OCP Global Summit 2026 — attendee guide

OCP Global Summit 2026 - San Jose McEnery Convention Center, Oct 12-15 (sold out).
Mon: breakouts 12-3:30, keynotes 4-6:05 (Lisa Su, Meta, Arm, Broadcom, NVIDIA), social 6:15.
Tue: keynotes 9-11:30 incl. ABB+Eaton "Shifting to DC power distribution" 10:50; expo opens 11:30; NVIDIA 800 VDC talk 4:10 + "Intelligent Power Smoothing at the Vera-Rubin Rack" 4:25.
Wed: Future Technologies Symposium; Vertiv "800 VDC grid-to-rack" 4:10; "DC Fault Manager" 4:45; Block Party 5-7.
Thu: NVIDIA MGX power rack 1:40; TI 800V DC/DC 2:45. Expo closes 2pm.
Guide: https://artifact.sahirm.com/a/ocp-global-summit-2026/

Today: Tue Oct 13 — your plan (9:00 → leave 1:25)

• 9:00–9:15OCP Welcome Address
  South Hall
• 9:15–9:35⭐ Microsoft: Reynold D’Sa, “Intelligence Infrastructure”
  South Hall · author of the Cobalt per-VM power-capping blog
• 9:35–10:50Oracle · WOCP · Lattice/AMI control plane · HPE · Intel
  South Hall · stay seated, expo is closed until 11:30
• 10:50–11:10⭐ ABB + Eaton: “Shifting to DC Power Distribution in Data Centers”
  South Hall · the must-see of the morning
• 11:10–11:30⭐ Google: Partha Ranganathan, “Intelligent Infrastructure for the Intelligence Era”
  South Hall
• 11:30–12:50Expo booths, eat on the go (no talks worth it in this window)
  Power booths, in order: Eaton, ABB, NVIDIA, then Vertiv, Delta, TI if time · grab food from the Front Yard lunch (until 1:00)
• 1:00–1:25⭐ Andy Bechtolsheim (Arista): “The Gigawatt AI Data Center Era”
  Room 210CG · walk over from the expo by 12:55
• 1:25Hit the road
  Drop the car, then the mover

What OCP is: 8 facts

1. Who. The Open Compute Project was started by Facebook (now Meta) in 2011 to open-source hyperscale data-centre hardware designs.
2. This year. 2026 Global Summit, Oct 12–15, San Jose McEnery Convention Center. Theme: “Scaling Innovation for the AI Era”.
3. Size. 22 technical tracks, executive talks, Innovation Village demos, the Future Technologies Symposium and co-located workshops.
4. Sold out. Registration sold out; no onsite registration. Your badge is the scarce thing this week.
5. Expo. Main Expo Hall = Halls 1–3; Expo Ballroom = Grand Ballroom 220A. Expo opens Tue 11:30 am, closes Thu 2 pm.
6. Power is a headline. Two dedicated power tracks (AI Open Data Center / Power Distribution, and Rack & Power): 800/1500 VDC, megawatt racks, direct-to-server DC, grounding, backup.
7. The big shift. AC → 800 VDC distribution for AI racks is the year’s power story. ABB + Eaton keynote it Tuesday; NVIDIA, Vertiv, TI and Delta all present on it.
8. Last San Jose year. 2027 moves to Moscone Center, San Francisco, Oct 4–7.

Day by day

Mon Oct 12 (today)

• 4:00–6:05 pmOpening keynotes (South Hall)
  Su (AMD), Janardhan (Meta, “Concrete to Code: AI infrastructure at gigawatt scale”), Awad (Arm), Khamisy (Broadcom), Buck (NVIDIA). Meta’s gigawatt talk is the one most about power.
• noon–3:30 pmBreakouts
  Optional; skim the app for anything in the two power tracks.
• 6:15–7:30 pmOpening Night Social
  Easiest networking of the week.

Tue Oct 13

• 9:00–11:30 amKeynotes
  Must-see: 10:50 ABB (Paolo Catapane) + Eaton (Peter De Bock): “Shifting to DC Power Distribution in Data Centers”. Also Microsoft (9:15), Google’s Partha Ranganathan (11:10).
• 11:30 am–7 pmExpo opens
  Walk the power booths first (list below).
• 12:50–1:05 pmDelta: Microgrids at Scale
  Resilient power infrastructure for AI data centres.
• 4:10–4:30 pmNVIDIA: 800 VDC AI Factories
  From architecture alignment to industry deployment.
• 4:25–4:40 pmNVIDIA: Intelligent Power Smoothing at the Vera-Rubin Rack
  The most power-quality talk of the week: smoothing synchronous GPU load swings that hit the grid as power oscillations. Note it overlaps the 800 VDC talk; prioritise this one.
• 1–5 pmCo-located workshops (DMTF, OIF, IPEC)
  Networking/optics/manageability; skip unless curious.
• 5–7 pmBooths & Bites
  Second pass on the expo.

