Walk the exhibition floor and one thing is immediately clear: this year's OCP had a single center of gravity, and it was Vera Rubin.

From OEM heavyweights like Dell and Supermicro, to Taiwanese supply-chain players such as Delta, Wiwynn and Pegatron, to liquid cooling and power vendors including Envicool, CoolIT, nVent and Vertiv, nearly every booth carried the words "NVIDIA Vera Rubin," "NVL72" or "MGX Ecosystem Partner."

On opening day, October 12, NVIDIA dropped the summit's biggest piece of news: through its open MGX architecture, it is formally contributing the Vera Rubin rack and compute tray designs to OCP. Any system vendor can now use those public blueprints to build AI factory infrastructure compatible with Vera Rubin.

What NVIDIA Opened Up

Speaking to a packed hall, Ian Buck, NVIDIA's vice president of hyperscale and high performance computing, devoted his keynote to a single theme: how open infrastructure becomes the key to scaling AI factories.

The vehicle this time is the third-generation MGX rack reference design, the MGX 1.2 specification, now released in final form. A few numbers are worth flagging:

  • Support for single-rack deployments of up to 250kW
  • Rack-level blind-mate liquid cooling, with no cables, hoses or fans inside the compute tray. Fully liquid-cooled, fully modular, serviceable by plugging and unplugging
  • A liquid-cooled 5,000A busbar aimed at higher-power AI systems
  • Native support for 19-inch equipment, with standardized power and cooling interfaces

"Everything we're contributing comes from multiple generations of NVIDIA experience across the entire stack," Joe DeLaere, senior manager for AI factories at NVIDIA, said on an analyst call, adding that the contributions will help the whole ecosystem build AI factories faster.

On the show floor, two full racks stood side by side at NVIDIA's booth: Vera Rubin NVL72 and the Vera CPU Rack. It was the most crowded photo spot in the hall. The former is the workhorse for training and large-scale inference; the latter targets the tool calling, orchestration and sandboxed execution of the agentic AI era. The Vera CPU rack is already shipping, and ASRock Rack's dual-Vera 2U system has appeared in ODM product catalogs.

The big screen at NVIDIA's own booth pushed a slogan about accelerating Time to First Token, laid out across four stages: strategic planning, agile deployment (claimed to cut time to live by 60%), solution optimization and intelligent operations. That is essentially the narrative spine of DSX AI Factory, which comes up later.

The Scope Widens: From Server Racks to the Power Grid

If OCP in past years was about servers and rack hardware standards, this year's conversation stretched across the full infrastructure stack: chips, networking, power, liquid cooling, data center buildings and the grid itself. Three moves stood out.

Open Data Center Specification v0.7

NVIDIA, AMD, Google and Meta jointly signed the release. Competitors putting their names on the same specification says plenty about how acute the industry's pain points have become. The draft, dated August 27, 2026, covers data center power delivery, cooling and physical layout, setting a common rack height (100 inches), maximum weight (8,000 pounds) and boundary conditions for power and cooling. In NVIDIA's framing, it gives system vendors, colocation providers and data center operators a set of "predefined templates" so they stop reworking custom designs for every customer.

OCP Grid Resilience Specification

Published jointly by Google, Meta, Microsoft and NVIDIA, it requires large AI facilities to stay connected during grid disturbances, respond automatically, and return to near-full load in under two seconds. Once an AI factory's load swings are large enough to affect grid stability, the data center stops being merely a consumer of electricity and becomes a participant in the grid ecosystem.

800VDC Power Steps Into the Spotlight

A company called Visionbay showed off K-1, an 800VDC AI factory design built for the Vera Rubin era. High-voltage DC power was everywhere on the floor this year, with Delta, Vertiv, Eaton and other power vendors all telling the same story.

One booth was worth lingering at: a microgrid scale model tying together a substation, renewables, a small modular reactor (SMR), energy storage, diesel and gas generator sets, solid oxide fuel cells and the data center building itself, with red for AC, blue for DC and green for communications. "From chip to grid" is no longer a slogan; it is what exhibitors are actually selling.

Ecosystem Roll Call: Everyone Lining Up Around Vera Rubin

One lap around the hall amounts to a walk through the entire AI infrastructure supply chain.

Dell Technologies led with "AI Factory with NVIDIA," opening up a PowerEdge rack to show the cabling, alongside PowerRack for Compute, built with AMD, and Dell Exascale Storage. The slogan "Beyond Cooling, Beyond Energy" sat in the most prominent spot.

Supermicro's blue booth was packed to the edges under the theme "Built To Accelerate: data center building blocks for scalable AI," showing the Helios rack developed with AMD and a full system built around Vera Rubin NVL72. Plenty of visitors pressed up against the glass case to photograph the NVL72 glowing green inside.

