Is liquid cooling mainstream in 2026?

Above roughly 30 kW per rack, yes — direct-to-chip is the default once accelerators pass 700 W and rack-scale systems pass 100 kW. Below 30 kW, sealed air-cooled micro data centers remain viable, and the turnkey liquid cabinet for the 20-50 kW middle ground is still an open market gap.

Has the Rubin generation started shipping?

Yes. NVIDIA launched Rubin at CES 2026 (January), ramped it into full production in May, and the first Vera Rubin NVL72 racks shipped September 23 (Supermicro), with CoreWeave running production workloads from September 30. The generation is defined by 2.3 kW-class GPUs, 180-220 kW racks, official 45 °C warm-water cooling and 100% liquid coverage — there is no air-cooled SKU.

What replaced 3M fluorinated cooling fluids?

For direct-to-chip, PG25 water-glycol — already the default — consolidated its position after 3M exited PFAS manufacturing at the end of 2025: it is cheap, non-flammable and outside PFAS regulation. Immersion cooling shifted toward engineered oils and remaining-chemistry suppliers, with fluid sourcing now a formal procurement risk item.

Digest · 7 min read · Updated 2026-10-12

State of liquid cooling, 2026

TL;DR

Above ~30 kW, liquid is no longer a choice — it is the physics. The Rubin generation — 2.3 kW-class GPUs in 100%-liquid racks fed with 45 °C warm water — started shipping in September 2026. Fluorinated fluids are gone (3M exited PFAS manufacturing at end-2025), leaving PG25 water-glycol as the default. Standards matured (OCP ORv3, SS 726:2026 for the tropics). What still does not exist: a turnkey, facility-free liquid cabinet in the 20-50 kW class — the gap this site tracks.

The power cliff became a plateau

When accelerators crossed 700 W and rack-scale systems crossed 100 kW (GB200-class racks), direct-to-chip moved from exotic to default in two years. The budgeting math is now routine: TDP sums × 1.2 for power, 60-80% capture for the liquid share, 1.1× for the room. What used to be a research topic is a checklist — this site's Formulas section is that checklist.

The Rubin generation is shipping

What was a roadmap item at the start of 2026 is now delivery news: NVIDIA launched Rubin at CES in January (in production on day one), announced the full-production ramp in May, and the first Vera Rubin NVL72 racks shipped September 23 (Supermicro — up to 6,000 racks/month at full production), with CoreWeave running production workloads on them from September 30. Engineering meaning: 2.3 kW-class GPUs (Max-P), 180-220 kW racks, 45 °C warm-water cooling made official, and 100% liquid coverage with zero fans — the first platform generation with no air-cooled SKU at all. AMD's Helios racks (MI455X) entered full production the same quarter, so the physics now applies across both ecosystems. The rack builder carries GB300 and Rubin rows with per-selection sources.

Fluids after PFAS

3M's exit from PFAS manufacturing (announced 2022, completed end-2025) removed the dominant fluorinated fluid supplier from the market. For direct-to-chip the answer was already sitting there: PG25 — cheap, non-flammable, unregulated as a chemical class. Immersion's fluid story got harder, which is one reason direct-to-chip consolidated its lead (architecture comparison).

Standards consolidated

The interoperability layer is essentially settled: OCP's UQD couplings and ORv3 blind-mate manifolds define how parts fit; ASHRAE's W-classes define how warm facility water may be; IEC 62368-1 Ed.4 carries the electrical-safety clauses. The 2026 addition is Singapore's SS 726 — the first national standard to take tropical operation seriously. Practical effect: the RFQ checklist now has named standards to demand on every line.

The tropics stopped being an edge case

Most of the world's next data center capacity is being built in warm climates, where the physics is least forgiving: capacity halves at 35 °C ambient, and wet-bulbs near 30 °C kill free cooling outright. Buyers in these markets should size against derated capacity, never nameplate — the single most common procurement mistake this site exists to prevent.

What still doesn't exist

Verified in our comparison data: major-vendor micro data centers stop at ~10 kW on air, and their liquid products start at 40-70 kW as components for AI halls. A turnkey, facility-free liquid cabinet for 20-50 kW — the hospital-edge, factory-edge, campus-edge case — remains essentially uncovered since the 2018 STULZ×CoolIT attempt. Whoever fills it defines the category.

Industry radar & milestones · Changelog · Cabinet comparison