Missed San Jose this year? You had plenty of company. The 2026 OCP Global Summit sold out months ahead, closed registration with no onsite option, and still left about 3,000 people on a waitlist.
The four days at the San Jose McEnery Convention Center (October 12 to 15) will shape how you buy servers, racks, power gear, and cooling over the next few years. Chipmakers launched racks built for AI agents. Microsoft revealed its third-generation inference chip. Power delivery moved toward 800 volts. And the Open Compute Project (OCP) published a 104-page paper warning that the industry keeps getting caught short on cooling.
This recap covers what was announced, contributed, and published at OCP Global Summit 2026, grouped by infrastructure layer. Each section ends with what it means for your planning.
Also Read: AI-Powered vs. Traditional SIEM: What Should Enterprises Choose in 2026?
OCP Global Summit 2026 at a Glance
OCP Global Summit 2026 ran under the theme “Scaling Innovation for the AI Era” and covered the full stack, from chip to grid. It was also the last summit in San Jose: from 2027, the event moves to the Moscone Center in San Francisco.
| Detail | What you need to know |
| Dates | October 12 to 15, 2026 (expanded to four days) |
| Venue | San Jose McEnery Convention Center, California |
| Program | 22 technical tracks, keynotes, Innovation Village, Future Technologies Symposium |
| Attendance | Roughly 15,000, plus a 3,000-person waitlist |
| Keynote companies | AMD, Arm, NVIDIA, Microsoft, Broadcom, Google, Meta, Intel, Oracle, HPE, ABB, Eaton, AMI/Lattice |
| Recordings | OCP posts session video and slides on its website after the show |
Three threads ran through almost every keynote:
- Agentic AI brings the CPU back. AI agents run constantly, call tools and juggle context, and that work lands on CPUs.
- Power and cooling set the pace. Four power specifications arrived in a single week, and the cooling discussion turned blunt.
- Readiness beats reaction. Specs are landing before the hardware that depends on them ships at volume.
OCP Global Summit 2026 expo hall at the San Jose McEnery Convention Center
Also Read: Best 5 Agentic Container Security Platforms in 2026
Key Announcements by Date
| Date | What landed |
| Oct 15 | OCP end-user value study reports up to 150x ROI for hyperscale members |
| Oct 14 | KIOXIA announces CM9-DLC liquid-cooled enterprise SSDs |
| Oct 13 | IDC forecast for OCP-recognized spending raised to ~$800B in 2029; AMI/Lattice keynote on rack management; Future Technologies Symposium |
| Oct 12 | Keynotes from AMD (Helion), Arm, NVIDIA (Vera rack, MGX 1.2) and Microsoft (Maia 300); OCP Nexus launch; Thermal Boundaries paper published |
| Oct 8 | Open Switch Platform Architecture (OSPA) draft spec announced ahead of the summit |
OCP Foundation Announcements
The OCP Foundation used the summit to put hard numbers behind open hardware and to release new tools and specs.
OCP Nexus: An AI Assistant for 15 Years of Specs
OCP launched OCP Nexus, an AI knowledge assistant you query in plain language. It searches specifications, reference designs, white papers, and past summit content, and it doubles as the event guide. If your team hunts through OCP documents during procurement, it’s worth a test drive.
IDC Forecast: OCP Spending Hits ~$800 Billion by 2029
The 2026 IDC Market Impact Study, commissioned by OCP, raised its 2029 forecast for OCP-recognized IT equipment by about 2.7x.
- 2029: about $800 billion, up from last year’s $295 billion forecast
- 2030: $892 billion, including about $814 billion in servers
- Growth: from about $154 billion in 2025, roughly 42% CAGR
- Share: about 51% of spending across the measured infrastructure categories by 2030
- AI-related: about $628 billion, or 70% of the 2030 total
End-User Value Study: Up to 150x ROI
A second OCP study found that hyperscalers investing about $10 million a year in OCP report more than $1.5 billion in value. Members also reported up to 15% faster deployment, up to 30% shorter qualification timelines, and twice as many qualified vendors.
Also Read: Best 6 Tools for Building an AI-Powered Software Factory on Your Existing Stack
OSPA: An Open Switch Platform Spec
Meta, Microsoft and Alibaba started the Open Switch Platform Architecture (OSPA), with input from Accton, Arista, Celestica, Cisco, NVIDIA, Wistron and others. SAI and SONiC already made switch software portable, but every switch SKU still needs custom platform code for thermals, inventory and firmware. OSPA standardizes seven areas: power management, hardware access, inventory, environmental control, firmware recovery, debuggability and security. The draft (Revision 1.0.0 candidate) is on GitHub.
