OCP Global Summit 2025 — Executive Sessions: Conference Intelligence

Conference: Open Compute Project (OCP) Global Summit 2025 — Executive Sessions track Location / dates: San Jose, CA · October 13–17, 2025 (largest OCP Summit to date — 11,000 attendees) Source set: 43 executive-session presentations · 94 organizations · 76 named individuals · 149 quantified metrics · 28 forward predictions

1. Executive Briefing — what an executive must know

  1. Power, not compute, is now the binding constraint on AI scale. GPU per-chip TDP is on an exponential ramp — NVIDIA H100 ~700 W → Blackwell >1,000 W → projected Feynman ~2,500 W+; AMD MI450 approaching ~2,000 W. Rack power is moving from a 2025 norm of ~130 kW toward a projected 1 MW per rack by 2027, while 75% of installed racks are still under 20 kW. Every infrastructure subsystem at the show — power delivery, cooling, networking topology — is being rearchitected around this single fact. Cisco stated it plainly: training frontier models like GPT-5 already exceeds a single data center's power envelope.

  2. 800 VDC and mandatory liquid cooling are the new baseline below the rack. A coordinated industry transition from 48–50 V to 800 VDC (Delta, LITEON, Flex, TE Connectivity, co-developed with NVIDIA/Meta/Microsoft/Google) is underway for rack loads above ~250 kW. Liquid cooling has crossed from option to requirement above ~120 kW, with a deep vendor bench (Envicool, Auras, Lenovo Neptune, MiTAC, Iceotope, Solidigm, Perstorp, TCT) and CDUs now scaling to 2.4–2.5 MW.

  3. Networking is fragmenting into three tiers, and the industry is consolidating on open Ethernet for two of them. Scale-up (UALink / OCP's new ESUN), scale-out (Ultra Ethernet — spec 1.0 shipped June 2025, 126+ members), and a newly-named scale-across tier (Cisco) for multi-datacenter training. Broadcom's Tomahawk 6 (102.4 Tbps, sampling June 2025) enables 128K-GPU two-tier fabrics at >40% power/optics savings vs. three-tier.

  4. Co-Packaged Optics crossed from "if" to "when." Meta's landmark field data — millions of hours at 400G, 5× performance vs. 3× optics, ~65% operational power savings — reframed CPO as inevitable. NVIDIA's Jensen Huang argues CPO can collapse training networks from three tiers to two. Consensus mass-deployment window: ~2028 (LightCounting forecasts >$5B CPO sales by 2030).

  5. Open standards and open firmware are the real competitive moat — and OCP is the venue. Google logged 70+ OCP contributions; AMI maintains an unfragmented OpenBMC LTS; new work spans openSFI, OCP SAFE, Caliptra/OCP Lock security, DC-MHS/PnP, and chiplet co-design (Rebellions + Samsung Foundry + Arm + ADTechnology on UCIe). Against the backdrop of the MIT finding that 95% of enterprise GenAI pilots fail, the winners are framing openness and interoperability — not raw silicon — as the differentiator.

Sign in to read the full article.

Sign In