Rising component costs may adjust prices on selected items without prior notice. Supply constraints may add shipping time — we’ll notify you if affected. OEM costs and supply tightening — contact us to secure stock and pricing. For projects or bulk orders, choose “Request for Quotation” on any item.
Call our specialists(65) 6978 3502 Email our sales teamsales@sourceit.com.sg

Servers & Power

The Last Air-Cooled Flagship? Why the Fans Stop Spinning After Blackwell Ultra

Today you can still order a flagship 8-GPU AI server, slide it into an ordinary air-cooled rack and switch it on. HPE's Compute XD690 does exactly that with eight Blackwell Ultra GPUs. No pipes, no coolant, no plumber.

Enjoy it while it lasts. On the evidence so far, Blackwell Ultra is the last NVIDIA flagship generation you can cool with air alone. The implications for air-cooled data halls are larger than most buyers have noticed.

Three generations of the 8-GPU server

Compare Hopper (H200) Blackwell Ultra (B300) Rubin (NVL8)
Air Air-only version Yes Common Yes In a tall chassis No None announced
Liquid Liquid version Optional Optional, and denser Required for the GPUs
📏 Typical chassis height About 5U
8U to 10U on air
2U, liquid
⚡ Power per server Lower Roughly 14 to 15 kW
Roughly 24 kW

Sources: HPE AI Servers, 2CRSi's Blackwell to Rubin power and cooling guide, Lenovo HGX Rubin NVL8 product guide. Power figures are third-party approximations and vary by vendor.

What changed with Rubin

The 8-GPU "HGX" server has been the air-cooled workhorse of enterprise AI since the H100. For the Rubin generation, vendors describe the equivalent system, HGX Rubin NVL8, as a 2U liquid-cooled server. Lenovo calls its version fully liquid-cooled. One system builder states plainly that there is no air-cooled version of any Rubin system.

The closest thing to a compromise is a hybrid design. ASUS, for example, has announced a model that liquid-cools the GPU board and uses air only for the CPUs. That reduces the plumbing, but it does not remove it. The GPUs still need a coolant loop.

NVIDIA itself has described Rubin servers as having sealed front panels and no internal fan noise. The fan, in other words, has been designed out.

Sources: ASUS HGX Rubin servers, DataCentreNews on Rubin servers

Why the fans finally lose

For years the industry has announced that air cooling has hit its limit, and for years engineers have pushed the limit back with bigger heat sinks, faster fans and taller chassis. The air-cooled B300 server is the proof: it exists, but it is 8U to 10U tall where its liquid-cooled sibling is half that or less.

Rubin moves the goalposts again. Third-party figures put the 8-GPU Rubin node at roughly 24 kW in 2U. One vendor describes its liquid-cooled Rubin server as fitting in a quarter of the rack space of the previous generation's air-cooled machine. At that density there is simply no room for the volume of air required.

Source: Cisco data sheet for the HGX Rubin NVL8 server

What this means if your hall is air-cooled

Your position What it means
Air Air-only hall, buying this year 🟢 An air-cooled B300 server such as the XD690 is the newest flagship you can host
Air Air-only hall, planning a refresh in two to three years 🟡 The refresh will probably need liquid. Start the facility conversation now.
📝 Signing a long colocation contract 🟡 Ask what the operator's liquid cooling roadmap is, in writing
Liquid Liquid already available 🟢 You have the widest choice in this generation and the next
🧮 Workloads fit on smaller GPUs 🟢 Air-cooled inference servers are not going away. This is a flagship problem.

Two unpopular conclusions

The air-cooled B300 server may age better than expected

Conventional wisdom says to avoid the last of anything. Here the logic may run the other way. If no air-cooled successor arrives, the air-cooled B300 server becomes the most capable machine an air-only hall can ever host. For organisations that cannot or will not add liquid, it is not a stopgap. It is the ceiling, and a ceiling holds its usefulness for a long time.

"We will stay on air" is now a strategy with an expiry date

Plenty of data halls have treated liquid cooling as optional, something for hyperscalers. For flagship AI hardware, that position has perhaps one generation left. This does not mean every hall needs a liquid loop tomorrow. It means that a five-year facilities plan with no liquid in it has quietly made a decision about which hardware the organisation can buy.

An important caveat: this is about flagship training and large-inference GPUs. Mainstream inference servers, such as 4U systems built around lower-power GPUs, continue to run on air and will for years.

Related on SourceIT

Share This Article
Facebook

More from our blog


← Older Post Newer Post →