ASUS Ships 220kW Rubin AI Racks as Power Limits Replace GPU Shortages

Ai4

Data center constraints have shifted from GPU availability to power and cooling, driving ASUS to offer hybrid liquid and air systems that bring up 220 kW racks in 30 minutes.

AI-assisted summary of Presentation by ASUS — original source. Check the original for full context.

Source published: 2026-08-14

Detailed notes

## Background This presentation was uploaded to YouTube by Ai4 on August 14, 2026. The video features Brian Ngo, Senior Product Marketing Manager at ASUS, presenting on AI infrastructure and server solutions.

## TL;DR The presentation highlights how the primary operational bottleneck in AI infrastructure has shifted from GPU availability to power and cooling constraints. ASUS showcases its enterprise hardware portfolio and rapid turnkey deployment capabilities, which can bring full data centers online in as little as 30 minutes. The session details next-generation liquid-cooled and hybrid server architectures, including NVIDIA GB300 and Vera Rubin racks operating up to 220 kW, while also addressing traditional facilities with 4U air-cooled alternatives.

## Atmosphere & style The presentation maintains a polished, authoritative, and corporate tech-focused tone. Brian Ngo delivers a pragmatically framed, forward-looking pitch that grounds high-level hardware stats in real-world data center logistics. It feels like an insider industry briefing designed to build confidence in enterprise longevity and operational speed.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=4s) The primary bottleneck in AI infrastructure has shifted from GPU availability to power density and thermal cooling plumbing. - [00:04](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=4s) ASUS leverages over 30 years of enterprise hardware history across CSP, non-CSP enterprise, and turnkey sectors to back long-term infrastructure decisions. - [00:04](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=4s) Hyperscaler learnings from massive 72-GPU liquid-cooled deployments directly inform standalone 4U enterprise server architectures. - [02:21](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=141s) ASUS achieved early market entry by shipping mass-production NVIDIA GB300 and B300 racks as of August 2025. - [02:21](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=141s) Turnkey data center deployment pre-planning and software integration allow full GB300 bring-up in as little as 30 minutes. - [05:23](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=323s) Reasoning models that think across thousands of tokens make inference a major cost driver, forcing hardware purpose-built for reasoning token efficiency. - [05:23](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=323s) Next-generation reasoning architectures deliver up to 10x lower token costs, 10x higher performance per megawatt, and 2.8x memory bandwidth increases. - [05:23](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=323s) Power requirements jump significantly between hardware generations, moving from 135 kW per rack (GB300) to 220 kW per rack (Vera Rubin). - [07:24](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=444s) For NVIDIA HGX Rubin deployments, ASUS offers a hybrid design with air-cooled CPUs, liquid-cooled GPUs, and a cableless motherboard that cuts setup time in half. - [07:24](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=444s) Modified MGX 48U racks can accommodate 9 compute trays housing 72 Rubin NVL8 GPUs while pushing rack-scale liquid cooling limits. - [09:09](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=549s) Air-cooled 4U MGX chassis supporting eight 600W GPUs enable high-density compute deployment without immediate facility liquid cooling upgrades. - [09:09](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=549s) Data center success relies heavily on total deployment ownership—including custom room cooling tuning, CDUs, and software stack validation—rather than simple hardware delivery.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=4s) The Shifting Constraint in AI Infrastructure

Two years ago, when asking data center teams what was holding back their AI roadmap, the answer was universally a single word: "GPUs." Today, asking that same team yields a much less glamorous answer: it's power and the plumbing required to carry heat away. The operational constraint has shifted dramatically, and ASUS has engineered its portfolio specifically to address this new reality.

ASUS highlights its track record to emphasize long-term stability rather than mere vanity—noting its inclusion on Fortune's World's Most Admired Companies list ten times, ranking among the top 25 global tech brands, and standing for 12 consecutive years as a top Taiwan global brand. In the infrastructure space, a decision made in 2026 must still be supported in 2032. Operating for over 30 years across multiple hype cycles, ASUS grows its server business across three core sectors: Cloud Service Providers (CSPs), non-CSP enterprise buyers, and full turnkey solutions. This unique position allows enterprise clients to benefit from direct learnings gathered while deploying massive 72-GPU liquid-cooled racks for hyperscalers.

Operationally, there is no longer any single generic "AI workload." Training a frontier model, fine-tuning it, and serving reasoning inference at scale represent three distinct machines with entirely different economics. Rather than pushing a single hero product, ASUS offers a broad portfolio tailored to specific workloads, facility constraints, and budgets.

- **Corporate Longevity:** 10x on Fortune's World's Most Admired Companies, top 25 global tech brand, 12 years running as a top Taiwan global brand, and over 30 years of hardware shipping history. - **Three Business Sectors:** Cloud Service Providers (CSPs), non-CSP enterprise buyers, and full turnkey solutions. - **Deployment Scaling:** Architectural insights from 72-GPU liquid-cooled CSP racks directly inform standalone 4U enterprise server designs.

[02:21](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=141s) Speed to Market and Turnkey Deployment

ASUS was among the first worldwide to ship NVIDIA GB300 and B300 racks in mass production as of August 2025. In the high-performance computing market, the gap between being "announced" and being "racked and running" is where roadmaps frequently die. Being early allows ASUS engineers to test, validate, and work through initial platform issues on aggressive timelines so that customers receive a platform that works seamlessly out of the box.

