Inference margins live and die on cost per GPU-hour. I400C40/I400C45 attacks all three inputs — up to 512 PCIe GPUs per container, PUE ≈ 1.05 on free cooling, zero building — RTX 4090/5090, RTX PRO 6000, H100, H200 fleets running in 75–90 days. 推理利润取决于单卡小时成本。I400C40/I400C45 同时压低三项输入 — 每箱最多 512 张 PCIe GPU、自然冷却气候下 PUE ≈ 1.05、无需土建大楼 — RTX 4090/5090、RTX PRO 6000、H100、H200 机群可在 75–90 天内投运。
You've seen the business case — modular blocks, fast monetization, inference-first economics. I400C40 and I400C45 are the Immersion Computing Units in that equation: a 40-ft or 45-ft container holding 8× A32 single-phase immersion tanks plus one 10 kW air-cooled rack. Servers sit fully submerged in dielectric fluid; each tank carries up to 50 kW with dual redundant CDUs, and all heat leaves the container on a warm-water loop at 32/37 °C. Everything is integrated and tested at the factory — on site you connect power, water, and network. 商业逻辑你已经看过——模块化区块、快速变现、以推理为先的经济模型。 I400C40 与 I400C45 就是这个等式里的浸没式计算单元：一台 40 或 45 英尺集装箱， 内置 8 个 A32 单相浸没液舱加 1 个 10 kW 风冷机柜。 服务器完全浸没于绝缘冷却液中；每个液舱最高承载 50 kW，配双路冗余 CDU， 全部热量经 32/37 °C 温水回路排出箱外。 所有系统在工厂完成集成与测试——现场只需接入电、水、网。
I400C45SUT50 (I400C45) · I400C40ST50 (I400C40)IT Load vs. Facility Load. 400 kW is what the servers draw. Total facility load is ~440–500 kW once cooling and auxiliaries are added — the gap is set by PUE, which depends on your site's climate. *Sites whose historical peak dry-bulb stays ≤ 24 °C run on pure dry coolers (free cooling, lowest PUE); hotter sites add a hybrid chiller. Always confirm which figure an "X MW" requirement refers to. IT 负荷与设施总负荷。400 kW 指服务器实际用电。计入冷却与辅助系统后， 设施总负荷约为 440–500 kW——差值由 PUE 决定，而 PUE 取决于站点气候。 *历史极端干球温度 ≤ 24 °C 的站点可纯靠干冷器运行（自然冷却，PUE 最低）； 更热的站点需加装混合冷机。遇到「X MW」的需求时，务必确认指的是哪一个口径。
The whole server — GPUs, VRMs, NICs, DIMMs — is submerged in dielectric fluid. No cold plates to match to each SKU, no fans to fail, no hot spots. Up to 50 kW per 32RU tank, with both air-cooled servers retrofitted for immersion and immersion-ready designs supported. 整台服务器——GPU、VRM、网卡、内存——全部浸没在绝缘冷却液中。 无需为每个 SKU 匹配冷板，没有会坏的风扇，也没有热点。 每个 32RU 液舱最高 50 kW，既支持风冷服务器改造浸没，也支持原生浸没设计。
Oil enters the tank at ≤ 35 °C and leaves at ≈ 43 °C; the facility loop runs at 32/37 °C. Where the historical peak dry-bulb stays ≤ 24 °C, dry coolers alone reject all heat — no compressors, PUE as low as ≈ 1.05. Hotter sites add a hybrid chiller that trims only the peaks. 冷却油以 ≤ 35 °C 进舱、约 43 °C 出舱；设施侧回路运行在 32/37 °C。 当站点历史极端干球温度 ≤ 24 °C 时，仅靠干冷器即可排走全部热量—— 无需压缩机，PUE 可低至约 1.05。更热的站点加装混合冷机，仅在峰值时补冷。
Each A32 tank carries dual CDUs in full 2N — two pumps, plate heat exchanger, and filtration per CDU, hot-swappable in minutes without draining the tank. A single pump or CDU failure never stops the servers. 每个 A32 液舱配备完全 2N 的双路 CDU——每路含两台泵、板式换热器 与过滤单元，无需排油即可在数分钟内热插拔更换。单台泵或单路 CDU 故障都不会让服务器停机。
Orbit the immersion container — tank rows, service aisle, and the oil loop — the exact unit your GPUs would live in.环绕查看浸没集装箱——液舱列、检修通道与油回路——正是你的 GPU 将要落户的那台设备。
Heavy USD is explicit download only. Naming and packaging rules: USD output spec.大体积 USD 仅提供显式下载。命名与封装规则见 USD 输出规格。
Both containers share the identical immersion core — 8× A32 tanks, the same oil loop, the same cooling rules. They differ in where the UPS lives. 两台集装箱共用完全相同的浸没核心——8 个 A32 液舱、同一条油回路、 同一套冷却规则。区别只在于 UPS 装在哪里。
| I400C40 | I400C45 | |
|---|---|---|
| Container箱体 | 40-ft40 英尺 | 45-ft (adds a dedicated power bay)45 英尺（增设独立配电舱） |
| IT CapacityIT 容量 | 360 kW rec. / 400 kW max推荐 360 kW / 最大 400 kW | 400 kW |
| UPS (EATON 9395XR-600, 600 kW)UPS（EATON 9395XR-600，600 kW） | External — customer-supplied外置 — 由客户提供 | Built-in power bay内置配电舱 |
| Battery Backup电池后备 | ~10 min (customer-supplied 2× 93LiG2)约 10 分钟（客户提供 2× 93LiG2） | ~20 min (built-in 2× 93LiG2)约 20 分钟（内置 2× 93LiG2） |
| UL ComplianceUL 合规 | — | UL compliant符合 UL |
| Best For最适合 | Lowest cost & size; sites with existing UPS成本与占地最低；已有 UPS 的站点 | Turnkey & regulated markets交钥匙交付与合规要求市场 |
Redundancy model. The container is the smallest redundancy unit: tanks are 2N on cooling internally, a single container has no internal IT redundancy, and system-level N / N+1 / 2N is built by adding containers. 冗余模型。集装箱是最小的冗余单元：液舱内部冷却为 2N， 单台集装箱内部不设 IT 冗余，系统级 N / N+1 / 2N 通过增加集装箱数量实现。
The full engineering model is live and interactive — every number computed from the same physics engine our design team uses, from tank-level oil flow to container PUE. 完整工程模型可在线交互——从液舱级油流量到整箱 PUE， 每个数字都来自我们设计团队所用的同一套物理引擎。
Walk the 8-tank layout of I400C40 and I400C45, power bay included.查看 I400C40 与 I400C45 的 8 舱布局，含配电舱。
Watch the oil and water loops move through the dual-CDU tank diagram.在双 CDU 液舱示意图中观察油路与水路的流动。
Oil flow per tank (≈ 11–12 m³/h), facility water, and the ΔT budget.单舱油流量（约 11–12 m³/h）、设施侧水量，以及 ΔT 预算。
Pick a PCIe server and a climate — get GPU count, binding constraint, and PUE band.选定 PCIe 服务器与气候条件，即可得到 GPU 数量、约束瓶颈与 PUE 区间。