NVIDIA MGX 1.2 Explained: What 250kW Racks Mean for GB300

NVIDIA contributed its MGX 1.2 rack and tray designs to the Open Compute Project at OCP 2026, and the headline number is 250kW per rack, nearly double GB200 NVL72's spec. That number alone doesn't tell you much. What matters is why Rubin's power envelope forced a new busbar and a new cooling connector, and whether 250kW is actually enough runway or just a bigger bucket that fills up again in a year.

TL;DR: What NVIDIA MGX 1.2 Changes for GB300 and Vera Rubin

• MGX 1.2 is NVIDIA's rack-and-cooling spec, given to OCP at the 2026 summit, supporting racks up to 250kW. Not a new GPU.
• GB200 NVL72 draws up to 120kW nominal. GB300 NVL72 runs 135kW TDP, peaking at 155kW, per Lenovo's rack-scale guide.
• Vera Rubin NVL72 is reported, not NVIDIA-confirmed, at 190kW to 230kW, still under the 250kW ceiling.
• New hardware: a liquid-cooled 5,000A busbar and blind-mate MQD/MQDB connectors that swap cold plates without leak risk.
• The ceiling expires fast: NVIDIA's 2027 Rubin Ultra "Kyber" rack is already disclosed at 600kW, beyond MGX 1.2's scope.
• Spheron lists GB300 NVL72 and Vera Rubin (R100) as reservable today. Reserve GB300 NVL72 capacity →

What NVIDIA Actually Opened at OCP: a Rack Spec, Not a New GPU

It's easy to read "NVIDIA MGX 1.2" next to "Vera Rubin" and assume a chip announcement. It isn't one. MGX is NVIDIA's modular rack-and-tray reference design, the physical and electrical blueprint that data center builders and ODMs use to actually house GPUs at rack scale, and version 1.2 is the update NVIDIA handed to the Open Compute Project at OCP 2026.

The rack is the thing that needed a new spec because the chip's own architecture made the server obsolete as a unit of measurement. Since GB200 NVL72, Blackwell's NVLink domain spans the whole rack, not a single box: 72 GPUs act as one coherent accelerator connected by NVLink switches, so the rack, not the server, is the smallest thing you can power, cool, or sell. A spec written for individual servers has nothing to say about a 72-GPU NVLink domain that lives inside one liquid-cooled enclosure, which is why the busbar, the cooling connectors, and the tray layout all had to be rewritten as rack-level problems rather than server-level ones.

What shipped is the plumbing, not the chip. Converge Digest reported from the summit that "MGX 1.2 adds blind-mate liquid cooling, a liquid-cooled 5,000A busbar, modular I/O bays, fanless liquid-cooled compute trays, native 19-inch equipment support and standardized interfaces." Wccftech reported that NVIDIA built the spec around Vera Rubin, stripping compute trays down so they carry no loose cables, hoses, or fans at all, and that the MGX ecosystem has grown to nearly 100 partners.

That ecosystem is the point. Giving away the rack blueprint means more vendors can build MGX-compliant racks without reverse-engineering NVIDIA's reference design from scratch, which is exactly what has to happen before Vera Rubin shows up across more than one or two cloud providers.

Alongside MGX 1.2, NVIDIA also published the Open Data Center Specification v0.7 and, jointly with AMD, Google, Meta, and Microsoft, an OCP Grid Resilience Specification, according to AI Weekly's summary of the summit. Two separate spec tracks, one event: electrical and cooling standards for the rack, and resilience standards for the grid connection feeding it. Both matter to the same buyer, for different reasons.

NVIDIA and Cisco also used the same OCP window to push 400Gb/s-per-lane USFP optics and DSX OS updates, per StorageReview's coverage, which is the networking half of the same story: faster racks need faster fabric between them, not just more power inside one rack.

GB200 NVL72 Power Consumption: Where the Rack Math Starts

GB200 NVL72 draws up to 120kW per rack on NVIDIA's own nominal spec, a figure confirmed in rax.ae's data center hosting guide, making it the first NVIDIA rack-scale product that forced most data centers to rethink power density from the ground up. That's well beyond what a conventional air-cooled server rack is built to carry, which is why GB200 NVL72 shows up in a facility as a dedicated electrical and cooling project, not a drop-in swap for an existing rack position.

For the full GB200 NVL72 spec table, including GPU count, memory, and NVLink bandwidth, see the GB200 NVL72 guide rather than re-deriving it here. The number that matters for this post is the 120kW baseline, because it's the floor every later generation has been measured against.