The real product extends far beyond bare metal and server boxes to what happens after the delivery truck leaves the facility. ASUS manages full turnkey deployments: delivering Cooling Distribution Units (CDUs) on time, tuning airflow and cooling paths specifically to the room rather than relying on standard reference designs, validating networking throughput, securing storage infrastructure, and ensuring the software management stack is fully running before handover. For customers ordering GB300 systems, ASUS software and pre-planning enable full data center bring-up in as little as 30 minutes—a process that typically takes competitors days or weeks.

``` +-------------------------------------------------------------------+ | ASUS Deployment Pipeline | +-------------------------------------------------------------------+ | 1. Pre-Planning & Custom Design | | - Airflow/Cooling Path Tuning to Physical Room Space | | - Hardware & Software Stack Integration | +-------------------------------------------------------------------+ | v | 2. On-Site Hardware & Cooling Delivery | | - Cooling Distribution Units (CDUs) Delivered | | - Bare Metal Racks Placed | +-------------------------------------------------------------------+ | v | 3. Validation & Software Handover | | - Network Throughput & Storage Infrastructure Secured | | - Management Software Stack Pre-Configured | | - Full Data Center Bring-Up in 30 Minutes | +-------------------------------------------------------------------+ ```

- **Production Timeline:** GB300 and B300 racks entered mass production shipping in August 2025. - **Data Center Bring-Up:** Turnkey setup completed in 30 minutes compared to traditional industry timelines of days or weeks.

[05:23](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=323s) Reasoning Models and Next-Gen Architecture Performance

The underlying nature of AI workloads has shifted due to reasoning models that do not merely generate immediate answers but "think" across thousands of tokens per query. This shift transforms inference from a minor operational expense into a dominant line item on cost sheets. Silicon is now purpose-built for reasoning rather than pure training throughput.

Compared to the NVIDIA Blackwell generation, next-generation architectures deliver massive jumps in key metrics: - **Token Cost Efficiency:** Up to 10x improvement in inferencing token cost. - **Power Efficiency:** Up to 10x performance per megawatt. - **Bandwidth:** Up to 2.8x increase in memory bandwidth.

At the full rack level, ASUS offers systems spanning current and upcoming architectural generations: - **NVIDIA GB300 Rack:** Includes 36 Grace CPUs and 72 B300 GPUs, drawing roughly 135 kW per rack, and is currently shipping. - **NVIDIA Vera Rubin Rack:** Features 36 Vera CPUs and 72 Rubin GPUs, operating at 220 kW per rack. It is entering mass production now and will ship next year.

[07:24](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=444s) Flexible Form Factors: Reference vs. Hybrid Cooling

To support NVIDIA HGX Rubin deployments, ASUS provides two distinct hardware paths based on facility constraints and risk profiles:

``` +-------------------------------------------------------------------+ | NVIDIA HGX Rubin System Options | +-------------------------------------------------------------------+ | 1. NVIDIA Reference Design | 2. ASUS Hybrid Design | | -------------------------- | --------------------- | | - Fully Liquid-Cooled | - Hybrid (Air CPU / Liq GPU) | | - 2U Compact Form Factor | - Cableless Board Design | | - Optimized for Power/PUE | - 2x Faster Setup Time | | - Reduced Integration Risk | - Lower Capex / Fast TTM | +-------------------------------------------------------------------+ ```

For engineers planning rack-scale infrastructure, ASUS offers a modified MGX 48U rack unit accommodating 9 compute trays and 72 Rubin NVL8 GPUs. The power density requirement jumps from 135 kW to 220 kW in a single hardware generation, highlighting the critical need to solve power and thermal constraints early in the facility design phase.

- **NVIDIA Reference Design:** 2U form factor, fully liquid-cooled, optimized for strict PUE and power density constraints. - **ASUS Design:** Hybrid cooling (air-cooled CPU combined with liquid-cooled GPU), completely cableless board design, providing double the setup speed. - **Rack Architecture:** Modified MGX 48U rack containing 9 compute trays and 72 Rubin NVL8 GPUs operating at up to 220 kW.

[09:09](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=549s) Air-Cooled Alternatives and Core Takeaways

For facilities where rack-scale liquid cooling is not currently feasible, ASUS offers three distinct 4U NVIDIA MGX air-cooled systems. These air-cooled chassis support eight 600W double-width GPUs each and are currently shipping. They allow data centers to deploy high-density compute immediately using traditional cooling infrastructure, with flexible processor options spanning Intel, NVIDIA, and AMD platforms.

Three central takeaways summarize the current state of AI infrastructure deployment: 1. **Power is the primary constraint:** Moving from 135 kW to 220 kW per rack in a single generation requires proactive fleet and facility planning. 2. **Flexible cooling architectures:** Organizations do not have to choose a single lane—they can deploy 4U air-cooled systems today while building out full liquid-cooled racks for future requirements. 3. **Integration determines success:** "The GPUs are the same everywhere. What actually differs is whether your partner owns the deployment or just the shipment."