Wed Oct 14

• 8 am–5 pmFuture Technologies Symposium
  Research track; dip in.
• 4:10–4:25 pmVertiv (Scott Armul): 800 VDC and the AI factory
  Grid-to-rack architectures as racks go 300 kW → 600 kW → MW-class.
• 4:45–5:00 pmThe DC Fault Manager
  Open fault coordination for 800 VDC data centres; protection is where PQ instrumentation fits.
• 5–7 pmBlock Party
  Networking.

Thu Oct 15

• 8 am–2 pmLast expo hours
  Follow-ups; the expo closes at 2 pm.
• 1:40–2:00 pmNVIDIA: MGX Power Rack Reference Design
• 2:45–3:05 pmTexas Instruments: 800V DC/DC design
  Topologies and trade-offs.

Slides (2 of 184 recorded sessions)

Your sessions + 800 VDC talks (2)

Session notes

Mon Oct 12

Attended Powering the AI Era: Building Sustainable and Efficient Infrastructure at Cloud Scale (presented by Microsoft)

Mon Oct 12 2:30 pm · 210AE · Brijesh Warrier (VP Software Engineering) · Nithish Mahalingam (Partner Director, Software Engineering)

Slides: Not in the app. QR codes on the slides link to their papers. Recording: 14-31-16.wav (20 min).

• Setup (Warrier): IEA sees data-centre power at ~2.5× by 2035 (up to 3×); AI servers +31% in 2026; hyperscaler capex ~$650B in 2026, so even 1% efficiency ≈ $6.5B.
• Core idea (Mahalingam): power has to become software-defined. Compute, network and storage are pooled and scheduled; power is still static, decided separately by facility, scheduler and hardware cap, which strands power.
• Three inputs: electrical topology, power telemetry, workload context (priority, flexibility). Observe → decide → act loop, with hardware capping as the safety backstop.
• Power steering (power to the work): budgets flow down row → rack → server → VM while telemetry flows up; plan for real demand, not worst case. For inference, prefill (power-hungry) and decode (memory-bound) get different hardware (their Splitwise paper).
• Power shaping (work to the power): expose per-power-domain headroom to the scheduler and the inference router so jobs land where there's slack. Azure sees some rows run hot while others run cold.
• Capping: priority-aware throttling. Software capping took seconds, so it moved into the Cobalt 200 CPU: up to 20% more oversubscription with no hit to critical work. Loops run at ms (capping), seconds (routing), slower (placement) and months (planning).
• Ask to OCP: open telemetry (standard power + capping signals), open topology, open controls, via the Hardware Management and Rack & Power groups. Also mentioned: capacitor-based backup units in the rack (separate OCP talks).
• Powerside angle: the whole stack runs on power telemetry at row/rack/server level, and they're asking OCP to standardise it. That's a direct opening for metering.

• Azure Cobalt per-VM power management blog

Tue Oct 13

Wed Oct 14

Thu Oct 15

Power booths to walk

Five questions to ask at every power booth

1. Where are you measuring power quality today in an 800 VDC rack: at the facility AC side, the rectifier, or on the DC bus?
2. How big are the synchronous GPU load steps you see at the rack and at the utility meter, and how fast?
3. What does the utility / grid operator ask you for on flicker, harmonics or oscillation before connecting an AI campus?
4. Who owns power telemetry, facilities or the IT/infra team, and what do they buy?
5. What is missing from today’s meters for DC distribution?

Keynotes

Reading: POLCA — Power Oversubscription in LLM Cloud Providers

• Microsoft Azure paper (Patel, Choukse, Zhang, Goiri, Warrier, Mahalingam, Bianchini), arXiv 2308.12908, Aug 2023.
• Power, not space, is the binding constraint for dense GPU clusters: each data centre has a fixed power budget set by generators and utility contracts.
• LLM inference clusters rarely hit their provisioned peak, so there is room to oversubscribe; training clusters only have ~10% headroom because synchronous training creates large, coordinated power peaks.
• POLCA uses GPU power capping and frequency scaling with two cluster-level thresholds, throttling lower-priority work first, to stay under the budget.
• Result in simulation on production traces: ~30% more servers in the same inference cluster with minimal performance loss and few throttling events.
• Why it matters at OCP: the same synchronous swings that limit oversubscription are what NVIDIA’s Vera-Rubin power smoothing and the 800 VDC work are trying to tame. That load profile is a power-quality problem as much as a capacity one.

Logistics

• Venue: San Jose McEnery Convention Center. Keynotes in South Hall; Expo Halls 1–3; Expo Ballroom 220A.
• Expo hours: Tue 11:30–7 · Wed 9–5 · Thu 8–2.
• Evenings: Opening Night Social Mon 6:15–7:30 · Booths & Bites Tue 5–7 · Block Party Wed 5–7.
• Co-located workshops (DMTF, OIF, IPEC Tue 1–5; SONiC Wed) are included with registration.

Sources

• Converge Digest preview (agenda, keynotes, tracks)
• NVIDIA at OCP 2026 (session times)
• Vertiv at OCP 2026
• POLCA paper (arXiv 2308.12908)
• Microsoft Research publication page