That 120kW figure also answers a related question people search alongside it: how much power an individual NVIDIA data center GPU draws, as opposed to the rack it sits in. The two numbers aren't the same thing, and mixing them up is the most common mistake in facility planning for Blackwell racks. NVIDIA doesn't publish a standalone per-chip wattage figure for the Blackwell or Vera Rubin generations the way vendors once did for single-GPU cards. Once 72 GPUs are wired into one NVLink domain, the rack becomes the thing NVIDIA measures and discloses, not the individual die, which is itself the practical answer to "how much power does the chip use": the question doesn't have a published answer at this scale, only a rack-level one. A facility that can power one GB200 NVL72 rack can't assume it can power two without re-running the electrical math, because 120kW times the rack count stacks fast against a building's total power allocation.

250kW Racks and the 5,000A Busbar: Why Rubin's Electrical Needs Outgrew GB200's

GB200 established the 120kW baseline. GB300 NVL72 pushed it to 135kW TDP, with peaks to 155kW depending on workload, per Lenovo's GB300 NVL72 rack-scale product guide. Vera Rubin NVL72 goes further still: NVIDIA hasn't published an official wattage, but analyst Ming-Chi Kuo's widely-cited estimate puts it at roughly 190kW in a power-saving Max-Q profile and 230kW in full-performance Max-P, a framing NVIDIA's own developer blog corroborates without confirming the exact number. Our Vera Rubin NVL72 guide covers the full preliminary spec table NVIDIA has published for compute, memory, and networking.

Each of those steps is a real jump in current, not just heat. More power at the same or similar voltage means more amps, and the physical copper that carries those amps has to grow with it or the rack itself grows to fit it. NVIDIA's answer in MGX 1.2 is a new liquid-cooled 5,000A busbar, which the same Wccftech report says delivers that current "without making the rack bigger." Liquid-cooling the busbar itself, not just the compute trays, is what makes that possible: running 5,000A through solid copper at rack density generates enough resistive heat that air cooling the bus bar alone would need more clearance than the rack has to spare.

This is the mechanical reason MGX 1.2 exists as a distinct spec revision rather than a minor rack refresh. GB200's 120kW could ride on the generation of rack electrical hardware NVIDIA was already shipping. Rubin's 190-230kW genuinely couldn't, which is why the busbar, not the GPU, is the part of this announcement that had to change first.

Blind-Mate Liquid Cooling: What Changes for Operators Hosting GB300 or Rubin

The second structural change in MGX 1.2 is how cold plates connect to compute trays in the first place. MGX 1.2 introduces rack-scale blind-mate liquid cooling through new MQD (MGX Quick Disconnect) and MQDB (blind-mate variant) connectors, per the same Wccftech report. Blind-mate means a technician slides a tray into the rack and the liquid-cooling connection makes itself, the way a server already self-connects to power and networking in a well-designed rack, instead of someone manually threading a hose fitting by hand.

For an operator hosting GB300 or Rubin racks today, this changes two things in practice:

• Tray swaps drop from hours to minutes. Manually disconnecting and reconnecting liquid cooling hoses on a dense rack is slow and carries real leak risk every time a fitting is reseated. Blind-mate connectors remove the manual step, which is also removing the most common point of operator error in liquid-cooled maintenance.
• Fanless compute trays become standard, per the same reporting, because the cooling path no longer depends on a technician getting a hose connection right by feel. That simplifies the tray itself and removes a moving part that used to fail.

Neither change is visible to a renter paying for GPU-hours. Both are the difference between a facility that can turn over a failed GB300 tray during a maintenance window and one that takes it down for half a shift. If you're evaluating a data center partner's readiness for next-generation racks rather than running one yourself, our GPU colocation guide covers what to check in a facility's cooling and power contract before you sign one.

GB300 Liquid Cooling Requirements, Now That They're Standardized

MGX 1.2 standardizes the connector. The thermal load it has to move is GB300-specific, and Lenovo's rack-scale product guide gives the clearest published numbers for what a facility actually has to supply.

Run a warmer facility loop, which is cheaper to operate because it needs less mechanical chilling, and the rack needs nearly three times the volume of water moving through it to shed the same heat.

Roughly 90% of a GB300 rack's heat is captured by the liquid loop, with the remaining 10% handled by air, per the same Lenovo guide. That ratio is what makes a rack this dense survivable in a standard data center hall at all: almost none of the heat is the facility's room-level air handling problem, nearly all of it leaves through pipes a cooling distribution unit has to be sized for in advance. A facility planning a GB300 retrofit has to pick its supply temperature strategy before it knows its real flow-rate requirement, not after.

For how GB300's liquid cooling and memory specs stack up against GB200 NVL72 on paper, including the pricing gap neither NVIDIA nor cloud providers publish officially, see the GB300 NVL72 vs GB200 NVL72 comparison.

How Much Runway Does 250kW Actually Buy? (The Part Nobody's Covering)

Every piece of coverage of MGX 1.2 states the 250kW figure as a standalone spec. Laid against the actual generational power curve, it reads differently: as one generation of headroom bought on purpose, not a ceiling meant to last.

