| Myth | Reality |
|---|---|
| Immersion cooling is the more advanced, forward-looking choice | Immersion currently captures heat more completely but stays niche because it requires redesigning around a submersion tank rather than fitting existing rack infrastructure |
| Liquid cooling is still an edge case for AI data centers | Liquid cooling adoption is projected around 37% in 2026 and is now the default design path for any facility hosting GPU-dense workloads |
| Cold plate and immersion are competing for the same winner-take-all market | The likely outcome is a dual-track market: cold plate for the majority, immersion for specialized extreme-density environments |
Air cooling could not keep pace with GPU power draw once accelerator TDPs crossed roughly a kilowatt per chip, and AI racks packing eight or more accelerators pushed density well past what air handlers in a conventional data center hall can remove. Liquid cooling stopped being an exotic option and became the baseline design assumption for new AI-dense facilities. The remaining question isn’t liquid versus air anymore — it’s which liquid cooling architecture, and that comes down to a choice between direct-to-chip cold plate and full immersion.
How the two architectures actually differ
Direct-to-chip liquid cooling routes coolant through a cold plate mounted directly onto the chip package, pulling heat away at the point it’s generated while the rest of the server — memory, power delivery, storage — stays in open air, cooled conventionally or left passively cooled. Immersion cooling submerges the entire server, board and all, in a dielectric fluid that doesn’t conduct electricity, so every component sheds heat into the surrounding liquid rather than just the hottest chip.
| Direct-to-chip (cold plate) | Immersion | |
|---|---|---|
| What’s cooled | The chip package specifically, via a mounted cold plate | The entire server, fully submerged in dielectric fluid |
| Heat capture completeness | Partial — contacts only one surface of the chip | More complete — every component sheds heat into the fluid |
| Integration effort | Fits into existing rack and facility infrastructure | Requires redesigning around submersion tanks |
| 2026 market position | Default, majority of new AI-dense deployments | Niche, specialized high-density and HPC environments |
| Vendor standard example | NVIDIA ships cold-plate cooling as standard on DGX H200 and MGX B200 systems | Growing interest as chips push past 2,000W envelopes |
Why cold plate won the “default” position despite capturing less heat
Immersion cooling is, on a pure thermodynamics basis, the more thorough solution — full submersion captures heat from every component, not just the chip package, and that matters more as total system power keeps climbing. But data centers don’t get built on thermodynamics alone; they get built on what fits the capital and operational constraints an organization already has. Direct-to-chip integrates into an existing rack form factor and facility power and plumbing design with far less structural change than immersion, which requires a fundamentally different physical layout built around submersion tanks rather than vertical racks. For an operator retrofitting existing halls or standing up new capacity on a tight timeline against a GPU shortage, that integration cost difference dominates the decision far more than the heat-capture completeness difference does.
- 1. Air cooling limit hitGPU TDPs cross roughly 1kW per chip; conventional air handlers in a data center hall can no longer keep pace.
- 2. Liquid becomes defaultLiquid cooling adoption reaches roughly 37% in 2026, the baseline assumption for new AI-dense builds.
- 3. Cold plate takes the majorityDirect-to-chip fits existing rack infrastructure, so it captures the bulk of new deployments despite partial heat capture.
- 4. Immersion holds the extreme-density nicheAs chips push past 2,000W, immersion becomes increasingly attractive for the specific facilities that need its more complete heat capture.
How the AI rack cooling market split by architecture in 2026
The dual-track outcome is a market structure decision, not a technology superiority decision — both architectures are mature, and each wins the segment suited to its integration profile.
Where the crossover point sits
As accelerator power envelopes keep climbing past the 2,000W mark and rack densities approach megawatt levels per rack, the calculus shifts. At that density, cold plate’s partial heat capture starts leaving enough residual heat in the rest of the system that immersion’s completeness becomes the more economical choice again, not just the more thorough one. This is why the most credible market outlook isn’t “one architecture wins,” but a durable dual-track split: cold plate serving the large majority of AI-dense deployments at today’s typical densities, immersion serving a growing minority of the most extreme, highest-density facilities where its heat-capture advantage actually pays for its integration cost.
Choose direct-to-chip Default
Right for the large majority of new AI-dense builds today — it fits existing rack infrastructure, has vendor-standard support (NVIDIA ships it standard on current DGX and MGX systems), and is faster to deploy against a tight capacity timeline.
Choose immersion Extreme density
Worth the integration cost specifically when accelerator power envelopes push past roughly 2,000W and rack density approaches megawatt levels, where partial heat capture starts leaving too much residual heat behind.
Stay on air cooling for new AI-dense builds Already obsolete
No longer viable once GPU TDPs cross roughly 1kW per chip — conventional air handlers cannot remove heat fast enough at modern AI rack densities.