- **4U Air-Cooled Systems:** Supports 8 x 600W double-width GPUs per chassis with CPU offerings across Intel, AMD, and NVIDIA. - **Power Scale:** Rapid jump from 135 kW to 220 kW per rack across consecutive hardware generations.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=4s) GPUs. — Brian Ngo

> [00:04](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=4s) it's power and the plumbing required to carry heat away. — Brian Ngo

> [02:21](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=141s) racked and running — Brian Ngo

> [05:23](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=323s) think — Brian Ngo

> [09:09](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=549s) Power is the primary constraint — Brian Ngo

> [09:09](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=549s) Flexible cooling architectures — Brian Ngo

> [09:09](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=549s) Integration determines success — Brian Ngo

> [09:09](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=549s) The GPUs are the same everywhere. What actually differs is whether your partner owns the deployment or just the shipment. — Brian Ngo

## Mentioned ASUS, Brian Ngo, Fortune, Cloud Service Providers, NVIDIA, NVIDIA GB300, NVIDIA B300, NVIDIA Blackwell, NVIDIA Vera Rubin, NVIDIA HGX Rubin, NVIDIA MGX, Grace CPU, Vera CPU, Rubin GPU, Rubin NVL8, Intel, AMD

## Video description Brian Ngo - Senior Product Marketing Manager, ASUS

[Source](https://www.youtube.com/watch?v=8xKpjgf2Q2c)

中文简报

电力限制取代GPU短缺 ASUS出货220kW Rubin AI机架数据中心的瓶颈已从GPU供应转向电力与散热，促使ASUS推出风液混合系统，可在30分钟内启动运行220 kW机架。本次演讲重点指出AI基础设施的核心运营瓶颈已从GPU可用性转变为电力和散热限制。ASUS展示了其企业级硬件产品组合与快速一站式交付能力，最快可在30分钟内使整个数据中心上线运行。内容详细介绍了下一代液冷及混合服务器架构，包括运行功率最高可达220 kW的NVIDIA GB300与Vera Rubin机架，同时也针对传统设施提供了4U风冷替代方案。 AI 基础设施的核心瓶颈已从 GPU 供应转向电力密度与散热管路，ASUS 结合 30 余年硬件积淀，将超大规模客户的 72-GPU 液冷实践深度融入企业级产品架构中。 (0:04) ASUS 于 2025 年 8 月实现 NVIDIA GB300 与 B300 机架的量产出货，通过交钥匙级全流程规划与软件预集成，使整套数据中心启动上线时间缩短至最快 30 分钟。 (2:21) 面对复杂推理模型带来的高昂成本，下一代架构可实现高达 10 倍的 Token 成本降幅及每兆瓦算力提升；同时单机架电力需求也从 GB300 的 135 kW 激增至 Vera Rubin 的 220 kW。 (5:23) 针对不同机房条件，ASUS 既提供支持 220 kW 的 Rubin 2U 全液冷与安装速度快 2 倍的无线缆混合散热方案，也支持搭载 8 颗 600W GPU 的 4U 传统风冷服务器灵活过渡。 (7:24)

Reading summary

ASUS argues that the primary operational bottleneck in AI data centers has shifted from GPU availability to power capacity and heat dissipation, driven by reasoning workloads that dramatically elevate rack power density. The Shift to Power Constraints and Reasoning Workloads (0:04) Whereas AI roadmaps were previously constrained by GPU chip availability, infrastructure operators now face facility limitations in power delivery and thermal extraction. This operational constraint is driven largely by the rise of reasoning models, which generate thousands of tokens per query to process answers, transforming inference into a dominant operational cost. According to ASUS, next-generation architectures designed for reasoning deliver up to a 10x improvement in inferencing token cost, a 10x increase in performance per megawatt, and a 2.8x expansion in memory bandwidth compared to the NVIDIA Blackwell generation. Hardware power requirements reflect this escalation in energy density. The currently shipping NVIDIA GB300 rack—integrating 36 Grace CPUs and 72 B300 GPUs—operates at roughly 135 kW per rack. Next-generation NVIDIA Vera Rubin racks, which pair 36 Vera CPUs with 72 Rubin GPUs, are entering mass production and will demand 220 kW per rack when shipping next year. Turnkey Deployment and Rapid Facility Bring-Up (2:21) To minimize the lag between hardware delivery and operational deployment, ASUS emphasizes turnkey facility integration. After beginning mass production shipments of NVIDIA GB300 and B300 racks in August 2025, the vendor structured its offerings around end-to-end rollouts, handling Cooling Distribution Unit (CDU) delivery, facility-specific airflow tuning, network throughput validation, and storage setup. ASUS claims that its pre-planning and software management stack allow clients to complete full data center bring-up in as little as 30 minutes, compared to traditional timelines of days or weeks. The company applies operational insights gathered from deploying 72-GPU liquid-cooled racks for hyperscale cloud service providers to standalone infrastructure across its non-CSP and turnkey customer bases. Cooling Flexibility Across Reference, Hybrid, and Air Systems (7:24) To support varied data center infrastructures, ASUS provides divergent hardware paths for high-density architectures. For NVIDIA HGX Rubin deployments, facilities can choose between a standard 2U fully liquid-cooled NVIDIA reference design optimized for power usage effectiveness (PUE) and power density, or an ASUS hybrid configuration that pairs an air-cooled CPU with a liquid-cooled GPU. ASUS asserts that its hybrid model uses a cableless board design to halve installation time and reduce capital expenditure in modified 48U MGX racks supporting 72 Rubin NVL8 GPUs at 220 kW. For facilities unable to accommodate liquid cooling, ASUS ships three 4U NVIDIA MGX air-cooled chassis options that support eight 600W double-width GPUs. These systems run on Intel, AMD, or NVIDIA processors, allowing operators to deploy compute density immediately using traditional cooling infrastructure while preparing for future liquid conversions. ASUS concludes that because silicon components are largely identical across hardware vendors, competitive differentiation depends on end-to-end deployment execution rather than component delivery alone. However, managing the steep power rise from 135 kW to 220 kW per rack presents an ongoing architectural challenge that requires operators to balance short-term air cooling with long-term liquid infrastructure planning.

