NVIDIA has joined AMD, Google, and Meta in backing a new open specification meant to standardize how the industry builds data centers for the AI boom, according to reports from CoStar, Fierce Network, and other outlets tracking the announcement. The effort, built around version 0.7 of what’s being called the Open Data Center Specification and jointly published at the OCP Global Summit on October 12, 2026, targets the physical layout, cooling infrastructure, and power architecture that AI compute clusters depend on. A second, related specification covering how facilities respond to power-grid disturbances was announced the same day, with NVIDIA, Google, Meta, and Microsoft as co-authors.
The timing is not an accident. Data center construction tied to AI has become one of the defining capital stories of 2026, with hyperscalers racing to add gigawatts of capacity while utilities, regulators, and chipmakers all scramble to keep pace. Standardizing how these facilities are designed, cooled, and wired into the grid is now treated as infrastructure policy, not just an engineering nicety. Yahoo Finance reported that the four companies are positioning the specification as a “global blueprint” for data center design, a framing that signals how central this coordination has become to the next phase of AI infrastructure buildout.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What the Open Data Center Specification v0.7 Actually Covers
Version 0.7 of the specification addresses three structural layers of a data center: the physical facility layout, the cooling infrastructure that keeps AI accelerators from overheating, and the power architecture that delivers electricity down to individual racks of servers. These are the categories where custom, vendor-specific engineering has historically slowed construction timelines and driven up costs, since every hyperscaler and every chip vendor has tended to specify its own rack dimensions, cooling loop designs, and busbar layouts. A copy of revision 0.7.0 identifies August 27, 2026 as its effective date and lists AMD, Google, Meta, and NVIDIA as authors, and the specification is guidance for facilities planned for 2027–2028.
By aligning on a shared specification, the participating companies are effectively trying to turn data center construction into something closer to a modular, repeatable process rather than a bespoke one. That matters because the current generation of AI accelerators draws enormously more power per rack than traditional cloud servers, and cooling systems built for yesterday’s assumptions do not scale cleanly to liquid-cooled, multi-kilowatt GPU racks. A shared spec means a cooling vendor or a power-distribution supplier can build one product that works across NVIDIA, AMD, Google, and Meta facilities instead of maintaining parallel product lines for each customer.
Business Insider’s coverage framed the effort explicitly around speed: the goal is to compress the time it takes to plan, permit, and build new AI-ready facilities. That framing lines up with what hyperscalers have been saying publicly for much of 2026 — that power and construction timelines, not chip supply, have become the binding constraint on how fast new AI capacity can come online.
The Grid Resilience Specification: A Separate but Connected Effort
Alongside the data center design specification, NVIDIA, Google, Meta, and Microsoft separately published the OCP Grid Resilience Specification, a document focused specifically on how data centers behave during utility-grid disturbances. According to the specification, facilities built to this standard are expected to stay connected to the grid during disruptions, respond automatically rather than requiring manual intervention, and restore near-full load in less than two seconds.
That two-second recovery window is a meaningful technical detail. AI training clusters pulling tens of megawatts can behave almost like an industrial load that snaps on and off as GPU utilization swings, and that kind of volatility can ripple through a local grid in ways that utilities have struggled to model. A grid operator dealing with a brief voltage sag or frequency event needs data centers to either ride through the disturbance or shed and restore load predictably, rather than tripping offline in a way that compounds instability elsewhere on the grid.
The Open Compute Project’s own announcement frames this work as part of a broader industry transition toward standardizing how AI infrastructure interacts with the grid and with low-voltage direct current (LVDC) power delivery inside facilities. That’s a notable shift: for years, data center power architecture was treated as an internal engineering decision each company made independently. Now it’s being treated as something closer to a shared utility-interconnection standard, similar to how telecom equipment or electrical code requirements get standardized across vendors.
NVIDIA’s MGX Architecture and the Open Compute Project
NVIDIA’s contribution to this standardization push goes beyond co-authoring specifications. The company has also contributed its Vera Rubin rack and compute-tray designs to the Open Compute Project through its open MGX architecture, effectively open-sourcing reference designs for how its next-generation GPU racks are built. More than 200 industry partners have already adopted NVIDIA’s MGX design platform, according to the company, spanning server makers, component suppliers, and systems integrators who build around that reference architecture rather than designing from scratch.
MGX has functioned as something like NVIDIA’s answer to the fragmentation problem in AI server design. Rather than every OEM building a unique chassis for each GPU generation, MGX gives partners a modular reference design they can adapt, which shortens the path from a new GPU announcement to shipped servers. Folding Vera Rubin’s rack and tray designs into OCP extends that same logic up a level, from individual servers to the surrounding data center infrastructure itself. The CoStar industrial real estate desk has been tracking how this kind of standardization intersects with the broader wave of AI-driven industrial construction, where speed-to-delivery on new facilities has become a competitive factor for landlords and developers as much as for the tech companies themselves.
