智算中心 · 托管与互联 · GPU 算力服务AI data centers · Colocation · GPU cloud
从整座液冷智算中心，到单卡 GPU 云服务器，Branpont 为您设计、建设、托管并运营训练与推理所需的基础设施。From an entire liquid-cooled AI data center to a single-GPU instance, Branpont designs, builds, hosts and runs the infrastructure behind your training and inference.
以上为示例设计指标，具体以项目合同与现场条件为准。Representative design targets; final figures depend on site and contract.
三种服务形态Three delivery models
三种形态共享同一套工程标准与同一支运维团队。您可以从任意一种开始，并随业务规模灵活组合。All three run on one engineering standard and one operations team. Start with any of them and combine them as you grow.
由我们为您规划、设计并建设面向 GPU 集群的智算中心，交付可上电、可运行的设施，并可长期代为运营。We plan, design and build an AI data center for your GPU clusters, hand it over powered and running, and can operate it for you long term.
把您的服务器放进高密度、运营商中立的数据中心，配套网络互联与 7×24 现场支持。Place your own servers in high-density, carrier-neutral facilities, with interconnection and 24/7 on-site support.
无需自购硬件，按需使用 GPU 云服务器、GPU 裸金属服务器与专属训练集群。Use GPU instances, bare-metal GPU servers and dedicated training clusters without buying hardware.
智算中心（AIDC）AI data centers (AIDC)
智算中心是为 GPU 集群而设计的数据中心：更高的单柜功率、液冷散热与无损网络。我们按设计·建设·运营模式交付，每个阶段都有明确的交付物与验收标准。An AI data center is built around GPU clusters: denser racks, liquid cooling and lossless networking. We deliver it design–build–operate, with clear deliverables and acceptance criteria at every phase.
评估可获得的电力容量（MW）、能评与选址政策、网络时延与水资源，测算 TCO，制定分期上电方案。We assess available power (MW), energy-review and siting requirements, network latency and water, then build a TCO model and a phased power-on plan.
统筹高密度供配电、液冷、楼板承重与组网，按 GB 50174 A 级与 Uptime Tier III 标准设计。High-density power, liquid cooling, floor loading and network topology, designed together to GB 50174 Class A and Uptime Tier III.
以 EPC 总承包模式统一负责设计、采购与施工；预制模块化缩短建设与上电周期，按里程碑验收。EPC delivery under one contract for design, procurement and construction. Prefabricated modules shorten time to power, with acceptance at each milestone.
L1–L5 分级调试，满载假负载测试与冗余切换演练，出具验收报告。Level 1–5 commissioning, full-load testing with load banks and failover drills, closed out with an acceptance report.
7×24 NOC 与 DCIM 监控，持续优化能效；也可为政府、国企或投资方提供智算中心代运营。24/7 NOC and DCIM monitoring with ongoing efficiency tuning — or we run the entire facility for government, state-owned or investor owners.
工程能力Engineering capabilities
液冷机位单柜设计功率 40–140 kW，可部署 GB300 NVL72 等整机柜系统；风冷区与液冷区分区规划。Liquid-cooled positions designed for 40–140 kW per rack, ready for rack-scale systems such as GB300 NVL72, with separate air- and liquid-cooled zones.
以冷板式液冷（DLC）为主、浸没式液冷按需，覆盖 CDU、二次侧管网、Manifold 与漏液监测。Direct-to-chip (cold-plate) liquid cooling first, immersion cooling where it fits — CDUs, secondary loop, manifolds and leak detection.
2N UPS 或 HVDC 供电、柴发 N+1，可选储能（BESS），并为 800 VDC 架构预留升级路径。2N UPS or HVDC, N+1 generators and optional battery storage (BESS), with an upgrade path to 800 VDC.
电力模块、液冷模块与预制机房在工厂集成、现场拼装，缩短建设与上电周期。Power, cooling and data-hall modules are integrated in the factory and assembled on site, shortening time to power.
楼体不动，通过风改液、加装 CDU 或背板空调（RDHx）提升单柜功率，可分区分批实施。Raise rack density in an existing building with air-to-liquid conversion, added CDUs or rear-door heat exchangers (RDHx), zone by zone.
通过绿电直连、绿证与 PPA 提高可再生能源占比；余热回收与算电协同进一步降低能耗与排放。Direct green-power supply, green certificates and PPAs raise the renewable share; heat reuse and grid-aware scheduling cut energy use and emissions further.
