2025 OCP Global Summit — Storage: Conference Intelligence

The Storage track at the 2025 OCP Global Summit (San Jose, CA, October 13–17) was dominated by a single underlying force: AI and GPU workloads are reshaping storage economics, power budgets, and access patterns. Across sixteen sessions, hyperscalers (Meta, Google, Microsoft), SSD and silicon vendors (Micron, Samsung, Kioxia, Solidigm, SanDisk, ScaleFlux, Pure Storage, Seagate, ARM, NVIDIA), and the standards/test ecosystem converged on a shared agenda: deploy QLC at scale, cut write amplification, harden drives against post-quantum threats, standardize multi-tenant performance isolation, and rethink the SSD entirely for GPU-initiated, IOPS-bound workloads. The OCP Datacenter NVMe SSD specification — now 220 pages and 475+ requirements, heading to spec 2.7 — is the connective tissue binding all of it.

1. Executive Briefing

QLC flash crossed into hyperscale production. Meta unveiled its first QLC storage server — a 2U Open Rack v3 chassis with 24 U.2-long slots accepting both standard NVMe SSDs and Pure Storage Direct Flash Modules — launching with 75TB DFMs, then 128TB SSDs and 150TB DFMs, delivering over 50 petabytes per rack at a 32 MB/s-per-TB bandwidth target. Built with Jabil as OEM in eight months and contributed to OCP as an Inspired Specification, QLC is now positioned as the power- and cost-efficient middle tier between HDD and TLC for read-heavy AI workloads.
Flexible Data Placement matured into a production write-amplification weapon. FDP is ratified in NVMe, in the OCP spec since 2.5/2.6, and shipping for 2–3 years with full Linux support. Vendors showed 70–80% WAF improvement through configuration alone: Kioxia's open-source U-Frog RocksDB plugin pushed drive WAF to ~1.0 (lifting throughput from ~8,000 to ~15,000 QPS); a no-code Linux-build experiment cut WAF from ~2.2 to ~1.0; and Solidigm's multiple reclaim group FDP cut a mixed hyperscale workload from 3.06 to 1.07 on a prototype — effectively doubling SSD service life.
Security became the dominant spec driver, with a hard 2027 deadline. Post-quantum cryptography (CNSA 2.0) is preferred in 2025 and mandated January 1, 2027, forcing dual-signed firmware, SPDM 1.4 multi-bank, and higher-bandwidth I3C sideband for PQC signatures 50x–500x larger than ECDSA. In parallel, Google and Microsoft advanced OCP L.O.C.K. 1.0 with Caliptra 2.1 and the finalized TCG Opal MEK-MPA multi-party authorization feature set — an open, cryptographically verifiable key-management base layer demoed live at the summit.
GPU-initiated storage opened a new IOPS-per-TCO SSD class. NVIDIA's SCADA (Scaled Accelerated Data Access) interface lets a GPU directly initiate fine-grained, massively concurrent NVMe access independent of the CPU. An H100 reached 96 million IOPS at only ~24% utilization (versus 43–45M for a same-generation CPU), and 48 Micron 9650 Gen6 drives scaled to nearly 250 million IOPS. An end-to-end GNN training workload run entirely from NVMe was only 20% slower than all-CPU-memory. The OCP Storage Next effort is gathering stakeholders to standardize these drives.
Telemetry, latency monitoring, and open test frameworks standardized fleet debuggability. Meta deployed OCP NVMe Latency Monitoring on thousands of production SSDs, publishing the first fleet-wide stall data (one unnamed vendor caused ~90% of >1-second stalls), and open-sourced its full SSD qualification suite via OCP Diag AutoVal — cutting firmware regression from 8 to 6 weeks. Micron showed how the open nvme-cli OCP plugin produces human-readable debug logs customers willingly share. The OCP NVMe SSD spec 2.7 was heading to the steering committee in November 2025.