精炼阅读摘要

华硕（ASUS）指出，当前人工智能基础设施建设的主要约束已从单纯获取GPU转变为应对电力供应和散热管道的物理限制；为此，华硕围绕不同工作负载与机房条件，推出涵盖风冷、混合液冷及交钥匙交付的多元产品线与服务架构。基础设施瓶颈转向电力与散热设施 (0:04) 华硕分析认为，数据中心团队的运作瓶颈已发生根本转变，两年前普遍面临的GPU短缺问题，如今已被供电能力与散热管道设施的物理约束所取代。同时，业界不再存在单一通用的“AI工作负载”，前沿基础模型训练、模型微调以及大规模推理服务分别对应着硬件需求与经济模型完全不同的计算设备。华硕据此放弃单一明星产品策略，转向按需定制的多元硬件组合。依托超过30年的硬件制造经验，华硕将其服务器业务划分为云服务提供商（CSP）、非CSP企业采购客户以及完整交钥匙解决方案三大板块。华硕强调，为超大规模客户部署72颗GPU液冷机架所积累的架构经验，能够直接转化并反哺于面向普通企业客户的单机4U服务器设计中，以确保客户基础设施决策具备跨周期的长期支持能力。交钥匙交付与数据中心快速上线 (2:21) 在高性能计算领域，硬件从发布到实际部署运行之间的延误常导致技术规划受阻。华硕于2025年8月开始量产出货NVIDIA GB300与B300机架，并指出提早进入量产使工程师能够在紧张的时间周期内完成初始平台测试与故障排查，从而向客户交付即开即用的成熟平台。华硕主张其服务核心延伸至硬件送达机房后的全流程交钥匙部署。该流程包括按时交付冷却分配单元（CDU）、针对机房特定空间调校气流与散热路径而非单纯依赖标准参考设计、现场验证网络吞吐能力、加固存储设施以及预置管理软件栈。华硕声称，借助前期的深度规划与软件协同，客户订购GB300系统后整座数据中心的上线调试时间可缩短至30分钟，而传统行业交付周期通常需要数天或数周。推理算力跃升与单机架功耗剧增 (5:23) 新型推理模型的运行机制要求在单次查询中生成数千个Token以完成思维链推导，导致推理成本在整体财务支出中的比重显著上升，芯片设计也由以往偏向训练吞吐量转为专门面向推理进行优化。华硕援引数据指出，相较于上一代NVIDIA Blackwell架构，下一代计算平台在推理Token成本效益和每兆瓦性能指标上均提升高达10倍，同时内存带宽提升达2.8倍。在机架级系统方面，当前正在出货的NVIDIA GB300机架配备36颗Grace CPU与72颗B300 GPU，单机架功耗约为135 kW；而处于量产阶段并计划于次年交付的下一代NVIDIA Vera Rubin机架配备36颗Vera CPU与72颗Rubin GPU，其功耗预计激增至220 kW。单代架构间的大幅功耗跨越，要求数据中心必须在机房规划初期解决配电和热负荷问题。兼顾液冷演进与风冷兼容的灵活架构 (7:24) 针对NVIDIA HGX Rubin的落地部署，华硕提供了两种不同风险与设施匹配的硬件路线：一种是2U全液冷NVIDIA参考设计，专门针对严苛的PUE和功率密度指标进行优化；另一种是华硕自研的混合冷却设计，采用CPU风冷搭配GPU液冷，并引入全无电缆主板设计，官方称其可将部署速度提升一倍，同时降低资本支出并缩短上市时间。针对机架级工程，华硕提供支持9个计算托盘与72颗Rubin NVL8 GPU的改款MGX 48U机架，以承载高达220 kW的功率密度。对于现阶段不具备整机架液冷改造条件的数据中心，华硕同步推出了三款支持8颗600W双宽GPU的4U NVIDIA MGX风冷服务器，目前已现货出货。该系列风冷系统支持Intel、AMD以及NVIDIA处理器，允许企业在沿用现有风冷机房设施的同时快速部署高密度算力，实现从传统散热到未来全液冷机架的平滑过渡。华硕总结认为，在各家硬件平台底层GPU高度趋同的背景下，基础设施供应商的实际竞争差异在于能否主导全流程的工程交付与机房落地，而非仅仅负责设备运输；而单机架功耗从135 kW向220 kW的快速跃升，也对数据中心长期的供电与散热协同规划提出了不可规避的工程挑战。