How MGX 1.2 Extends the Vera Rubin Rack Design
The updated MGX 1.2 specification expands the open architecture for NVIDIA Vera Rubin AI factories with rack-level blind-mate liquid cooling, interoperable rack-level interconnect infrastructure, and a liquid-cooled 5,000-amp busbar designed for higher-power AI systems. It also adds native support for 19-inch equipment and higher-density rack designs. These are the same power and cooling layers that the facility specification addresses, so the rack-level and facility-level work are meant to be read together.
MGX 1.2 also increases modularity through configurable input/output slots, fanless liquid-cooled compute trays, standardized power and cooling interfaces, and common infrastructure, which NVIDIA says shortens assembly, testing, and maintenance time and simplifies manufacturing and field servicing. The architecture supports rack deployments of up to 250 kW, while making maintenance easier, deployment simpler, and infrastructure reuse more practical for ecosystem partners building interoperable accelerated computing platforms.
For operators, colocation providers, and system makers, the practical step is to check any rack and tray design against both the MGX 1.2 interfaces and the facility specification before committing to a layout. NVIDIA’s DSX reference design is described as compatible with the new specifications, and the stated goal of the specification is to give builders predefined templates instead of bespoke facility designs for each customer. Teams planning liquid-cooled capacity should treat the busbar, cooling, and rack-density details as inputs to early design decisions rather than late retrofits.
Why Four Rivals Are Cooperating on Infrastructure
NVIDIA, AMD, Google, and Meta compete directly and fiercely in AI chips, cloud services, and AI models. That makes their cooperation on a shared data center specification notable, though it fits a pattern that’s common in capital-intensive infrastructure industries: companies compete on the product layer while cooperating on the plumbing layer, because nobody benefits from fragmented standards when the fragmentation itself slows everyone down.
The Open Compute Project, the umbrella organization under which much of this standardization work happens, was originally founded by Meta (then Facebook) in 2011 to open-source its own data center and server designs. Since then, it’s grown into the venue where competitors regularly collaborate on hardware specifications, from rack dimensions to power delivery to networking interconnects. The current round of specifications extends that same cooperative model to the much higher-stakes territory of AI-scale power and cooling, where the dollar amounts involved in getting it wrong are far larger than in the server-rack standardization fights of a decade ago.
For chipmakers like NVIDIA and AMD specifically, there’s also a practical supply-side incentive. If every hyperscaler builds data centers around incompatible facility specs, chip vendors have to validate their hardware against a longer list of bespoke environments. A shared specification narrows that validation matrix, which can shorten the time between a new GPU generation shipping and it actually being deployed at scale inside customer facilities.
The Power Bottleneck Behind the Push
This standardization effort doesn’t exist in a vacuum. It’s a direct response to the power and grid-capacity constraints that have become the central bottleneck in AI data center expansion through 2026. Utilities across multiple US regions have flagged concerns about interconnection queues, transformer shortages, and the sheer unpredictability of load from AI training clusters that can swing dramatically based on workload. Regulatory bodies have also stepped in: FERC’s approval of a PJM data center registry earlier this year, which tracks large data center loads while stopping short of mandating power curtailment, reflects the same underlying tension between grid stability and AI growth that the new OCP grid resilience spec is trying to address from the facility side rather than the regulatory side.
Seen that way, the data center specification and the grid resilience specification are two halves of the same problem. One standardizes how facilities are physically built so construction moves faster. The other standardizes how those facilities behave electrically so grid operators can plan around them with more confidence. Both are attempts to remove friction that has been slowing the pace at which new AI compute capacity actually comes online, even as chip supply and model development continue to accelerate.
Market Impact: What This Means for the Data Center Supply Chain
For the broader data center supply chain, standardization at this level tends to favor scale players over niche specialists. Companies that can manufacture cooling systems, power distribution units, and rack hardware to a single shared specification can sell into NVIDIA, AMD, Google, and Meta facilities with one product line instead of four. That’s a meaningful efficiency gain for component suppliers, and it’s part of why networking and optics vendors have been racing to align their own product roadmaps with OCP standards. The push also follows a similar standardization effort already underway in AI networking hardware, where Credo’s recent SRMO optics push for AI racks and Marvell’s 102.4 Tbps switch debut, both unveiled at OCP events, point to the same underlying logic: a shared standard shortens the path from announcement to shipped product and avoids duplicating R&D across customer-specific variants.
