (MRC). No other vendor in the Ethernet ecosystem — not Broadcom, not Intel, not Arista — can match that full-stack integration. The risk for the industry is that this integration creates a de facto monoculture. If 90% of large-scale AI training runs on NVIDIA GPUs connected by NVIDIA NICs through NVIDIA switches using an NVIDIA-designed protocol, the "open" label on the OCP specification becomes cosmetic. Enterprise buyers should stress-test their vendor-diversity requirements against this reality before committing to multi-year infrastructure contracts.

Capital Implications

For CFOs approving nine- and ten-figure data-centre builds, MRC on Spectrum-X shifts the calculus in favour of Ethernet. InfiniBand's premium — historically 20–30% above comparable Ethernet deployments in total cost of ownership, according to Business20Channel.tv's prior infrastructure cost analyses — becomes harder to justify when Ethernet can deliver comparable multi-path resilience. That said, InfiniBand retains a latency edge in tightly coupled workloads, and NVIDIA's Quantum-X800 roadmap suggests InfiniBand is far from sunset. The likely outcome is bifurcation: InfiniBand for the very largest NVL72-based SuperPODs, Spectrum-X Ethernet for the broader market of 1,000-to-50,000-GPU clusters that most enterprises will actually build.

Why This Matters for Industry Stakeholders

Network architects evaluating 2026–2028 fabric designs should treat MRC's OCP release as a procurement inflection point. Specifically: (1) any new AI cluster above 1,000 GPUs should now include Spectrum-X Ethernet with MRC in the evaluation matrix alongside InfiniBand, something that was not credible 18 months ago; (2) operations teams should demand MRC-level telemetry from any competing Ethernet vendor — if a supplier cannot match per-path visibility and microsecond failover, the operational cost delta will erode any upfront hardware savings; (3) compliance officers in regulated sectors should note that MRC's fine-grained path-level control creates an auditable record of data movement, a feature increasingly relevant under the EU AI Act's transparency requirements for high-risk systems. The concrete risk is that organisations delaying fabric decisions until UEC's specification matures may find themselves 12–18 months behind competitors who adopt MRC-based Spectrum-X in 2026 and begin training frontier models on proven infrastructure.

Forward Outlook

Three developments will determine MRC's trajectory over the next 12 months. First, the Ultra Ethernet Consortium's response: UEC members including Broadcom, AMD, Intel and Arista Networks must decide whether to adopt, adapt, or compete with MRC's OCP specification. A fragmented response would benefit NVIDIA; a unified counter-specification could slow MRC adoption among customers wary of NVIDIA's full-stack dominance. Second, NVIDIA's own InfiniBand roadmap creates internal tension — the company must articulate a clear segmentation between Spectrum-X and Quantum-X800 to avoid cannibalising its higher-margin InfiniBand sales. Third, real-world benchmarks from independent parties will be decisive. The OpenAI, Microsoft and Oracle testimonials are powerful, but the broader enterprise market requires third-party performance data — ideally from organisations such as the MLCommons consortium — before committing billions. The open question is whether MRC's head start is durable or merely temporal. If UEC delivers a production-grade multi-path transport by Q1 2027, NVIDIA's first-mover advantage narrows. If UEC slips — as industry-wide specification efforts often do — MRC on Spectrum-X could become the TCP/IP of AI networking: technically open, practically dominated by one vendor's implementation. That outcome would reshape data-centre procurement for a decade, and the industry has roughly six months to influence which way it breaks.

Key Takeaways

• MRC, now an open OCP specification as of 6 May 2026, is the first multi-path RDMA transport proven at gigascale AI factories operated by OpenAI, Microsoft and Oracle.
• Hardware-speed failover in microseconds and dynamic congestion avoidance address the two primary failure modes that stall large-scale GPU training — link failure and path congestion.
• NVIDIA's decision to open MRC pre-empts the Ultra Ethernet Consortium's rival specification, creating a first-mover advantage that competitors must now respond to within 12 months.
• Enterprise buyers gain a credible Ethernet alternative to InfiniBand for AI clusters, but should scrutinise the degree of full-stack NVIDIA dependence that Spectrum-X entails.
• Regulated sectors — financial services, healthcare, government — benefit from MRC's per-path telemetry and auditability, aligning with EU AI Act transparency mandates.

References & Bibliography

[1] NVIDIA. (2026, May 6). NVIDIA Spectrum-X — the Open, AI-Native Ethernet Fabric — Sets the Standard for Gigascale AI, Now With MRC. https://blogs.nvidia.com/blog/spectrum-x-ethernet-mrc/

[2] OpenAI. (2026). OpenAI Official Website. https://openai.com/

[3] Microsoft. (2026). Microsoft Official Website. https://www.microsoft.com/

[4] Oracle. (2026). Oracle Cloud Infrastructure. https://www.oracle.com/cloud/

[5] Open Compute Project. (2026). OCP Specifications. https://www.opencompute.org/

[6] Ultra Ethernet Consortium. (2025). UEC Overview and Roadmap. https://ultraethernet.org/

[7] Broadcom. (2026). Broadcom Networking Solutions. https://www.broadcom.com/

[8] Dell'Oro Group. (2026, Q1). Data Centre Switching Market Report. https://www.delloro.com/

[9] NVIDIA. (2026). NVIDIA InfiniBand Networking. https://www.nvidia.com/en-gb/networking/infiniband/

[10] NVIDIA. (2026). NVIDIA Networking Solutions. https://www.nvidia.com/en-gb/networking/

[11] NVIDIA. (2026). CUDA Toolkit Developer Resources. https://developer.nvidia.com/cuda-toolkit

[12] Google Cloud. (2026). Google Cloud Platform. https://cloud.google.com/

[13] IEEE. (2026). IEEE 802.3 Ethernet Working Group. https://www.ieee.org/

[14] Arista Networks. (2026). Arista Data Centre Networking. https://www.arista.com/

[15] MLCommons. (2026). MLCommons Benchmarking Consortium. https://mlcommons.org/

[16] UK Government. (2025, January). AI Opportunities Action Plan. https://www.gov.uk/government/publications/ai-opportunities-action-plan

[17] European Commission. (2024). EU AI Act — Regulation 2024/1689. https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai

[18] US Department of Defense. (2026). Chief Digital and Artificial Intelligence Office. https://www.ai.mil/

[19] Business20Channel.tv. (2026). AI Category Coverage. https://business20channel.tv/?category=AI

[20] Oracle. (2026). Oracle Cloud Compute. https://www.oracle.com/cloud/compute/