Detailed notes — zh

## Background 本内容由频道 Ai4 于 2026 年 8 月 14 日发布在 YouTube 平台。页面简介显示，本次演讲由华硕（ASUS）资深产品营销经理 Brian Ngo 主讲，重点展示华硕在 AI 基础设施领域的最新布局与解决方案。 ## TL;DR 华硕在本次演讲中剖析了当前 AI 算力瓶颈从单纯的 GPU 供应向机房供电与散热水务的转移。针对这一挑战，华硕展示了其从成熟出货的 GB300 NVL72（135 kW）跨越至次世代 Vera Rubin NVL72（220 kW）的超级机架产品演进，并提供 NVIDIA 全液冷参考设计、华硕自主混合散热以及兼容存量机房的 4U 风冷 MGX 系统等多轨方案。主讲人强调，在硬件趋于同质化的时代，超越裸机销售的端到端软硬件集成与现场快速部署能力，才是保障 AI 路线图落地的真正核心壁垒。 ## Atmosphere & style 整场演讲呈现出务实、沉稳且极具工程实战导向的专业氛围。主讲人剥离了浮夸的 AI 概念炒作，直击数据中心电力与水务散热等现实痛点，以条理清晰的技术对比与犀利的行业金句直击决策核心。整场分享节奏紧凑有力，兼具对底层物理边界的冷静敬畏与对交付落地能力的笃定自信，给观众带来扎实、可靠的技术信任感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=4s) AI 算力基础设施的核心瓶颈已从单纯的 GPU 供应，转移至机房电力配给与散热水务管网的物理限制。 - [00:04](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=4s) 华硕横跨 CSP 云服务商与企业级市场的业务布局，使其能将超算液冷经验直接沉淀并应用于标准企业服务器中。 - [00:04](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=4s) AI 负载已细分为模型预训练、微调与大规模推理三种不同场景，需要多元产品组合而非单一通用机型来匹配。 - [02:21](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=141s) 率先实现规模化量产的核心价值在于提前验证并排除工程缺陷，消除从发布到实际上架运转的落地鸿沟。 - [02:21](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=141s) 涵盖 CDU 部署、现场定制风道与软硬件协同调优的全栈交付方案，可将机架点亮上线时间压缩至 30 分钟。 - [05:23](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=323s) 推理模型转向深度思考链导致单次查询 Token 消耗激增，推理已成为数据中心算力支出的主导性成本。 - [05:23](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=323s) 新一代计算架构在单查询 Token 成本效率和每兆瓦能效比上提升最高达 10 倍，内存带宽提升达 2.8 倍以打破内存墙。 - [05:23](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=323s) 单机架功耗从 GB300 NVL72 的 135 kW 激增至 Vera Rubin NVL72 的 220 kW，供电规划已成不可逾越的硬约束。 - [07:24](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=444s) 针对 HGX Rubin 提供纯液冷（追求极致 PUE）与混合散热（CPU 风冷+GPU 液冷，加速建置并控制成本）两条落地路径。 - [07:24](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=444s) 针对无法进行全机架液冷改造的存量数据中心，4U MGX 风冷方案支持 8 块 600W GPU 并兼容多平台异构算力。 - [09:50](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=590s) 算力硬件在物理层面呈现同质化趋势，端到端的工程集成与生产级稳定交付能力才是硬件厂商的核心差异化壁垒。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=4s) 算力瓶颈的转移与华硕的技术积淀演讲伊始，主讲人抛出了一个鲜明的行业对比：两年前如果去问数据中心团队推进 AI 路线图的最大阻碍是什么，对方总会脱口而出同一个词——GPU。然而今天再去问同一个团队，得到的回答往往褪去了高科技的光环，变得极其现实而接地气：“是电力，以及带走热量的管道水务（power and the plumbing that carries the heat away）”。算力基础设施的核心瓶颈已经发生转移，如何应对供电与散热的严苛挑战，正是华硕在硬件与架构层面持续攻坚的核心方向。在展示产品前，主讲人花半分钟梳理了华硕的技术底蕴与商业信誉：华硕十次入选《财富》杂志“全球最受赞赏公司”，跻身全球最佳科技品牌前 25 强，并连续 12 年蝉联台湾国际品牌榜首。这些荣誉背后的核心意义不在于虚名，而在于“长治久安的可靠性（longevity）”。企业在 2026 年做出的基础设施决策，必须确保在 2032 年仍能获得全面的支持与保障。华硕在硬件制造领域历经 30 多年风雨与多轮行业周期，其服务器业务横跨三大板块：云服务提供商（CSP）、非 CSP 企业级客户以及全栈解决方案。这种业务结构让华硕兼收并蓄——只服务超大规模数据中心的厂商往往只懂超级架构，而只做传统企业的厂商又无法接触技术前沿；华硕恰恰横跨两界，能将为 CSP 部署 72 颗 GPU 液冷机架积累的实战经验，直接沉淀并融入到面向企业机房的 4U 服务器当中。