On the construction side, developers and real estate investors who build and lease data center shells have an incentive to build to the new specification as well, since doing so widens the pool of potential hyperscale tenants who can move into a facility without requiring custom retrofits. That dynamic has already started to show up in how industrial real estate firms describe their AI-focused development pipelines, and it suggests the specification’s influence will extend well beyond the four companies that authored it.
| Specification | Lead Contributors | Primary Focus | Key Technical Detail |
|---|---|---|---|
| Open Data Center Specification v0.7 | NVIDIA, AMD, Google, Meta | Facility layout, cooling, power architecture | Open spec for standardized data center physical design |
| OCP Grid Resilience Specification | NVIDIA, Google, Meta, Microsoft | Grid-disturbance response | Restore near-full load in under 2 seconds |
| NVIDIA MGX Architecture | NVIDIA + 200+ partners | Rack and compute-tray reference design | Vera Rubin rack/tray designs contributed to OCP |
Historical Context: From Facebook’s Open Compute Project to AI-Scale Infrastructure
The Open Compute Project’s roots trace back to Facebook’s 2011 decision to open-source the designs of its own custom-built servers and data centers, a move intended to cut costs by eliminating proprietary vendor markups on commodity hardware. Over the following decade, OCP grew to include contributions from Microsoft, Google, Intel, and eventually NVIDIA and AMD, expanding from server motherboard specs into networking switches, rack power shelves, and cooling systems.
What’s different about the current round of specifications is the scale of what’s at stake. A decade ago, OCP standardization was mostly about shaving costs off commodity cloud server deployments. Today’s specifications are being written for facilities that house GPU racks drawing many times the power density of the servers OCP originally standardized, feeding AI training and inference workloads that represent some of the largest capital expenditures in the history of the technology industry. The Data Center Dynamics construction channel has tracked this shift extensively, documenting how AI-driven demand has pushed data center construction into a category of infrastructure project more comparable to utility-scale power plants than to the office-park data centers of the 2010s.
Competitive Comparison: How the Big Four Approach Infrastructure Differently
Even as NVIDIA, AMD, Google, and Meta align on a shared specification, their underlying motivations and infrastructure strategies differ in ways worth separating out. NVIDIA’s interest is largely about ensuring its GPU hardware, especially upcoming Vera Rubin-generation systems, can be deployed as fast as possible across the widest possible range of customer facilities, which is also why it has pushed MGX adoption so aggressively. AMD’s participation reflects a similar supply-side logic, aiming to ensure its own accelerators aren’t locked out of facilities designed primarily around NVIDIA’s specifications. The same day the specification was published, October 12, 2026, AMD introduced AMD Helion, an open, OCP-specification-based rack blueprint for agentic-AI infrastructure that builds on its existing AMD Helios rack-scale solution.
Google and Meta, as hyperscale operators rather than chip vendors, have a different incentive: both companies run enormous owned-and-leased data center fleets and have the most to gain from a standardized supply chain that lowers construction costs and shortens build timelines across every site, regardless of which chip vendor’s hardware ends up inside. Microsoft’s involvement specifically in the grid resilience specification, rather than the broader data center design spec, suggests its near-term priority is grid interconnection and reliability rather than facility layout standardization, which tracks with Microsoft’s own public statements about power availability being a gating factor for its AI capacity plans. That urgency also shows up in the capital numbers: Microsoft, Meta, Google, and Amazon’s combined AI infrastructure capex has climbed toward $725 billion, a scale of spending that makes shaving even a few percentage points off construction timelines worth pursuing through shared specifications.
| Company | Role in Standardization Push | Primary Infrastructure Incentive |
|---|---|---|
| NVIDIA | Co-author of both specs; contributed MGX/Vera Rubin designs to OCP | Faster, wider deployment of GPU hardware across customer facilities |
| AMD | Co-author of Open Data Center Specification v0.7 | Ensuring accelerators aren’t excluded from standardized facility designs |
| Co-author of both specs | Lower construction costs and faster builds across its data center fleet | |
| Meta | Co-author of both specs; OCP’s founding member | Standardized supply chain across its AI infrastructure buildout |
| Microsoft | Co-author of Grid Resilience Specification only | Grid interconnection reliability for AI capacity expansion |
What This Means for Data Center Construction Timelines
The practical test of this standardization push will be whether it actually shortens the time between breaking ground on a new AI data center and bringing GPU racks online. Construction delays tied to custom cooling loop engineering, power architecture validation, and grid interconnection studies have all been cited across the industry as factors stretching build timelines well beyond what chip availability alone would suggest. If component suppliers can build to a single specification instead of negotiating bespoke requirements with each hyperscaler, procurement and manufacturing lead times for cooling systems and power distribution hardware should, in principle, compress.