GB200 at 120kW, GB300 at 135-155kW, and Vera Rubin's reported 190-230kW all fit comfortably under MGX 1.2's 250kW ceiling. That's the point: NVIDIA sized the spec to cover the generation it was built for with margin left over, which is consistent with AI Weekly's framing of MGX 1.2 as the vehicle NVIDIA chose specifically to open Vera Rubin's rack blueprints. What the spec doesn't cover is already on NVIDIA's own roadmap. Tom's Hardware reported that NVIDIA showed off Rubin Ultra "Kyber" racks built around 600,000 watts, due in 2027.

600kW is 2.4 times MGX 1.2's own 250kW ceiling. Whatever rack-and-tray spec eventually houses Kyber will need a new busbar rating, a new cooling path, and very likely a new spec number, because 250kW genuinely cannot stretch to cover it. Read plainly: MGX 1.2 solves Vera Rubin's power problem for the 2026-2027 window and buys the industry roughly one generation of runway. It is not evidence that rack power has stopped climbing, and operators retrofitting for 250kW today should plan their electrical headroom with Kyber's number in mind, not just Rubin's.

Here's the arithmetic a facility operator actually has to run this quarter. Take a flat 2MW power allocation, a realistic AI-ready budget for a mid-size colocation hall. At GB200's 120kW nominal draw, that 2MW fits 16 racks. At GB300's 155kW peak, the same 2MW ceiling fits 12 racks, a 25% drop in rack count for the same power budget and no change in floor space. At Vera Rubin's reported 230kW Max-P, it drops to 8 racks, half of what GB200 could fit in the identical hall. MGX 1.2's 250kW limit doesn't move any of those numbers; what moves is how many racks of a given generation a fixed power allocation can actually energize, and that number, not the GPU count on a spec sheet, is what decides how much compute capacity a facility can actually bring to market at once.

What an Open Rack Standard Means for GB300 and Vera Rubin Cloud Availability

An open, published rack spec changes who can build compliant infrastructure, not just how fast any one builder can do it. Before MGX 1.2 went to OCP, a data center wanting to host Rubin-class density had to work from whatever NVIDIA shared directly or reverse-engineer a reference design. With the spec public and nearly 100 partners already building to it, more facilities can get to a compliant 250kW rack in parallel instead of queuing behind NVIDIA's own direct engineering support, one data center at a time.

That doesn't collapse the timeline to zero. MGX 1.2 standardizes the electrical and cooling interface; it doesn't manufacture 5,000A busbars or qualify a facility's chiller plant for 230kW racks overnight. Each operator still has to retrofit power distribution, requalify its liquid cooling loop against GB300's or Rubin's actual flow-rate numbers, and get its own facility certified. What the open spec removes is the design uncertainty: a builder working to a published standard isn't guessing at tolerances NVIDIA hasn't shared.

For where Vera Rubin cloud availability and cost per token actually stand right now, independent of any single facility's retrofit progress, see the Vera Rubin NVL72 cloud availability guide.

Getting GB300 or Vera Rubin Capacity on Spheron

MGX 1.2 being open doesn't mean every data center partner is equally ready for it today. Facility-by-facility retrofit timing is exactly why sourcing GB300 or Rubin capacity through a single operator's readiness is a risk a renter doesn't need to carry directly.

Spheron's pricing page lists GB300 NVL72, GB200 NVL72, and Rubin (R100) as reserve-only line items, with no published hourly rate, the same as every provider checked for this post and for our existing GB300-vs-GB200 and Vera Rubin coverage. Reserving one of these racks routes through Spheron's Supplier Matchmaking: Spheron sources from its certified data center network, negotiates pricing, and handles setup, rather than requiring you to vet any single facility's busbar rating or cold-plate readiness yourself. Because that sourcing spans a network rather than one owned fleet, one partner being mid-retrofit on MGX 1.2 compliance doesn't block the reservation; Spheron routes it to a partner that's ready.

That's also the honest limit here: Spheron doesn't operate or publish its own MGX 1.2-compliant facility specs. The power and cooling retrofit described in this post is a decision made by the underlying data center partner, not something a renter configures through Spheron directly. If you need rack-scale capacity you can book by the hour today rather than reserve ahead, standalone B300 and B200 GPUs are available on-demand now, outside the rack-scale reservation path, at $12.42/hr and $9.58/hr per GPU respectively as of 14 Oct 2026.

Pricing fluctuates based on GPU availability. Spheron rates above are live as of 14 Oct 2026; other providers reflect their most recent published rates and may have changed. Check current GPU pricing → for live rates.

If you operate a facility weighing an MGX 1.2 retrofit and want to list that capacity rather than rent it, Spheron's partner program routes vetted renter demand to data centers that have already made the investment.

MGX 1.2 standardizes what a Rubin-class rack needs; actually having one ready to book is still a facility-by-facility race. Spheron's certified network and reservation path mean GB300 and Vera Rubin capacity doesn't wait on any single operator's retrofit finishing.

Reserve GB300 or Vera Rubin capacity on Spheron →