主讲人强调，高喊“All in AI”十分轻巧，但落地到运营层面必须面对现实：行业里早已不存在通用的“单一 AI 工作负载”。训练前沿大模型、微调模型以及大规模推理服务，在硬件特性和经济模型上完全是三种截然不同的机器。华硕并未押注于某一款“单打冠军”产品，而是提供完整的产品组合，旨在帮助客户在工作负载、机房物理条件与预算之间找到最佳平衡。 - 华硕具备逾 30 年硬件研发交付经验，确保 IT 投资具备跨越周期的长期支持能力。 - 业务覆盖云服务商、非 CSP 企业与完整解决方案，实现顶尖液冷超算经验向标准企业级硬件的直接转化。 - AI 工作负载已精细分化为前沿预训练、微调与大规模推理，必须以多元化产品矩阵精准匹配不同机房条件与财务模型。 --- ### [02:21](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=141s) 现货量产优势与全栈交付部署在交付时效上，华硕早在去年 8 月便已成为全球首批实现 GB300 和 B300 机架大规模量产出货的厂商之一。主讲人直言不讳地指出：“在当前市场中，从产品发布到实际上架运转之间的鸿沟，正是路线图的折戟之地（the gap between announced and racked and running is where road maps go to die）。”率先量产出货并非为了炫耀，其真正价值在于华硕的工程师已经在别人的交付时间表上提前完成了平台的测试、验证，并排除了所有设计缺陷；客户最终拿到手的，必然是成熟可靠的稳定版本。然而，硬件与裸机机架只是交付的基础，“货车卸货离开之后的事情”才是见真章的关键。华硕提供的不仅是单纯的机柜，更涵盖了全套工程级实施方案：冷却液分配单元（CDU）准时入场并调试到位；根据客户现场机房的具体空间定制气流组织与散热路径，而非生搬硬套公版参考设计；网络吞吐能力逐一验证，存储架构全面加固，整套软硬件管理平台在移交机架钥匙前就已在实际负载下稳定运行。对于算力部署而言，这套定制化的实施全流程才是核心产品。市场上任何人都能卖 GPU，但提供端到端生产级系统集成的厂商凤毛麟角。凭借预先规划与软硬件无缝协作，部分订购 GB300 的客户在数据中心现场仅耗时 30 分钟便完成了整套系统的点亮上线；而在业内，这一复杂的调测过程通常需要花费数天乃至数周时间。 - 率先实现 GB300/B300 量产出货，在前期代际磨合中消化工程试错成本，交付即用产品。 - 提供超越裸机的完整现场交付：CDU 按期部署、针对现场房间定制风道热流、网络/存储安全调优与管理软件预装。 - 软硬件深度协同的标准化方案，将传统需数天至数周的数据中心点亮周期大幅压缩至 30 分钟。 --- ### [05:23](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=323s) 推理时代来临：GB300 与 Vera Rubin 机架方案演进随着大模型技术演进，算力底座面临的负载性质发生了根本性重塑。当下的推理模型不再只是直接吐出答案，而是具备深度的逻辑思考过程，单个查询往往需要消耗数千个 Token。这导致推理环节从过去的边缘辅助开销，演变成了企业账单上占据统治地位的核心成本项。面对推理算力需求的爆发，芯片的底层设计理念也必须从单纯追求训练吞吐量转向专为推理优化。相较于前代 NVIDIA Blackwell 平台，新一代架构在关键性能指标上迎来了跨越式提升： - 每次推理查询的 Token 成本效率最高提升至 10 倍。 - 每兆瓦（MW）电力消耗带来的算力性能最高提升至 10 倍。 - 内存带宽提升达 2.8 倍，有效消除推理长文本与思考链时的内存墙瓶颈。在全机架级解决方案中，华硕目前正规模化出货 GB300 NVL72 机架，该系统整合了 36 颗 Grace CPU 与 72 颗 B300 GPU，单机架设计功耗约为 135 kW。与此同时，华硕正在推进下一代基于 Vera Rubin 架构的超级机架量产，整机集成 36 颗 Vera CPU 与 72 颗 Rubin GPU，单机架功耗激增至 220 kW，目前已进入量产阶段并计划于明年全面出货。 ``` [ 机架级演进概览 ] +--------------------------------------------------------------------+ | GB300 NVL72 (现已出货) | | 架构: 36x Grace CPU + 72x B300 GPU | | 功耗: 约 135 kW / 机架 | +--------------------------------------------------------------------+ │ ▼ (代际跨越: 功耗与密度跃迁) +--------------------------------------------------------------------+ | Vera Rubin NVL72 (正步入量产，次年发货) | | 架构: 36x Vera CPU + 72x Rubin GPU | | 功耗: 约 220 kW / 机架 | +--------------------------------------------------------------------+ ``` --- ### [07:24](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=444s) 散热架构选择与风冷/液冷产品线布局针对 NVIDIA HGX Rubin 系统，华硕提供了两条并行的采购路径。主讲人指出这两种架构并无绝对的优劣之分，而是针对不同的机房物理边界量身定制： 1. **NVIDIA 官方参考设计**：采用全液冷方案与更紧凑的 2U 外形规格，经过 NVIDIA 官方全套认证，集成风险低，特别适用于机房供电与部署密度（PUE 与空间约束）极为苛刻的现代化数据中心场景。 