OCP 2026：Agentic AI 推动数据中心从芯片走向电网

NVIDIA 进一步开放 AI Factory 架构

NVIDIA 的 Ian Buck 在开幕主题演讲环节结束时进行了广泛介绍，涵盖 Vera Rubin 系统、开放机架设计、网络、电力输送以及 AI Factory 软件。

该公司宣布，基于其第三代 MGX 1.2 参考架构，向 OCP 作出更多贡献。

该规范包括高密度液冷计算系统、改进的冷却连接、标准化电力接口以及模块化机械设计。

Buck 强调，这些进展与其说仅关乎计算性能，不如说同样关乎制造效率和可维护性。

他介绍称，NVIDIA 通过简化机械结构和提升自动化装配程度，将计算托盘的装配时间从 Grace Blackwell 一代的大约 90 分钟缩短至 Vera Rubin 的约 1 分钟。

在数千个机架的规模下，这些制造环节的改进能够对部署进度产生实质影响。

NVIDIA 还预览了下一代 Vera Rubin Ultra 架构，该架构旨在通过采用共封装光学技术的先进 NVLink 网络互连 576 块 GPU。

开放数据中心与电网韧性规范

但对于数据中心开发商和运营商而言，NVIDIA 对 OCP 最具影响力的贡献，或许并不在计算机架内部。

NVIDIA、AMD、Google 和 Meta 已发布 Open Data Center Specification，目前版本为 0.7，旨在为 AI Factory 设施建立通用设计原则。

该规范涉及电力架构、冷却基础设施和物理布局，帮助设备供应商与开发商围绕共享的基础设施需求开展协调。

这项共享设施相关工作与供应商特定项目同步推进，例如 NVIDIA 的 DSX Ready。该项目为支持其 AI Factory 设计的电池储能和液冷设备制定了认证要求。

NVIDIA 还与 Google、Meta 和 Microsoft 一同发布了 OCP Grid Resilience Specification，针对大型 AI 部署在电网扰动期间的运行行为提出规范。

该规范明确了设施应保持并网、自动响应扰动，以及快速恢复至正常运行负载的预期要求。

这些举措反映出，将 AI Factory 视为一个集成的电气与计算系统正变得愈发重要。

Buck 明确阐述了更广泛的目标：

“AI Factory 必须成为电网的一部分，不仅是消费者，也应成为贡献者。”

他描述了一种未来：计算工作负载可以根据电力系统状况动态响应，在必要时转移或降低需求，而非始终表现为缺乏弹性的固定负载。

NVIDIA 还重点介绍了 Visionbay 计划中的 K-1 AI Factory 架构，该架构采用 800 伏直流配电，并集成 Vertiv 提供的配套电力与热管理技术。

更高电压的配电方式有望减少铜材需求和电能转换环节，同时支持密度日益提升的计算部署。

据 Buck 表示，相比某些传统替代方案，800VDC 方案可将铜材需求减少约 45%。

其意义超越了单个机架。随着 AI 部署迈向数百兆瓦乃至最终吉瓦级规模，电力分配将成为制约设施设计、部署速度和运营效率的核心因素。

将电力作为计算资源进行管理

NVIDIA 还提出了 AI Factory 优化的另一重要维度：利用软件从既有电力基础设施中释放更高的计算生产力。

Buck 介绍了 DSX MaxLPS，这是一种动态电源管理方法，可依据实际工作负载需求分配电力容量，而不是假设每个处理器都会持续以其额定最大功耗运行。

其前提十分直接：数据中心可能拥有固定的公用事业供电配额，但其计算设备并不一定会在每个时刻都消耗全部配额。

这种差异创造了容纳额外工作负载或计算系统的机会，前提是在电力限制范围内对用电进行管理。

在此前与 Nscale 于其冰岛设施发布的一项联合评估中，NVIDIA 展示了在同样的 264.4 千瓦已配置电力预算下，将受管 GPU 数量从 140 块增加至 192 块。