功率、标准与周期为示例设计指标，具体以项目合同与现场条件为准。Power ratings, standards and timelines are representative design targets; final figures depend on site and contract.
托管与互联Colocation & interconnection
从单个机柜到整个数据机房：零售型、批发型与定制化建设（BTS）托管，配套运营商中立的网络互联。From a single rack to an entire data hall: retail, wholesale and build-to-suit colocation, with carrier-neutral interconnection.
以单机柜、机笼或包间为单位，含电力、制冷与安防，带宽按需选配。By the rack, cage or private suite, with power, cooling and security included; add bandwidth as you need it.
为 GPU 服务器与整机柜系统提供液冷机位与配套组网。Liquid-cooled positions for GPU servers and rack-scale systems, with the networking to match.
整租数据机房或模块，由您自主部署与运营，签订长期合约。Lease a whole data hall or module and run it your way, on a long-term contract.
按云厂商或大模型公司的规格专建园区或楼栋，也可只交付土建与供电（Powered Shell）。A campus or building built to your specification — or a powered shell you fit out yourself.
网络与互联Network & interconnection
运营商中立的接入让您自主选择上游线路；交叉连接、云专线与数据中心互联让训练数据在园区、公有云与站点之间高速流动。Carrier-neutral access lets you choose your upstream. Cross-connects, cloud on-ramps and DCI move training data quickly between campuses, public clouds and your sites.
单柜功率与起租规模为示例指标，具体以园区资源与合同为准。Rack densities and minimums are representative; final terms depend on the site and contract.
运维服务Managed services
远程协助、专家现场支持与驻场代维覆盖设备的全生命周期；GPU、InfiniBand 与液冷节点由专业工程师处理。Remote hands, smart hands and resident engineers cover the whole hardware lifecycle, with specialists for GPU, InfiniBand and liquid-cooled nodes.
GPU 算力服务GPU cloud
GPU 算力服务（算力租赁）让您无需自购硬件，即可获得训练与推理所需的算力：小规模任务分钟级开通，大规模训练使用专属集群，资源与计费方式按需组合。GPU cloud gives you compute for training and inference without buying hardware. Small jobs start in minutes, large runs get dedicated clusters, and you mix resources and billing to fit.
01 — 训练Training
面向大模型预训练与大规模后训练：整机独享、无损网络、托管调度，可按周、月或年预留。For pre-training and large-scale post-training: dedicated nodes, lossless networking and managed scheduling, reserved by the week, month or year.
千卡级集群以周为单位交付，可预约锁定容量。Clusters of a thousand GPUs or more are delivered in weeks; reserve ahead to lock in capacity.
02 — 推理与开发Inference & development
面向推理、微调与开发调试：分钟级开通、随开随停，按秒、按卡时或按词元（Token）计费。For inference, fine-tuning and development: ready in minutes, stop anytime, billed by the second, the GPU-hour or the token.
按量实例可转为包年包月，长期负载的成本更可预期。When a workload settles, move it to a subscription for more predictable cost.
平台能力Platform capabilities
NVIDIA 与国产 AI 加速卡统一纳管，按负载选型，降低对单一供应商的依赖。NVIDIA and domestic AI accelerators under one control plane. Choose by workload and reduce reliance on any single supplier.
通过控制台或 OpenAPI 开通与管理资源，内置监控、配额与成本分析。Provision and manage resources from the console or OpenAPI, with monitoring, quotas and cost analytics built in.
并行文件系统与 S3 兼容对象存储，支持 GPUDirect Storage 与检查点高速读写。A parallel file system plus S3-compatible object storage, with GPUDirect Storage and fast checkpointing.
断点续训、故障节点自动隔离与自愈，提升有效训练时长（goodput）。Checkpoint-restart plus automatic isolation and recovery of failed nodes, for more goodput from every run.
VPC 与多租户隔离，可选物理隔离的专属资源池。VPC and multi-tenant isolation, with physically isolated dedicated pools on request.
昇腾等国产 AI 加速卡资源池，提供框架与算子迁移适配，满足信创要求。Pools of Ascend and other domestic AI accelerators, with framework and operator porting for Xinchuang requirements.
GPU 型号、数量、可用区域（境内 / 境外）与交付周期以实际供应及出口管制政策为准。GPU models, quantities, regions (mainland China or international) and lead times depend on actual supply and export-control policy.
机型与资源Accelerators
从 PCIe 推理卡到液冷整机柜，按训练、微调或推理选择合适的 GPU 与形态。From PCIe inference cards to liquid-cooled racks, choose the GPU and form factor for training, fine-tuning or inference.