2. **华硕自主混合散热设计（Hybrid Cooling）**：采用“CPU 风冷 + GPU 液冷”的解耦方案。该设计兼顾了实用性与交付速度，使现场搭建速度提高一倍。主板均采用全免线缆设计（cableless board），是希望规避过高 Capex、快速实现业务上线的折中利器。两者的决策逻辑非常直接：若首要考核指标是极限电力利用率（PUE）与极高密度，应选择纯液冷参考设计；若首要考量是投资回报周期与上市速度（Time to Market），混合散热设计则是更务实的选择。针对机房工程师群体，主讲人详细介绍了基于上述架构演进的修改版 MGX 48U 机架：机柜内部署了 9 组计算托盘（Compute Trays），总计搭载 72 颗 Rubin NVL8 GPU，机架整体功耗承载高达 220 kW。单机架功耗从 135 kW 飙升至 220 kW 的事实再次印证：数据中心在规划机架设计时，电力供应与配电拓扑是必须前置判定的生死线。对于目前尚不具备机架级液冷改造条件的企业机房，华硕提供了现成的风冷过渡方案。华硕旗下的三款 4U NVIDIA MGX 系统现已现货出货，每台支持 8 张 600W 双宽 GPU，完全采用风冷散热，能够直接部署在现有数据中心机架中。系统支持灵活的异构硬件配置，全面兼容 Intel、NVIDIA 和 AMD 的多重算力组合。 - HGX Rubin 纯液冷版：2U 高度，全液冷散热，降低官方验证风险，适配极致能效比与高密机房。 - HGX Rubin 混合冷却版：CPU 风冷结合 GPU 液冷，免线缆内部连接，建置速度加快 1 倍，显著降低部署门槛与初始投资。 - 存量机房风冷方案：4U NVIDIA MGX 现货供应，支持 8 块 600W 双宽 GPU，兼容 Intel、AMD 与 NVIDIA 平台，免除液冷改造包袱。 --- ### [09:50](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=590s) 总结与三大核心启示在演讲结尾，主讲人总结了基础设施决策者必须掌握的三条关键认知：第一，**电力是不可逾越的核心约束**。短短一个产品代际内，机架功耗就从 135 kW 飙升至 220 kW。这种指数级的供能跃升要求企业在规划算力集群和物理机房基础设施时，必须提前数年进行长远规划与电力储备。第二，**不必在单一技术路径上孤注一掷**。企业现阶段无需勉强做非黑即白的选择，完全可以采取务实的混合演进策略——今天为现有业务部署即开即用的 4U 风冷 MGX 系统，同时为明天更长远的大模型需求储备全机架液冷数据中心。第三，也是主讲人最为着重强调的一点：**“集成才是最难的部分（Integration is the hard part）”**。在当下的算力市场中，各家拿到的 GPU 芯片并无本质区别，算力硬件在物理层面是同质化的；真正的核心差异在于，你的硬件合作伙伴究竟是仅仅负责把包装箱发到你的门口，还是能够全面接管后续复杂的部署工程，确保整套先进系统无缝、稳健地在生产环境中真正运转起来。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=4s) 是电力，以及带走热量的管道水务（power and the plumbing that carries the heat away） — Brian Ngo > [00:04](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=4s) 长治久安的可靠性（longevity） — Brian Ngo > [02:21](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=141s) 在当前市场中，从产品发布到实际上架运转之间的鸿沟，正是路线图的折戟之地（the gap between announced and racked and running is where road maps go to die） — Brian Ngo > [02:21](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=141s) 货车卸货离开之后的事情 — Brian Ngo > [09:50](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=590s) 电力是不可逾越的核心约束 — Brian Ngo > [09:50](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=590s) 不必在单一技术路径上孤注一掷 — Brian Ngo > [09:50](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=590s) 集成才是最难的部分（Integration is the hard part） — Brian Ngo > [00:04](https://www.youtube.com/watch?v=8xKpjgf2Q2c&t=4s) 单一 AI 工作负载 — Brian Ngo ## Mentioned ASUS (华硕), Brian Ngo, Ai4, NVIDIA, NVIDIA Blackwell, NVIDIA MGX, GB300, B300, GB300 NVL72, Vera Rubin, Vera Rubin NVL72, HGX Rubin, Grace CPU, B300 GPU, Vera CPU, Rubin GPU, Rubin NVL8, Intel, AMD, 《财富》杂志 (Fortune) ## Video description Brian Ngo - Senior Product Marketing Manager, ASUS [Source](https://www.youtube.com/watch?v=8xKpjgf2Q2c)

