Air cooling is generally sufficient for H100 servers, since most 8x H100 SXM and PCIe systems are designed and validated by NVIDIA and OEMs to run on standard front to back airflow with adequate rack level cooling capacity. The practical constraint is not the chip itself but the data center's ability to remove roughly 10 kilowatts of heat per rack unit at that density, which requires higher CRAC or CRAH capacity, hot aisle containment, and more airflow than older 3 to 5 kilowatt racks were built for. Liquid cooling, whether direct to chip cold plates or rear door heat exchangers, becomes more attractive when multiple 8 GPU nodes are packed into a single rack, pushing rack density well above what air alone can economically dissipate, or when ambient noise and fan power consumption need to be reduced. Newer and hotter GPUs like the B200 push closer to the point where liquid cooling stops being optional, but the H100 generation still works well in properly provisioned air cooled data centers. The right choice depends on existing facility cooling capacity and planned rack density. Nanobase AI assesses a customer's data center thermal capacity before recommending air or liquid cooling for an H100 deployment.
The real constraint is the room, not the chip
H100 servers are validated by NVIDIA and OEMs to run on standard front-to-back air cooling, so the question is rarely whether the chip needs liquid — it is whether the data center can remove roughly 10 kW of heat per rack unit at the density you plan to deploy. Older data centers built around 3–5 kW racks were simply not designed for this, which is why the cooling conversation is really a facility capacity conversation more than a GPU specification one.
Where the crossover point sits
| Deployment density | Cooling approach | Typical fit |
|---|---|---|
| 1 node per rack (~8–11 kW) | Air, with hot aisle containment | Most enterprise data centers with upgraded CRAC/CRAH capacity |
| 2–3 nodes per rack (~20–30 kW) | Air (high-capacity) or hybrid rear-door heat exchanger | Facilities with reinforced cooling infrastructure |
| 4+ nodes per rack (~40 kW+) | Direct-to-chip liquid cooling | Purpose-built AI data centers |
| B200/GB200-class density | Liquid cooling effectively required | New builds designed around Blackwell |
Most single-node-per-rack H100 deployments never need to cross into liquid cooling. The decision shifts once an organization tries to pack multiple 8-GPU nodes into the same rack footprint to save floor space, which is when air cooling stops being economical even though it may still technically work.
What changes if you choose liquid anyway
Liquid cooling, whether direct-to-chip cold plates or rear-door heat exchangers, is not only a density solution. It also reduces fan power draw and noise, which matters for facilities near office space, and it can allow denser rack layouts that reduce floor space and cabling runs. The tradeoff is added infrastructure: a coolant distribution unit, plumbing, and often a facility water loop that many existing data centers were not built with, representing a real capital project layered on top of the server purchase itself.
A practical decision path
- Confirm current CRAC/CRAH tonnage and hot aisle containment capacity against planned rack count and density.
- Calculate total heat load per rack at your planned node-per-rack density (roughly 10 kW per H100 SXM node).
- If total per-rack heat load stays within existing cooling capacity with reasonable headroom, air cooling is the simpler and cheaper choice.
- If density plans require 3+ nodes per rack, or the facility already runs close to its cooling ceiling, evaluate liquid cooling infrastructure costs against the alternative of leasing more rack space to stay air-cooled.
- Revisit this decision before any Blackwell-generation upgrade, since B200 and especially GB200 NVL72 push much closer to requiring liquid cooling outright.
Frequently asked questions
Do all H100 servers ship with air cooling by default?
Most OEM H100 servers, including standard HGX and DGX H100 configurations, are designed for air cooling out of the box. Liquid-cooled variants exist but are typically chosen deliberately for density or noise reasons rather than being required by the GPU itself.
At what rack density does liquid cooling become necessary rather than optional?
There is no universal number since it depends on the specific facility's cooling capacity, but many data centers find liquid cooling becomes the more economical choice somewhere between 20 and 40 kW per rack, well below what B200 and GB200-class systems demand.
Does liquid cooling improve H100 performance?
Liquid cooling does not increase an H100's rated clock speeds, but it can reduce thermal throttling risk in poorly ventilated environments and lower fan noise and power draw, which matters more for facility operations than for raw GPU throughput.
Is retrofitting liquid cooling into an existing air-cooled data center expensive?
It is a genuine infrastructure project involving plumbing, leak detection, and often a coolant distribution unit tied into a facility water loop, so it should be budgeted and planned as a capital project rather than treated as a simple hardware swap.
How Nanobase AI helps
Nanobase AI, an enterprise AI engineering company with engineering headquarters in Silicon Valley, assesses a customer's actual data center thermal capacity before recommending air or liquid cooling for an H100 deployment, avoiding both under-provisioned facilities and unnecessary liquid cooling spend. Learn more about our GPU infrastructure services.
Ready to discuss your project? Contact Nanobase AI or email hello@bumu.tech.