NVIDIA
H200
大模型训练与推理LLM training and inference
NVIDIA
B200
大规模训练与高吞吐推理Large-scale training and high-throughput inference
NVIDIA
GB300 NVL72
超大模型训练与推理Training and serving very large models
NVIDIA
L40S
推理、微调与图形渲染Inference, fine-tuning and graphics
AMD
Instinct MI355X
大显存训练与推理Memory-heavy training and inference
国产 AI 加速卡Domestic AI accelerators
昇腾等Ascend and others
政企与信创场景Public sector and Xinchuang workloads
规格为厂商公开参数。机型、数量与可用区域（境内 / 境外）以实际供应为准；部分型号受出口管制政策约束，境内区域仅提供可合规供应的型号。Specifications are vendors' published figures. Models, quantities and regions (mainland China or international) depend on actual supply. Some models are subject to export controls; mainland regions offer only models that can be lawfully supplied there.
技术规格Specifications
供电、制冷、网络与安全均按行业通用标准描述；设计值与实测值分开披露，可用性按月统计。Power, cooling, network and security are described against industry standards. Design and measured values are reported separately, and availability is measured monthly.
| 供电Power | |
|---|---|
| 市电接入Utility feed | 双路市电Dual utility feeds |
| UPS / HVDC | 2N UPS 或 240 V / 336 V HVDC2N UPS or 240 V / 336 V HVDC |
| 后备电源Backup generation | 柴油发电机组 N+1，后备时长按项目设计N+1 diesel generators; runtime designed per project |
| 储能Energy storage | 可选储能系统（BESS）Optional battery energy storage (BESS) |
| 制冷Cooling | |
| 液冷Liquid cooling | 冷板式液冷（DLC）为主，浸没式液冷按需Direct-to-chip (cold-plate) first; immersion on request |
| CDU | N+1 冗余N+1 redundant |
| 供液温度Supply temperature | 温水冷却，参照 ASHRAE W32–W45Warm-water cooling, per ASHRAE W32–W45 |
| 风冷Air cooling | 冷 / 热通道封闭Hot- and cold-aisle containment |
| 网络Network | |
| 公网接入Internet access | BGP 多线（电信 / 联通 / 移动），运营商中立Multi-carrier BGP (China Telecom, China Unicom, China Mobile); carrier-neutral |
| 计算网络Compute fabric | InfiniBand NDR 或 400GbE RoCEv2，单节点 8 × 400GNDR InfiniBand or 400GbE RoCEv2; 8 × 400G per node |
| 拓扑Topology | 轨道优化 + 1:1 无收敛胖树Rail-optimized, non-blocking (1:1) fat-tree |
| 互联Interconnection | 交叉连接 · 云专线 · 数据中心互联（DCI）Cross-connects · Cloud on-ramps · DCI |
| 安全Security | |
| 物理安防Physical security | 7×24 安保、多级门禁、视频监控24/7 guards, layered access control, CCTV |
| 消防Fire protection | 极早期烟雾探测 + 气体灭火Very early smoke detection with gas suppression |
| 合规Compliance | 可在等保三级环境交付Can be delivered in MLPS Level 3 environments |
| 运维Operations | |
| 监控Monitoring | 7×24 NOC、DCIM 与动环监控24/7 NOC, DCIM and environmental monitoring |
| 现场支持On-site support | 远程协助与专家现场支持Remote hands and smart hands |
| SLA | 按月统计与报告，不含计划维护Measured and reported monthly, excluding planned maintenance |
以上为示例设计指标，具体园区规格以项目合同与现场条件为准。Representative design targets; site-specific figures depend on contract and site conditions.
典型负载Reference workloads
以下为常见 AI 负载的参考配置。最终方案由解决方案架构师根据您的模型、数据与时间表评估确定。Starting points for common AI workloads. A solutions architect sizes the final setup around your models, data and timeline.
从零训练或继续预训练基础模型，重点关注线性加速比与长时间稳定运行。Training foundation models from scratch or continuing pre-training, where scaling efficiency and long, stable runs matter most.
用领域数据完成 SFT、LoRA 或 DPO，让模型更贴合您的业务。Adapt a model to your domain with SFT, LoRA or DPO on your own data.
在线与离线推理，按时延与吞吐目标做容量规划与压测。Online and batch inference, sized and load-tested against your latency and throughput targets.