Source quotes

[0:04] Brian Ngo: 是电力，以及带走热量的管道水务（power and the plumbing that carries the heat away） [0:04] Brian Ngo: 长治久安的可靠性（longevity） [2:21] Brian Ngo: 在当前市场中，从产品发布到实际上架运转之间的鸿沟，正是路线图的折戟之地（the gap between announced and racked and running is where road maps go to die） [2:21] Brian Ngo: 货车卸货离开之后的事情 [9:50] Brian Ngo: 电力是不可逾越的核心约束 [9:50] Brian Ngo: 不必在单一技术路径上孤注一掷 [9:50] Brian Ngo: 集成才是最难的部分（Integration is the hard part） [0:04] Brian Ngo: 单一 AI 工作负载

Source quotes

[0:04] Brian Ngo: GPUs. [0:04] Brian Ngo: it's power and the plumbing required to carry heat away. [2:21] Brian Ngo: racked and running [5:23] Brian Ngo: think [9:09] Brian Ngo: Power is the primary constraint [9:09] Brian Ngo: Flexible cooling architectures [9:09] Brian Ngo: Integration determines success [9:09] Brian Ngo: The GPUs are the same everywhere. What actually differs is whether your partner owns the deployment or just the shipment.

Original transcript

[0:04] Yeah. So, two years ago when I asked that data center team uh what was holding back their AI road [0:09] map, the answer was always the same word, GPUs. [snorts] Ask that same team today and you get a [0:16] much less uh glamorous answer. It's a power and the plumbing that carries the heat away. [0:23] That's the story of the next uh 10 minutes. The constraint moved. I want to show you where it [0:28] moved to and uh what we have built at ASUS to stay ahead of it. Uh so 30 seconds of our credentials [0:36] then I won't mention it again. So 10 times on the fortune's world's most admired companies [0:43] uh top 25 of the world's uh best tech brands and 12 years running as a top Taiwan global brand. [0:51] The reason this matters is not vanity, it's longevity. The infrastructure decision you [0:57] make in 2026 has to still be supported in 2032. We have been shipping hardware for over 30 years [1:06] through uh several hype cycles and we are still here. Our server business grows on three different [1:13] sectors. The cloud service providers, the nonCSP enterprise buyers and full solutions. I mention [1:20] this mix because it changes what you get. A vendor who only sells to hyperscalers builds for hypers [1:28] scale problems. A vendor who only sells enterprise never sees the frontier. We sit in both. [1:35] So what we learned deploying 72 GPU liquid cooled racks for a CSP shows up in a 4U box that lands in [1:44] your building. So all in AI is an easy thing to say. Here is what it means operationally. [1:51] There's no such thing as an AI workload anymore. Training a frontier model, fine-tuning it, and [1:58] serving reasoning inference at scale are three different machines with three different economics. [2:05] So, we do not lead with just one hero product, but we lead with a full portfolio. [2:10] And the honest question is which one fits your workload, your building, and your budget? [2:21] So, here's a claim you can go out and check. Uh, ASUS was among the first in the world to ship uh [2:27] GB300 and B300 uh racks uh in mass production as of uh August of last year. In this market, [2:36] the gap between announced and racked and running is where road maps go to die. Being early is [2:42] not a brag. It means our engineers have already tested, validated and made the mistakes on this [2:47] platform on someone else's schedule. But the version you get is uh the version that works. [3:00] Everything so far has been about the boxes, the bare metal, the racks. This slide [3:06] is about what happens after the truck leaves. Cooling distribution units delivered on time, [3:12] air flow and cooling paths tuned to your room, not just a reference design. And networking [3:18] throughput validated, storage infrastructure secured, and the management software stack [3:25] actually running before we hand over the keys to your rack. Once the design is finalized, our [3:32] team delivers a customized implementation plan and manages every step of the deployment. [3:39] Honestly, I would argue this is the real product. Anyone could sell you a GPU, but not [3:45] the full uh production setup that we offer. So, uh some of our customers ordered the GB300 and with [3:53] our software solution, we were able to get them up and running in 30 minutes at a data center. [3:59] Uh this problem usually takes days to weeks for most, but we're able to get you up and running in [4:05] 30 minutes by uh pre-planning everything and uh yeah, software and hardware seamlessly. [4:15] Okay, here I'll play a video. [4:17] [music] [4:24] [music] [4:31] [music] [4:38] Heat. Heat. [music] [4:56] [music] [5:03] [music] [5:10] [music] [5:17] [music] [5:23] So the workload changed underneath us. Reasoning models don't just answer now. They think sometimes [5:30] thousands of tokens per query. U that turns inference from a cheap afterthought into a [5:37] dominant line on your cost sheet. The performance numbers on this slide are measured against Nvidia [5:43] Blackwell. The reason the generational gain is this steep is that the silicon is now designed [5:50] for reasoning not just training throughput. So you get up to 10 times uh token cost inferencing uh [5:58] 10 times performance per megawatt and uh 2.8 times uh memory bandwidth as well. [6:09] Yeah. So these are the full uh rack level solutions. On the right hand side is the GB300 [6:17] uh shipping now. It'll be uh 32 36 grace CPU 7200 B300 GPUs at roughly 135 kilowatts per [6:28] rack and is shipping now. The Vera Rubin on the right is 36 Vera CPUs, 72 Reuben GPUs, [6:37] and 220 kilowatts. It's entering mass production right now and will be shipping next year. [6:50] So another video here. [6:52] [music] [7:02] [music] [7:13] [music] [7:24] So, there are two paths to buy the Nvidia HGX Reuben from us. And I'm going to stand [7:30] here and tell you uh neither of them is better. They're for different use cases. So, on the left, [7:37] we'll have the Nvidia reference design. Uh this is Nvidia validated, so reduce integration risk. [7:44] It's a smaller 2U form factor with uh when power and density are your main constraints. [7:51] On the right is our design. It will be a hybrid cooling design. So offers a air cooled CPU [7:57] and then a liquid cooled GPU. So this flexible option gives you a more uh practical and [8:06] uh twice as fast uh setup time as well. Uh both are fully cableless board. So you choose based [8:13] on your constraint. If it's power and PUE, you'll go with the liquid cooled. If time to market and [8:20] capex is your wrist, then you'll go with the hybrid uh liquid cooled air cooled version. [8:28] Yeah. So this is the full uh rack level design based on this uh for the engineers in the room. So [8:36] this is a modified MGX 48 rack unit uh with nine compute trays in 72 uh Reuben NVL8 uh GPUs here. [8:48] So the key is the this will use a 220 kilowatt uh power uh [8:54] constraint here as well. So 135 to 220 kilowatt and one generation. [9:00] So power is something that uh you have to decide when you're designing your racks here. [9:09] And if rack scale liquid cooling is not happening at your facility currently this year, [9:14] then this is your honest answer right here. We have three for you uh Nvidia MGX systems. [9:21] Uh they support eight 600 watt uh double width uh GPUs in each and they're all shipping now. [9:28] So this is fully air cooled and you can uh set it up in your data centers or [9:32] rack level uh solutions already right now. We'll have a mix of flavors depending on your CPU [9:39] uh GPUs uh you want to put in but uh yeah we have the full offering through Intel, Nvidia and AMD. [9:50] All right, last slide. Uh so three things I want you to take away. One, power is the constraint. [9:56] So 135 kilowatt to 20 kilowatt in one generation. Uh this is something you have to plan ahead [10:04] before you uh plan your full uh fleet and facility. Two, [10:08] you don't you do not have to pick a lane. Air cool for you for the ones you're building today [10:14] and full liquid cooling racks for the ones you build tomorrow. And then lastly is the one I would [10:21] underline. Integration is the hard part. The GPUs are the same everywhere. What actually differs [10:28] is whether your partner owns the deployment or just the shipment. All right, thank you everyone.