Missed SanDisk And Micron? 10 Stocks For The Next AI Wave – Carrier Global (NYSE:CARR), Micron Technology
Anyone who caught the memory shortage early on is sitting on the trade of the cycle. But, after such a massive rally, anyone still chasing it is could be buying the top.
The capital that powered that move is already looking for what comes next.
And the next AI backlog story is forming somewhere else entirely — in the pipes, plates and pumps that keep AI factories from melting.
You Missed The Memory Trade, Don’t Miss The Liquid Cooling One
According to 22V Research the answer is liquid cooling — the plumbing that keeps Nvidia Corp. (NASDAQ:NVDA)‘s next-generation racks from cooking themselves alive.
In a note shared Wednesday by analyst Dauvin Peterson, liquid cooling is seen as “the next significant AI-connected order and backlog story.”
The reasoning is mechanical, not narrative. Air can no longer pull heat off the chips fast enough.
The threshold where conventional air cooling fails is around 41.3 kilowatts per rack. Nvidia’s Vera Rubin 72 and Vera Rubin 144 racks are designed for 155 and 200 kilowatts.
The speculative Kyber rack is rumored to breach 500. A one-megawatt rack is no longer hypothetical.
At those densities, water — not air — becomes the only option. Vera Rubin systems shipping later this year arrive with integrated cooling trays, piping and connectors as standard.
There is no air-cooled version.
The transition is a hardware constraint baked into the next generation of AI silicon.
Why The Order Books Are Already Moving
The early signal that this trade is real is the order books at companies that sell the equipment.
Carrier CEO David Gitlin told analysts the company has begun booking meaningful orders for 2027, with capacity still available to take more incremental business this year.
Trane Technologies PLC (NYSE:TT) followed with its own data point. Total first-quarter bookings hit a record $6.7 billion, up 27%, with backlog now at $10.7 billion — up more than 30% versus year-end 2025.
Trane has flagged that customer lead times have stretched from six to nine months historically out to twelve to eighteen months, as buyers scramble for certainty of supply. The company also closed its acquisition of Liquidstack in March, anchoring it inside the two-phase liquid cooling race.
The order curve is steepening.
The Math Behind A Gigawatt AI Factory
Think of an AI factory the way you would think of a power plant. The basic unit is the megawatt — roughly enough electricity to run a thousand American homes.
The next generation of AI factories will hit one full gigawatt.
Building one at gigawatt scale costs around $75 billion today. The cost-per-megawatt math is what makes the number real.
A year ago, one megawatt of AI compute cost roughly $10 million to build out.
Today it runs over $20 million — doubled in twelve months, mostly because the chips inside got more expensive and more power-hungry.
Compute alone now eats more than 60% of the total bill. Inside that $20-million-per-megawatt budget, thermal management accounts for 10% to 20%.
The liquid cooling kit specifically — pipes, cold plates, coolant distribution units — runs $1.0 million to $1.5 million per megawatt.
Now scale it up. 22V estimates the United States will add 50 gigawatts of new AI compute between now and 2030.
Multiply by the per-megawatt cooling cost, add a redundancy buffer, and the U.S. liquid cooling market alone lands between $55 billion and $82.5 billion over four years. Widen the lens to the full thermal envelope and the number stretches toward $200 billion.
That is what makes cooling load-bearing rather than ancillary.
A chip that throttles because it cannot dissipate heat is a chip earning a fraction of its potential. The decision is no longer whether to liquid-cool. It is which vendor wins the contract.
10 Stocks Benefiting From The Liquid Cooling Trade
The 22V note identifies a concentrated group as primary beneficiaries.
The table below pairs each name with the latest analyst consensus from Benzinga’s coverage, alongside current price and implied potential to consensus.
Image: Shutterstock