Common mistakes
Common mistake
Choosing immersion cooling because it sounds like the more advanced technology, without accounting for the facility redesign it requires. For most deployments at current density levels, that integration cost isn’t justified by the heat-capture advantage.
Common mistake
Assuming a cooling architecture decision made today will still be optimal once next-generation accelerators push past current power envelopes. The crossover point favoring immersion moves as chip TDPs rise, and a facility designed around cold plate at today’s density may need a costly redesign sooner than planned.
What worked
Operators that planned furthest ahead selected cold plate for their current build while explicitly designing facility power and plumbing infrastructure with enough headroom to support an immersion retrofit later, rather than treating the current architecture choice as permanent.
Glossary
- Direct-to-chip cooling (cold plate)
- A liquid cooling method where coolant is routed through a plate mounted directly onto a chip package, removing heat at its source while the rest of the server remains air-cooled.
- Immersion cooling
- A liquid cooling method where an entire server is submerged in a non-conductive dielectric fluid, capturing heat from every component rather than just the hottest chip.
- TDP (thermal design power)
- The maximum amount of heat a chip is expected to generate under normal workloads, used to size the cooling system required to remove it.
- Dual-track cooling market
- The expected 2026-era market structure where cold plate serves the majority of AI-dense deployments and immersion serves a growing niche of the most extreme-density facilities.
Key takeaways
- Direct-to-chip cold plate cooling is the default choice for AI-dense data centers in 2026, not immersion, despite immersion capturing heat more completely.
- Liquid cooling overall is projected to reach roughly 37% adoption in 2026, having become the baseline design assumption once GPU TDPs crossed roughly 1kW per chip.
- Cold plate wins the majority of deployments because it integrates into existing rack and facility infrastructure with far less structural change than immersion requires.
- NVIDIA ships cold-plate cooling as a standard configuration option on its current DGX H200 and MGX B200 systems, reinforcing cold plate’s default status.
- Immersion cooling becomes increasingly attractive as accelerator power envelopes push past roughly 2,000W and rack densities approach megawatt levels, where its heat-capture advantage starts to outweigh its integration cost.
- The most credible market outlook is a durable dual-track split rather than one architecture displacing the other — cold plate for the majority, immersion for the extreme-density minority.
- A cooling architecture decision made today can become suboptimal as next-generation chip TDPs rise, so facilities benefit from designing in headroom for a later immersion retrofit.
FAQs
Is immersion cooling better than direct-to-chip cooling for AI data centers?
It captures heat more completely since the entire server is submerged rather than just the chip, but direct-to-chip remains the 2026 default because it integrates into existing rack infrastructure with far less structural change, making it the more practical choice at typical current densities.
How much of the AI data center market uses liquid cooling in 2026?
Liquid cooling adoption is projected at roughly 37% in 2026, having become the baseline design assumption for any facility hosting GPU-dense AI workloads once accelerator power draw crossed roughly a kilowatt per chip.
Why did cold plate become the dominant liquid cooling method instead of immersion?
Because it fits into existing rack and facility infrastructure with much less redesign than immersion, which requires building around submersion tanks rather than conventional vertical racks — an important advantage for operators building capacity quickly against a tight GPU supply timeline.
When does immersion cooling become the better choice?
As accelerator power envelopes push past roughly 2,000W and rack densities approach megawatt levels, where cold plate’s partial heat capture starts leaving too much residual heat in the rest of the system, making immersion’s more complete heat capture worth its integration cost.
Do major AI hardware vendors support liquid cooling as standard?
Yes. NVIDIA offers cold-plate liquid cooling as a standard configuration option on its current DGX H200 and MGX B200 systems, reflecting cold plate’s position as the default architecture for AI-dense deployments.
Will one cooling architecture eventually replace the other?
The most likely outcome, based on current industry analysis, is a durable dual-track market rather than one displacing the other: cold plate serving the majority of deployments, and immersion serving a growing niche of the highest-density, most extreme facilities.
Can a data center switch from cold plate to immersion cooling later?
It requires a significant facility redesign, since immersion needs submersion tanks rather than conventional racks, which is why forward-planning operators build extra power and plumbing headroom into cold-plate facilities to make a later retrofit less disruptive.
Why is air cooling no longer viable for new AI data centers?
Because GPU thermal design power has crossed roughly 1 kilowatt per chip, and modern AI racks packing eight or more accelerators generate more heat than conventional air handlers in a data center hall can remove.
References
- Gottog Power — “Liquid Cooling In Data Centers Explodes In 2026”
- JETCOOL — “Direct Liquid Cooling vs. Immersion Cooling for Data Centers”
Related reading on The Tech Trends: Data Center Power Constraints, AI Data Center Buildout, GPU Shortage Economics, and NVIDIA vs AMD vs Custom Silicon.