That compression won’t be instantaneous. Specifications take time to work through supplier qualification processes, and existing facilities built to older, non-standardized designs will continue operating as-is rather than being retrofitted. The more realistic near-term effect is that new facility designs starting construction from this point forward are more likely to adopt the shared specification, gradually shifting the installed base over the coming years rather than transforming it overnight. Cooling, in particular, remains a major swing factor: Trane Technologies’ recent stock surge on the strength of AI data center cooling demand and AMD’s 32,768-core Helion rack design for agentic AI workloads, which AMD said was under development with ecosystem partners and planned for availability in November 2026, both illustrate how much engineering is still happening around, rather than strictly inside, the new specification’s boundaries.
Predictions: Where This Standardization Effort Goes Next
- More chip vendors and cloud providers will likely join the Open Data Center Specification as formal contributors over the next 12 months, following the pattern of OCP’s historical growth from a handful of founding members to a much broader coalition.
- Expect follow-on specifications addressing networking interconnects and optics standardization to be published under the same OCP umbrella, extending the same logic from facility design down to the switch and cable layer.
- Utilities and grid operators in major data center markets are likely to reference the OCP Grid Resilience Specification’s two-second restoration benchmark in future interconnection agreements, turning what started as an industry-authored technical document into a de facto regulatory reference point.
- Component suppliers serving the data center cooling and power distribution markets will increasingly market compliance with the new specification as a selling point, similar to how OCP compliance has functioned in server hardware procurement for the past decade.
- NVIDIA’s MGX adoption numbers are likely to keep climbing past the current 200-plus partner count as Vera Rubin-generation hardware moves toward broader availability, since partners building to MGX today are positioning themselves to support that next GPU generation without a redesign.
The Bigger Picture: Infrastructure as the New AI Battleground
This specification push is part of a broader trend in 2026 where the competitive battleground for AI leadership has shifted partly away from model benchmarks and toward the physical infrastructure needed to train and run those models at scale. Power availability, construction speed, and grid reliability have all become limiting factors that no amount of chip performance can bypass, which is why companies that compete fiercely on AI products are nonetheless finding reasons to cooperate on the underlying plumbing.
It also reflects a maturing recognition that AI data centers are no longer a niche category of real estate or electrical engineering. They are increasingly treated as critical infrastructure, closely watched by grid operators, regulators, and now, apparently, the chipmakers and hyperscalers themselves, who have concluded that a fragmented approach to building that infrastructure is no longer sustainable given the pace and scale of AI-driven demand.
Frequently Asked Questions
What is the Open Data Center Specification v0.7?
It’s an open specification developed with contributions from NVIDIA, AMD, Google, and Meta that covers data center facility layout, cooling infrastructure, and power architecture, aimed at standardizing how AI-focused data centers are physically designed and built.
What is the OCP Grid Resilience Specification?
It’s a separate specification, co-authored by NVIDIA, Google, Meta, and Microsoft, that defines how data centers should respond to utility-grid disturbances, including staying connected during disruptions and restoring near-full load in under two seconds.
Why are NVIDIA, AMD, Google, and Meta cooperating when they compete directly?
Infrastructure standardization is a case where competitors can benefit from shared standards even while competing on products. A common specification lowers costs for component suppliers, speeds up construction, and reduces the validation burden chip vendors face when deploying hardware across different customer facilities.
What is NVIDIA’s MGX architecture?
MGX is NVIDIA’s open, modular reference architecture for GPU server and rack design. NVIDIA has contributed its Vera Rubin rack and compute-tray designs to the Open Compute Project through MGX, and more than 200 industry partners have adopted the platform.
Does this standardization effort solve the AI data center power shortage?
Not directly. The specifications address how facilities are designed and how they interact with the grid during disturbances, but they don’t create new power generation capacity. They’re aimed at removing engineering and construction friction, which is a separate problem from the underlying shortage of available grid power in many data center markets.
Is this related to the Open Compute Project (OCP)?
Yes. Both specifications were published under the Open Compute Project, the industry body originally founded by Meta in 2011 to open-source data center and server hardware designs. OCP has since become the standard venue for competing companies to collaborate on infrastructure specifications.
Will this change how fast new AI data centers can be built?
Potentially, over time. Standardized specifications can shorten supplier qualification and design validation timelines for new facilities, though the effect will likely show up gradually in newly started projects rather than immediately across the existing base of data centers.
Who else might join the specification going forward?
Reports have not named additional companies beyond NVIDIA, AMD, Google, Meta, and Microsoft as of this announcement, but OCP’s history suggests other chipmakers, cloud providers, and infrastructure suppliers are likely to adopt or contribute to the specification over time.
