If AI Racks Reach 250 kW, Where Does Cash Stick?

ANALYSIS · MEMBERS CONTINUE · 5 min free / 14 min full · Data as at 16 Oct 2026

This week in San Jose, the open rack design for NVIDIA's next AI systems was published for up to 250 kW a rack, fed by a 5,000-amp liquid-cooled busbar. On the same Thursday, grid rules for AI data centres passed their final industry ballot, three months after one fault pushed about 3,900 MW of data centre load onto backup power in PJM. If racks really reach 250 kW, where does the cash stick?

In short

• The question: if AI racks reach 250 kW, where does cash stick?
• The answer: Scenario B. Cash sticks first with the makers of power conversion and electrical gear that sit between the grid connection and the rack, then with halls built for dense power from day one, then with the chip buyers who fill them. Halls designed for lighter racks, and operating sites that must meet new ride-through rules, pay last
• Why: the MGX 1.2 rack published at the Open Compute Project summit on 12 October 2026 runs to 250 kW, but its busbar is still a 48-volt part carrying up to 5,000 amps. That much current is where the new money goes: 800-volt power racks from the second half of 2026, row power centres in 2027, and the protection gear around them
• The size of the book: Vertiv guides US$13.8–14.2 billion of 2026 sales, and customer prepayments on its balance sheet doubled to US$3.6 billion in six months. Eaton's Electrical Americas orders are up 41% on a rolling basis. Delta Electronics says AI products will pass 25% of its 2026 revenue
• Behind the wall: who gets paid, in order · the current problem in plain numbers · three scenarios, arithmetic shown · what the grid rules cost an operating hall · six practical takeaways · Members' FAQ
• What would break it: 800-volt gear slipping while ride-through retrofits pile up (A), or lower-power chips and open standards turning power gear into a price fight (C)
• Nothing here is a recommendation. This maps what this week's specifications and rules already assume, with the working shown

What's in this piece: the numbers that matter · why the power path became the clock · what the map assumes in one line · members from here · who gets paid, in order · three scenarios, arithmetic shown · our read · what would change the view, with dates · practical takeaways · FAQ · references

The numbers that matter

Chart 1 — What a 250 kW rack asks of the power path. Second Order arithmetic from the MGX 1.2 specification and NVIDIA's 800-volt roadmap. Education only.

Why did the power path become the clock?

A rack of AI chips needs three things to arrive together: the chips, a hall that can cool them, and power delivered to the rack at the right voltage. The AI Data Centers layer shows where each sits. Our rack-scale AI explainer covers why the whole rack is now the unit of compute. This piece is about the last few metres of cable.

On 12 October, at the Open Compute Project Global Summit, NVIDIA contributed its Vera Rubin rack and tray designs as MGX 1.2. The rack supports deployments up to 250 kW. Its busbar, the thick conductor running down the back of the rack, is the same width as before but deeper and liquid-cooled, so it can carry up to 5,000 amps at a nominal 48 volts. That is the physics problem in one line: at low voltage, more power means much more current, and current means copper, heat and conversion losses.

NVIDIA's answer is 800-volt direct current, developed with Google and Microsoft through the Open Compute Project. On 11 August it said an 800-volt power rack that slots into existing halls arrives in the second half of 2026, a row power centre feeding up to 2 MW per row follows in 2027, and a facility-scale DC power block comes after that.

The second half of the clock sits outside the fence. PJM has now logged three events in which data centres switched to backup generation after a normal grid fault, the largest about 3,900 MW on 22 July 2026. On 15 October, the new reliability standards for computational loads passed their final ballot at NERC, the grid's standards body. FERC has ordered them filed by 31 December. So the money question is not whether racks get denser. It is who gets paid to move that much power the last few metres, and to keep it on when the grid wobbles.

What does the map already assume?

Three answers compete:

1. The power room binds harder. 800-volt gear slips toward 2028, operators push more current through low-voltage busbars, and ride-through retrofits land on operating halls at the same time
2. The power path binds in steps. Hybrid power racks slot into existing halls from late 2026, 800-volt volume builds through 2027, and the grid rules are filed on time. Gear makers collect first; dense new halls next
3. The bind eases. Lower-power chips and inference designs cut current per rack, open standards make power gear interchangeable, and scarcity pricing thins

A tip sheet would pick a power stock. Analysis asks which answer this week's documents already need.

Our read, in one line: the map already assumes Scenario B. Racks head to 250 kW while the rack itself still runs on a 48-volt, 5,000-amp busbar, so cash sticks first with the makers of conversion and electrical gear that move power from the grid connection to the rack, then with halls designed for dense power from day one, then with the chip buyers who fill them. Halls built for lighter racks, and operating sites that must meet ride-through rules, pay last. The stuck step is the power path inside the fence, not the chip. The scenarios and the arithmetic are below.

The free preview ends here.

—— Still to read · ~9 of 14 minutes ——

MEMBERS CONTINUE HERE

• Who gets paid, in order, and who waits
• The current problem in plain numbers
• Three scenarios with the arithmetic
• What the grid rules cost an operating hall
• Six practical takeaways
• Members' FAQ