新川サーバー | エンタープライズ サーバー ハードウェア サプライヤー
The Rack That Consumes 40 kW and Spits Out 120 Degrees of Heat
A single rack of eight Dell XE9680 servers with H100 GPUs draws approximately 40 kW at full load — nearly the same power consumption as 12 average American homes. At 40 kW per rack, the air cooling system required to remove that heat is enormous: 12+ air handlers per row, raised-floor plenum pressure adjustments, and cold-aisle temperatures that need to drop to 16-18 degrees Celsius just to keep GPU inlet temperatures under 40 degrees. At 40 kW per rack, air cooling is not inefficient. It is structurally incapable.
で 2026, the thermal tipping point for data centers is clear: racks above 30-40 kW require liquid cooling. Every hyperscaler has crossed this threshold. Enterprise data centers are now crossing it as AI infrastructure deployments drive per-rack power density from 8-12 kW (typical enterprise) に 40-100 kW (AIトレーニング). This article covers what liquid cooling means for your next server purchase.
The Three Cooling Technologies in 2026
| テクノロジー | How It Works | Max Rack Density Supported | Maturity | Server Impact |
|---|---|---|---|---|
| Air Cooling (CRAC/air handlers) | Cold air pushed through raised floor or hot-aisle containment. Servers with heatsinks and fans. | Up to ~30 kW per rack (with hot/cold aisle containment) | Mature, well-understood, cheap | None — standard servers work |
| Direct-to-Chip (Cold Plate) Liquid Cooling | Coolant (water or dielectric fluid) pumped through cold plates attached directly to CPUs, GPU, and memory modules. Removes 70-80 percent of heat at the source. Air handles the rest. | Up to ~100 kW per rack (limited by power distribution, not cooling) | Production-ready. NVIDIA, デル, xフュージョン, HPE all ship direct-to-chip solutions. | Requires liquid-cooled server variant (cold plates installed at factory). Air-cooled servers cannot be retrofit easily. |
| Immersion Cooling | Entire server submerged in dielectric coolant. 100 percent of heat captured at source. No fans required. | まで 150+ kW per rack | Emerging. Grindr, eMerge Americas, and several AI startups deploying. Oil-based fluids preferred. | Requires immersion-certified server. Major redesign. Warranty implications. |
Server-Level Liquid Cooling Options in 2026
| プラットフォーム | Air-Cooled Option | Direct-to-Chip Liquid-Cooled Option | Notes |
|---|---|---|---|
| Dell XE9680 (8x H100) | Available (8,500W peak, air-cooled with high airflow) | XE9680 with Dell Liquid Cooling Solution (cold plates on GPUs + CPUs) | Air-cooled works up to ~8 GPUs at 700W each. Liquid cooling recommended for sustained training loads. |
| Dell R770 (2x Xeon 6, up to 2x 450W GPU) | Available (standard air-cooled) | Not typically liquid-cooled (below thermal threshold) | R770 runs fine air-cooled. Liquid cooling not needed below ~2 kW per node. |
| xFusion G5500 V7 (10x H100/A100) | Available (10,000W peak, air-cooled with GPU riser airflow) | G5500 V7 with xFusion liquid cooling (cold plates) | 10x 700W H100 = 7 kW GPU heat. Air cooling marginal for sustained 100 percent utilization. |
| HPE Cray EX (exascale class) | 利用不可 (liquid only) | Exclusive liquid cooling | The first exascale systems were liquid-cooled by necessity. This is the reference point for where enterprise is heading. |
The Economics: Air vs Liquid for AI Workloads
| Cost Category | Air-Cooled (40 kW rack, 8x XE9680) | Direct-to-Chip (40 kW rack, 8x XE9680 liquid) | Savings with Liquid |
|---|---|---|---|
| Initial cooling infrastructure (per rack) | $18,000 (4x air handlers, containment, raised floor) | $25,000 (CDU + rack manifolds + coolant distribution) | Liquid costs more upfront: +$7,000 |
| Annual electricity — cooling (at $0.12/kWh) | $30,000 (cooling consumes ~35 percent of IT power in air-cooled environments) | $10,000 (liquid cooling consumes ~10 percent of IT power) | Liquid saves $20,000/year |
| Annual electricity — IT (same IT load) | $42,000 | $42,000 (same) | Same |
| Space efficiency | 1 row of racks per 8 air handlers | 2x rack density in same footprint | 50 percent more racks per square meter |
| 5-Year TCO (cooling only) | $168,000 | $75,000 (initial + 5 years electricity) | Liquid saves $93,000 per rack over 5 年 |
| Break-even point | – | ~5 months (initial premium offset by electricity savings) | Liquid pays for itself in under 6 months for AI racks |
The economics flip decisively at AI densities. An air-cooled 40 kW rack costs $168,000 in cooling over 5 年. A liquid-cooled rack costs $75,000 — a $93,000 saving per rack. For a 20-rack AI infrastructure deployment, つまり $1.86 million in cooling savings over 5 年. The liquid cooling premium is paid back in under 6 months.
What This Means for Your Next Server Purchase
- If you are buying GPU servers for AI training, ask for the liquid-cooled variant. The XE9680 and G5500 V7 both offer direct-to-chip liquid cooling. The premium is 5-8 percent of server cost; the payback is under 6 months at full utilization. Air-cooled GPU servers will be the exception, not the rule, by 2028.
- If you are buying general-purpose servers (R760, R770, 2288H V7), air cooling is still correct. Below ~2 kW per node, liquid cooling does not pay back. The R770 with dual 450W GPUs sits at the edge — evaluate liquid cooling only if you plan sustained 100 percent GPU utilization.
- Plan your data center power and cooling infrastructure for the peak, not the average. A rack that runs at 8 kW today with air cooling may need to host an 8x GPU server at 40 kW next year. Design your raised floor, power distribution, and aisle containment for 30-40 kW per rack capability even if you deploy at 10 kW today.
Source AI-Ready Server Infrastructure Through Xincuan
We supply Dell XE9680 and xFusion G5500 V7 GPU servers in both air-cooled and direct-to-chip liquid-cooled configurations. 工場直送価格, 3-年保証, そして世界的な発送. Our engineers can help you evaluate whether liquid cooling pays back for your specific AI workload density.
Request a cooling strategy consultation and quote
Explore: AI Servers | Dell PowerEdge | xフュージョン