大规模数据清洗与预处理，与公有云及本地系统统一管控。Large-scale data cleaning and preparation, managed together with your public cloud and on-premises systems.
参考配置仅供选型参考，不构成性能承诺；最终配置以方案评估为准。Reference configurations are a starting point, not a performance commitment; final sizing follows an assessment.
计费模式Pricing models
四种计费模式可组合使用。价格取决于机型、规模、区域与合同期限，请联系我们获取报价。Combine any of the four. Pricing depends on hardware, scale, region and term — contact us for a quote.
弹性Flexible
按卡时计费 · 随开随停Per GPU-hour · Start and stop anytime
稳定Steady
按月或按年 · 锁定资源与单价Monthly or annual · Locked resources and rate
适合大规模训练For large-scale training
1–3 年承诺 · 容量保障1–3 year commitment · Guaranteed capacity
设施Facilities
机柜·月或 kW·月Per rack or per kW, monthly
开通流程Getting started
解决方案架构师了解您的模型规模、数据量、时间表与合规要求。A solutions architect reviews your model size, data volume, timeline and compliance needs.
给出机型、网络与存储配置，以及区域、交付时间与计费方案。You get a hardware, network and storage configuration, with region, delivery date and billing options.
实例分钟级开通；千卡级集群以周为单位交付，交付前完成烤机与基准测试。Instances start in minutes. Thousand-GPU clusters are delivered in weeks, burned in and benchmarked before handover.
按约定指标验收，随后进入 7×24 运维与按月 SLA 报告。Acceptance against agreed metrics, then 24/7 operations with monthly SLA reports.
承诺期限与交付周期为示例，具体以正式报价与合同为准。Commitment terms and lead times are representative; final terms are set in your quote and contract.
常见问题FAQ
我们通过自建与合作数据中心资源提供服务。具体园区、可用机型与交付周期取决于您的需求与当期资源，请联系解决方案团队获取最新信息。
We serve customers from both our own and partner data center capacity. Specific campuses, available hardware and lead times depend on your requirements and current capacity — contact our solutions team for the latest details.
单卡或单机实例可分钟级开通；千卡级集群通常以周为单位交付，具体取决于机型供应、网络规模与验收要求。建议提前预约，锁定容量与交付窗口。集群交付前均完成烤机与 NCCL 等基准测试。
Single-GPU and single-node instances start in minutes. Thousand-GPU clusters are typically delivered in weeks, depending on hardware supply, fabric size and acceptance requirements. Reserve early to lock in capacity and a delivery window. Every cluster is burned in and benchmarked, including NCCL tests, before handover.
GPU 算力服务支持按量计费（按卡时）、包年包月与预留容量（1–3 年承诺）；推理服务可按秒计费，模型 API 按百万词元计费。托管按机柜·月或 kW·月计费，带宽与交叉连接另计。
GPU cloud is available on demand (per GPU-hour), by monthly or annual subscription, or as reserved capacity on a 1–3 year commitment. Inference endpoints can bill per second, and the model API per million tokens. Colocation is billed per rack or per kW each month, with bandwidth and cross-connects billed separately.
支持。我们提供昇腾等国产 AI 加速卡资源，并协助完成框架、算子与模型的迁移适配，满足信创要求。国产算力与 NVIDIA 资源可在同一平台统一纳管。
Yes. We offer Ascend and other domestic AI accelerators, and help port frameworks, operators and models to them for Xinchuang requirements. Domestic and NVIDIA resources are managed on the same platform.
不一定。改造可分区、分批实施，并配合负载迁移，尽量减少对在运业务的影响；是否需要短时停机取决于现场条件，我们会在现场勘察后给出具体实施方案。
Not necessarily. Retrofits can run zone by zone alongside workload migration to keep disruption to a minimum. Whether any brief downtime is needed depends on the site — we'll set out the approach after a survey.
可以。零售型机柜托管以单机柜起步，也提供高密度液冷机位；您可按需叠加 BGP 带宽、交叉连接、云专线，以及远程协助、部署上架等运维服务。
Yes. Retail colocation starts at a single rack, and high-density liquid-cooled positions are available too. Add BGP bandwidth, cross-connects, cloud on-ramps and services such as remote hands or rack-and-stack as you need them.
告诉我们您需要的机型、规模、区域与时间表，解决方案架构师将为您给出配置建议与报价。Tell us the hardware, scale, region and timeline you need, and a solutions architect will come back with a configuration and a quote.
或直接发送邮件至Or email contact@branpont.com