What it means for you: if you buy OCP-recognized gear, you’ll have more vendors and shorter qualification cycles to work with. If you run SONiC or FBOSS, OSPA could cut onboarding time for new switch hardware.
Compute: CPU Racks Built for AI Agents
The biggest shift at OCP Global Summit 2026 came from processors that sit beside the accelerator. An AI agent works in a loop: it reasons, calls a tool, checks the result, and feeds it back into the next step. Scheduling those tool calls, sandboxes and orchestration tasks is CPU work, and vendors arrived with dedicated CPU racks for it.
Arm’s Mohamed Awad framed the market in three CPU roles: the head node feeding the accelerator, the cloud-native CPU serving users, and the agentic CPU acting as a “virtual user” that never sleeps. Arm also published a white paper on CPU performance for agentic AI, developed with Fujitsu, Meta, and OpenAI.
NVIDIA Vera CPU Rack and MGX 1.2
NVIDIA’s Vera CPU rack, an MGX design with 256 liquid-cooled CPUs, is now shipping. NVIDIA says a rack offers over 20,000 cores and more than 2x the agentic performance of the previous Grace generation (SDxCentral).
NVIDIA also contributed its third-generation MGX 1.2 rack architecture to OCP, including Vera Rubin rack and compute-tray designs. MGX 1.2 adds:
- Blind-mate liquid cooling at rack scale
- A liquid-cooled 5,000 A busbar
- Native 19-inch equipment support and fanless compute trays
- Support for racks up to 250 kW
NVIDIA’s Ian Buck said Vera Rubin compute-tray assembly dropped from 90 minutes to one minute.
AMD Helion
AMD CEO Lisa Su introduced Helion, an OCP-based rack reference design for agentic AI. It uses an ORv3 rack, follows DC-MHS, and comes in liquid- or air-cooled versions.
- CPU-only configuration: up to 32,768 cores from 6th Gen EPYC “Venice” processors, aimed at hyperscalers running agents separately from inference
- Enterprise configuration: EPYC CPU nodes, Instinct MI350P GPU nodes and a token router in one rack
- Partners: Cisco, Dell, Hon Hai, HPE, Lenovo, QCT, Supermicro and Wiwynn
- Availability: November 2026, with the spec to be contributed to OCP
Su also showed the 6th Gen EPYC “Verano,” AMD’s first EPYC with an LPDDR memory interface, due in 2027.
Intel GPU D270
Intel detailed the GPU D270 (codenamed Crescent Island), a 350 W air-cooled PCIe 5.0 inference card. It skips HBM in favor of LPDDR5X, with up to 480 GB per card in partner designs. Intel is sampling it, with no launch date announced.
| Vendor | Product | What it is | Availability |
| NVIDIA | Vera CPU rack | 256-CPU MGX rack for agent workloads | Shipping now |
| AMD | Helion | OCP rack, CPU-only or CPU + GPU | November 2026 |
| Intel | GPU D270 | Air-cooled inference card, LPDDR5X | Sampling |
| Arm | Agentic CPU white paper | Performance framework with Fujitsu, Meta, OpenAI | Published |
What it means for you: if you’re sizing infrastructure for AI agents, budget for CPU capacity alongside GPUs. Agent execution and model inference can now run on separate racks, which gives you more room to match spend to workload.
Accelerators: Microsoft Maia 300
Microsoft unveiled Maia 300, its third-generation in-house inference accelerator, and confirmed that Maia 200 is now in production powering Copilot. Microsoft plans to deploy the Maia family at multi-gigawatt scale starting in 2027 and make it the main inference hardware for its own services.
Maia 300 targets long-context, memory-heavy, and decode-heavy workloads, including multi-step agentic tasks. It uses separate chiplets for compute, memory controllers, and I/O.
| Spec | Maia 300 (vendor-reported or press-reported) |
| Memory | 288 GB HBM4 |
| Memory bandwidth | ~18 TB/s |
| FP8 / FP4 compute | 6.9 / 13.8 PFLOPS |
| TDP | 950 W |
| Rack throughput | Up to 7x Maia 200 in a 96-accelerator rack |
| Efficiency claim | 40% more tokens per megawatt than leading GPU systems |
| Production | 2027 |
Treat these figures as vendor claims until independent benchmarks appear. No outside party has published test results yet.
What it means for you: Maia 300 won’t be something you buy, but it shapes Azure’s inference pricing and capacity from 2027. It also signals that hyperscalers are designing memory bandwidth first for agent and long-context serving.
