Weekend Edition · Fri 16 Oct 2026
AI's Infrastructure Race Moves Beyond Chips to Power, Cooling and Grid Resilience
This week's Open Compute Project summit put the plumbing of AI at centre stage: power distribution, liquid cooling and how data centres behave on the grid. The announcements are mostly specifications and company claims; deployment and measured benefits come later.
Chanan Zevin - Chief Editor and Head of Desks
Published
Why specifications matter
The Open Compute Project, founded by Meta, is an industry body that publishes shared hardware designs. Its Global Summit in Silicon Valley this week drew a reported record of about 15,000 attendees, according to CRN. The pitch for common specifications is simple: if racks, power shelves, cooling connectors and telemetry follow agreed interfaces, operators can buy from several suppliers and suppliers can build to one design rather than many.
In a post on October 12, OCP leaders wrote that power densities are "on track to reach and then surpass 1MW in a single rack," and described the data centre as "one large chip to grid design, deployment and operations problem."
What was published
Nvidia contributed version 1.2 of its MGX rack and tray specification. According to the published document and the company's description reported by CRN, it covers liquid-cooled racks with a 48-volt DC, 5,000-amp busbar, blind-mate liquid cooling and standard quick-disconnect couplings, for rack deployments up to 250 kW.
OCP also published the Mt. Diablo Disaggregated Power Rack specification, version 1.0, an air-cooled rack that houses power equipment separately from compute and delivers ±400 volts DC, with battery backup and telemetry. OCP's own summary of the year lists a white paper on 800-volt DC distribution and a path to 1,500 volts, guidelines for energy-storage safety, a requirements specification for solid-state transformers converting medium voltage to 800 volts DC, and a specification covering how hyperscale sites ride through grid disturbances and recover active power.
Company announcements
Chipmakers tied products to the shared designs. CRN reported that AMD unveiled a reference rack design called Helion, based on an OCP specification, with server vendors planning configurations from November; Intel showed Xeon racks built on the Open Rack v3 standard that can pack more than 46,000 cores into a 133.45 kW power envelope, by its account.
Expected benefits, such as faster assembly, easier service and better efficiency, are the companies' claims. Nvidia said MGX 1.2 "improves assembly, test and service times"; we found no independent measurement of that yet.
Specifications are not deployed equipment
A published specification is an agreed design, not equipment in service. OCP's post itself notes that busway and layout designs for 800-volt DC to the rack are "still incomplete" and that on-site generation remains "an unsolved problem." The grid ride-through specification sets requirements; whether sites meet them in real disturbances will only be known from operating data.
The investment numbers are forecasts too: OCP cites an IDC study projecting more than $890 billion of spending on OCP-recognised IT equipment in 2030.
Why it matters
Power and cooling now limit how fast AI capacity can be built as much as chip supply does. Common interfaces could lower cost and speed construction, and grid-behaviour standards matter to utilities deciding whether to connect very large loads. Independent evidence will come from deployments, not summit slides.