InfiniBand vs Ethernet — What Matters This Week (May 2026)

1. Top 3 Things Happening (User View)

1. **Cornelis Networks published a roadmap committing the 800 Gbps CN6000 (2026) to Ultra Ethernet, abandoning proprietary Omni-Path** — the last major proprietary HPC-fabric holdout joining the standards track. Why now: UEC 1.0 shipped 2025-06-11 with a ~1M-endpoint transport target (Products module), which is the specific capability shift that made a proprietary 500k-endpoint fabric uneconomical to keep alive.

• Why it matters to: IBETH-HPC-FABRIC-01 (HPC Fabric Architect / Cluster Lead)
• Action: Re-run your AllReduce TCO with CN6000-on-Ultra-Ethernet on the dual-source list before the next RFP bundles fabric to compute.

1. **Meta's ~1GW Prometheus Ethernet/RoCE cluster comes online in 2026, explicitly not InfiniBand; Hyperion scaling toward 5GW** (Meta Platforms company module). The single largest AI-fabric buyer is voting Ethernet at multi-gigawatt scale — a public reference that did not exist before 2025.

• Why it matters to: IBETH-HYPER-PROCURE-01 (Hyperscaler procurement lead)
• Action: Use Meta's RoCE-at-1GW as the proof point to break NVIDIA's bundled-fabric discount lock-in in this quarter's negotiation.

1. **HPE closed its $14B all-cash Juniper acquisition (2025-07-02) and reported Q3 FY25 revenue ~$9.1B, +19% YoY** (HPE company module). The Ethernet-AI franchise just got repriced upward by a real M&A clearing event, not a roadmap promise.

• Why it matters to: IBETH-STANDARDS-01 (UEC / IBTA committee participant)
• Action: Re-weight committee influence forecasts — a doubled HPE/Juniper networking arm shifts UEC voting gravity; brief your product team this week.

2. Strategic Implications (Our View)

1. The proprietary-low-latency-fabric category is collapsing into standards-track Ethernet, not coexisting with it.

• Evidence: Products module — "the contest is between *proprietary low-latency switched fabrics* and *standards-track Ethernet-AI*"; Cornelis CN6000 pivot confirms direction.
• Implication for next 30 days: Reframe the topic as "InfiniBand's moat is now a one-generation lead, not a category." Stop publishing IB-vs-Eth as a permanent binary.

1. The decisive battle is moving *underneath* the IB/Ethernet debate to optical I/O scale-up.

• Evidence: Products module — "optical I/O quietly redefining the scale-up layer underneath both"; Ayar Labs (TeraPHY+SuperNova) and Enfabrica (ACF) hiring SerDes/224G PAM4 with zero business signals yet.
• Implication for next 30 days: Add an explicit scale-up/optical layer to coverage so we are early on the next inflection, not narrating a settled scale-out fight.

1. Hiring is a cold-start blind spot for this topic and must not drive conclusions.

• Evidence: Hiring module — "treat 30 as a cold-start floor, not growth"; 30/30 roles 100% unspecified seniority.
• Implication for next 30 days: Gate all hiring-derived claims; lean on capex/M&A signals (Meta $66–72B, HPE $14B) as leading indicators instead.

3. Newsletter Issue Spotlight

• Lead story: Cornelis Networks committing the 800 Gbps CN6000 (2026) to Ultra Ethernet and abandoning proprietary Omni-Path.
• 3-second hook: The last proprietary AI fabric vendor just defected to Ethernet.
• Image direction: V16 isometric 3D — a single proprietary HPC switch rack lowering a small white flag onto a glowing Ethernet superhighway carrying a dense stream of ~1M endpoint nodes flowing past it; cool teal/blue palette, one warm amber accent on the flag.
• Subject line: Why Cornelis just gave up on proprietary fabric

4. Reverse-Hype Watch

• Overhyped right now: InfiniBand's latency moat as a *permanent category advantage*. Cornelis's pivot plus Meta's explicit 1GW non-InfiniBand RoCE build show it is now a one-generation lead, not a structural lock-in.
• Underrated right now: Optical I/O scale-up (Ayar Labs TeraPHY/SuperNova, Enfabrica ACF). No business signals yet, but Intel/AMD/GlobalFoundries backing plus 224G PAM4 hiring means this redefines the layer both IB and Ethernet sit on top of.

5. Pattern Library Updates

• Success pattern observed: A single concrete capitulation event (a proprietary vendor publicly defecting to the standard) makes a stronger, more shareable lead than abstract market-size framing. Evidence: Cornelis module — "Published roadmap committing the 800 Gbps CN6000 (2026) to Ultra Ethernet, pivoting away from prop[rietary]."
• Failure pattern observed: Treating hiring volume as a leading indicator on a cold-start topic produced a false read (30 roles, 100% unspecified seniority, no prior baseline). Avoidance: require a ≥2-quarter baseline before any hiring-derived claim; fall back to capex/M&A signals when absent.