Networking: Scale-Up Ethernet, Optics and Open Switches
Ethernet consolidated its hold on AI scale-up networking, and optics pushed further into the rack.
ESUN and the Life of Copper
Broadcom’s Asad Khamisy said five or more hyperscalers now run Ethernet scale-up networks in production. The OCP Ethernet for Scale-Up Networking (ESUN) effort has grown to more than 175 organizations since it launched at the 2025 summit.
On cabling, Khamisy said new modulation can double in-rack bandwidth while keeping copper’s 2-meter reach. Beyond about 1,000 XPUs, operators are moving to hybrid designs: copper inside the rack and optics between racks.
Optical Specs and Form Factors
- OCI MSA: Microsoft’s Reynold D’Sa pointed to the Optical Compute Interconnect MSA, the first optical spec for multi-rack scale-up systems.
- USFP: NVIDIA and Cisco described a pluggable optics form factor with a path to 400 Gb/s per lane.
- Silicon photonics: Lightmatter and Dell presented an Open Silicon Photonics for AI Systems workstream inside OCP’s Modular Hardware System (MHS) project.
- Optical circuit switching: Voyant Photonics revealed plans for a chip-scale optical circuit switching engine for AI data center networks.
Switches and Fabrics
- Meta is adopting NVIDIA Spectrum-6 102.4 Tb/s Ethernet switches.
- Marvell showcased its Teralynx T100 102.4T switch and second-generation Structera CXL memory expanders.
- Nokia demonstrated Multipath Reliable Connection (MRC) and integrated IP-optical designs for distributed AI fabrics.
- OSPA (covered above) targets the switch platform layer underneath all of this.
What it means for you: if you’re building AI clusters on Ethernet, ESUN is the spec to track. Plan cable plants for copper inside racks and optics across them once you pass roughly a thousand accelerators.
Power: 800 VDC and Grid Resilience
Power delivery produced four new specifications in one week, according to TechArena’s recap. As racks climb from about 100 kW toward 1 MW, 54 VDC distribution can’t carry the current without massive copper. The answer the community is converging on is 800 VDC.
| Announcement | Who | What it does |
| First 800 VDC AI factory (K-1) | Visionbay, with NVIDIA and Vertiv | Built on the in-development OCP 800 VDC standard; cuts copper use by about 45% and scales racks toward 1 MW |
| Mt. Diablo 800 V architecture | Microsoft | Moves power conversion out of the compute rack; contributed to OCP at versions 0.7 and 1.0 |
| Grid interface specification | Microsoft, with utilities | Common interface between AI data centers and the grid, released the month before the summit |
| OCP Grid Resilience Specification | Google, Meta, Microsoft, NVIDIA | Targets restoring near-full load within two seconds of a grid disturbance |
Meta also described its Prometheus cluster in Ohio, which spans more than 30 structures. Meta said prefabricated fabric buildings can bring capacity online in three to five months.
What it means for you: if you’re planning AI capacity for 2027 and beyond, ask your colocation provider and power vendors about their 800 VDC roadmap now. Grid resilience is also becoming a design requirement: utilities want large AI loads that ride through disturbances instead of dropping off.
Cooling: The Thermal Boundaries Paper
The most quoted document of the week was “Thermal Boundaries in AI Infrastructure,” a 104-page paper OCP published on October 12. Authors from Promersion, Oracle, Meta and Avayla argue that the industry keeps adopting new cooling methods late and paying for it.
Three Ideas Worth Knowing
- Least-disruption law: the industry adopts a new cooling method only when the current one can no longer do the job, because each new method brings dependencies people prefer to defer.
- Preparation debt: when air cooling hit its limit on the hottest accelerators, few downstream vendors were ready, and supply of cold plates, manifolds and coolant distribution units (CDUs) fell short.
- Cold plate has a ceiling too: the paper projects single-phase cold plate’s physics limit near 4,000 W for a reticle-class die, while frontier package power doubles roughly every two years.
The authors don’t name a single successor. They recommend keeping a toolbox ready: air, cold plate, immersion, two-phase and microfluidics, interoperable and matched to the heat load and the site. They also flag commissioning, the step between building a multi-technology thermal system and trusting it with production load, as a gap no single vendor owns.
Cooling Hardware on Show
- Microsoft’s third-generation Mt. Atlas heat exchanger nearly doubles cooling capacity in the same footprint. Microsoft now runs several independent water loops at different temperatures to serve liquid- and air-cooled racks side by side.
- NVIDIA’s MGX 1.2 contribution standardizes flow rates and quick disconnects so OCP members can build compatible parts.