The Taiwanese supply chain was impossible to miss. Delta put up a full-rack display branded "NVIDIA Vera Rubin Platform," covering the power chain from 800V DC down to the chip. Wiwynn's screens looped a rapid teardown of AMD's Helios rack-scale system. Pegatron took its own co-branded booth with NVIDIA, showing three product lines: the liquid-cooled Vera Rubin platform, Vera CPU and Groq 3 LPX. HPE, GIGABYTE and xFusion also brought full systems built around the latest NVIDIA and AMD platforms.

On the AMD side, the Helios full rack was the clear star, with a crowd permanently gathered under its vertical light sign. Worth remembering: AMD is also a co-signer of Open Data Center Specification v0.7. On open standards, AMD and NVIDIA have chosen the same side.

Google Cloud displayed a rack codenamed Mt Diablo. On open rack design, Google has long been one of OCP's most active champions.

On the liquid cooling and power side, Envicool's "CoolInside full-chain liquid cooling" booth carried a sign with real weight: OCP Project Deschutes Open CDU, co-signed by Envicool and Google. On open-source CDUs, Chinese vendors have pushed into the core circle of standard setting. Vertiv showed a cutaway data center model with red and blue piping making the liquid cooling path plain to see, with "NVIDIA 2026 MGX Ecosystem Partner" and "OCP Platinum" plaques sitting side by side on the base. CoolIT, nVent and Eaton booths drew similar crowds.

Software Is Opening Up Too

Beyond hardware, NVIDIA released four open-source software components for DSX AI Factory: switch configuration (Config Manager), GPU cluster stress testing (Cluster Readiness Engine), rack firmware management (Rack Management Service) and bare-metal provisioning (Cluster API Provider).

The components come out of NVIDIA's operational experience running DGX Cloud, and the goal is blunt: let cloud providers turn a pile of hardware into a production-ready cluster within hours. At the core, DSX OS acts as the operating system layer, co-designing compute, networking, power, cooling, facilities and software as a single system. What NVIDIA wants is to turn the entire AI factory into a repeatable, copy-and-paste standardized product.

The networking announcements came just as thick. Meta said it will use NVIDIA Spectrum-6, a 102.4Tb/s Ethernet switch chip, in its Vera Rubin AI factories, under the internal project name N-IceCube. Google's Falcon Transport will land on future ConnectX-10 SuperNICs and BlueField DPUs. NVIDIA and Cisco jointly proposed the USFP optical module standard, the industry's first pluggable module running 400Gb/s per lane, with roughly four times the bandwidth density of existing 200G per-lane modules.

A $892 Billion Prize, and One Sober Caveat

On the summit's second day, October 13, the OCP Foundation published its annual market impact study, commissioned from IDC. The numbers are striking:

  • Global spending on OCP standards-related infrastructure is projected to grow from about $154 billion in 2025 to $892 billion in 2030, a compound annual growth rate of roughly 42%
  • The 2029 forecast was revised sharply upward from last year's $295 billion to $800 billion, a 2.7x increase in a single year and a rare move even by the standards of IDC's forecasting history
  • By 2030, OCP-related equipment is expected to account for about 51% of total spending across the infrastructure categories tracked, with AI-related spending at roughly $628 billion, or about 70% of the total. Servers remain the largest category at about $814 billion

To be clear, this is a market outlook commissioned by OCP, measuring spending on equipment that "conforms to OCP specifications" rather than actual orders or shipments. The direction of travel, though, is not in doubt: open standards are turning from an idealistic position into a highly dependable business.

What NVIDIA opened up here is the physical design of the rack and tray: power, cooling, mechanical structure. The Rubin GPU itself, NVLink interconnect and the CUDA software stack all remain firmly in NVIDIA's hands. The body is open; the soul stays closed. The elegance of the play is this: by turning the surrounding infrastructure into an industry standard, NVIDIA gets the entire supply chain, OEMs, ODMs, rack makers, power vendors, liquid cooling suppliers, to stock up and invest ahead of time around its own dimensions and technology roadmap. The more the ecosystem flourishes, the deeper the moat.

For Chinese players in the supply chain, that is both an opportunity and a constraint. Racks, power supplies, liquid cooling, busbars, structural parts: Rubin-compatible solutions will sustain a long list of companies, and Envicool's name appearing on an OCP specification is the clearest proof. But everyone also knows where the ceiling sits.

Walking out of the hall, the San Jose sunset was just right. In front of the green OCP sphere by the entrance, people were still queuing to take photos. Nobody knows what next year's AI factory will look like. What is certain is that its blueprints are being turned, page by page, into public standards right here.