- Vertiv presented an update on OCP’s modular technology cooling workstream and joined a coolant panel with Dow, AMD, ChemTreat and Microsoft.
Liquid-cooled AI server rack showing coolant manifolds discussed at OCP Global Summit 2026
What it means for you: don’t plan a facility around one cooling method. Specify water loops, floor loading and service access that can support cold plate today and two-phase or immersion later, and qualify more than one supplier for CDUs and manifolds.
Memory, Storage and Rack Management
Agents carry context from step to step, and that context has to live somewhere. Memory and storage vendors used the summit to show where it should sit.
Agentic AI Changes the Memory Math
At MiTAC’s Executive Tech Forum, moderated by TechArena’s Allyson Klein, four vendors laid out the problem layer by layer:
- SK hynix: agents overtook humans as the main source of token traffic early in 2026, and more than 85% of agent tokens come from cached prompts. SK hynix described new memory tiers, including high-bandwidth flash (HBF) and CXL-pooled memory in the multi-terabyte range.
- Solidigm: an 18-token user prompt grew to 42,018 tokens once system prompts, retrieval, tool schemas and history were added, needing about 13.1 GB of KV cache. Storing KV cache on SSD cut time to first token from 31 seconds to 2.8 seconds in one test with Metrum AI.
- Intel: positioned the CPU as the orchestrator for agent workflows.
- MiTAC: cited capacity for 5,000 racks per month and power readiness from 480 VAC up to 800 VDC.
Storage Hardware
KIOXIA added direct liquid-cooled (DLC) E3.S models to its CM9 enterprise SSD line. The CM9-DLC supports the OCP Datacenter NVMe SSD v2.5 spec, targets context caching, and ships in 1 DWPD and 3 DWPD endurance options. It’s sampling to select customers.
Rack Management and the Control Plane
Lattice CEO Ford Tamer and Sanjoy Maity of AMI used their keynote to push for a common management and control model for AI racks. NVIDIA extended its DSX OS with four open-source components, including a Rack Management Service. With OSPA covering switches, open management is moving from single servers to whole racks.
Future Technologies Symposium Winners
- Best AI Paper: “How to Increase Tokens per Watt by 30 percent”
- Best Sustainability Paper: Inlyte Energy and LoadCrest, on iron-sodium batteries for emissions-free backup and load-leveling at AI data centers
What it means for you: plan agent infrastructure around session context, concurrency and where KV cache lives. Model size and GPU count alone will underestimate your memory and storage bill.
What OCP Global Summit 2026 Means for Your Roadmap
You don’t need to run a hyperscale fleet to act on what came out of San Jose. OCP designs now reach well beyond OCP’s own membership, and IDC expects them to account for about half of infrastructure spending in its categories by 2030.
- Size agent workloads separately. Model agent execution (CPU, memory, KV cache) apart from model inference (accelerators). Helion and the Vera rack both assume this split.
- Ask vendors about OCP alignment. Request ORv3, DC-MHS and MGX compatibility in RFPs to widen your supplier pool and shorten qualification.
- Put 800 VDC on your facility roadmap. Even if your next build stays at lower densities, confirm your colo or design partner has an 800 VDC and grid resilience plan.
- Design cooling as a toolbox. Leave room for cold plate, two-phase and immersion, and budget time for commissioning mixed thermal systems.
- Watch the open specs. ESUN, the OCI MSA, OSPA and the Grid Resilience Specification will shape what vendors sell you in 2027.
| If you are a… | Start with |
| CIO / CTO | IDC forecast, 150x ROI study, 800 VDC roadmap |
| Cloud or infrastructure architect | Helion, Vera rack, MGX 1.2, Maia 300 |
| Network engineer | ESUN, OCI MSA, OSPA, USFP |
| Data center / facilities lead | Thermal Boundaries paper, Mt. Diablo, Grid Resilience Specification |
| IT buyer | OCP-recognized product options, KIOXIA CM9-DLC, partner Helion systems |
Conclusion
OCP Global Summit 2026 showed an industry preparing for the next wall before it hits. AI agents pulled CPUs back into the AI rack, power moved toward 800 VDC with grid-level resilience, and OCP’s own thermal paper told everyone to get ready for every cooling method at once.
Most of these specs arrived before the hardware that depends on them ships at volume. That timing gives you a window: the teams that plan for agent-ready compute, 800 VDC power and mixed cooling now will deploy faster when the 2027 parts arrive. The summit moves to San Francisco next year, and session recordings from San Jose are your best way to go deeper in the meantime.