在所评估的工作负载配置下，该测试实现了约 49% 的聚合推理吞吐量提升。

尽管这些结果仅适用于该测试环境，且不应假定能够在所有部署中统一复现，但它们指向了一项重要的新兴策略。

AI Factory 运营商或可通过协同工作负载调度、动态电源管理以及更精细化的基础设施控制，提升受限电力容量的生产效率。

这为业界熟悉的提升电源使用效率的追求引入了另一个维度。

目标不再只是降低辅助基础设施消耗的能源，而是在既有可用电力容量内提高有效计算产出。

光网络与开放传输生态系统

NVIDIA 还详细介绍了其 Spectrum-X 网络架构的进展，包括共封装光学和多平面以太网交换。

该公司表示，其新的 Spectrum-X Photonics Multiplane 交换机将在降低大型 AI 网络复杂度的同时，提升能效和故障响应能力。

Buck 强调了硬件辅助故障检测能力，能够在约 2.6 毫秒内识别网络问题。

NVIDIA 还宣布扩大与 Meta 在网络领域的合作，包括为这家超大规模云服务商的新型液冷 IceCube 交换平台提供 Spectrum-6 技术。

此外，NVIDIA 与 Google 宣布，双方正在推动未来的 NVIDIA ConnectX-10 SuperNIC 和 BlueField DPU 支持 Google 的 Falcon 传输技术。

这些进展进一步凸显了通用网络接口日益扩大的重要性，尽管主要供应商仍通过不同的芯片、交换架构和软件实现展开竞争。

业界日益复杂的 AI 网络无法承受完全孤立的硬件生态系统所带来的限制。

新的基础设施方程式

OCP 2026 的开幕主题演讲展示了 AI 正在如何彻底改变数据中心开发背后的工程假设。

向 Agentic 工作负载的转变，正在为 CPU 基础设施创造新的需求，同时加速对 GPU 和专用芯片的投资。

吉瓦级项目正在改变人们对建设周期、冷却、网络架构和运营可靠性的预期。

而曾主要被视为支撑性需求的电力基础设施，如今正变得与计算系统设计和软件编排密不可分。

不过，本届峰会最具启示性的特点或许在于，彼此竞争的科技公司正承认它们面临一个共同的工程难题。

AMD、NVIDIA、Arm、Broadcom、Meta 以及更广泛的 OCP 社区，仍在处理器、网络技术、软件和集成系统领域激烈竞争。

与此同时，AI 基础设施在物理和运营层面的复杂性日益要求采用通用接口、可复用设计与协作规范。

这种张力——专有计算优势与开放基础设施需求之间的张力——如今已处于 AI Factory 建设浪潮的核心。

对于数据中心开发商、运营商及其供应商而言，其影响远远超出下一代服务器硬件。

随着 AI 计算从单个机架扩展到地理上分布广泛、功耗密集的基础设施，行业部署和运营容量的能力将越来越取决于整个系统能否被高效地协同工程化。

开放计算的下一个前沿不再局限于服务器机架，而是从芯片延伸至公用事业电网。

NVIDIA Opens More of the AI Factory Architecture

NVIDIA's Ian Buck closed the opening keynote program with an extensive presentation covering Vera Rubin systems, open rack designs, networking, power delivery and AI factory software.

The company announced additional contributions to OCP based on its third-generation MGX 1.2 reference architecture.

The specification includes high-density liquid-cooled computing systems, improved cooling connections, standardized power interfaces and modular mechanical designs.

Buck emphasized that these developments are as much about manufacturing and serviceability as they are about computational performance.

He described how NVIDIA reduced the assembly time for a compute tray from approximately 90 minutes in the Grace Blackwell generation to roughly one minute for Vera Rubin through simplified mechanical construction and more automated assembly.

At the scale of thousands of racks, those manufacturing improvements can materially influence deployment schedules.

NVIDIA also previewed its next-generation Vera Rubin Ultra architecture, designed to interconnect 576 GPUs using an advanced NVLink fabric with co-packaged optics.

Open Data Center and Grid Resilience Specifications

For data center developers and operators, however, NVIDIA's most consequential OCP contributions may lie outside the computing rack.

NVIDIA, AMD, Google and Meta have published an Open Data Center Specification , currently at version 0.7, intended to establish common design principles for AI factory facilities.

The specification addresses power architecture, cooling infrastructure and physical layout, helping equipment suppliers and developers coordinate around shared infrastructure requirements.

That shared facilities work unfolds alongside vendor-specific programs such as NVIDIA's DSX Ready, which establishes qualification requirements for battery energy storage and liquid cooling equipment supporting its AI factory designs.

NVIDIA also joined Google, Meta and Microsoft in publishing an OCP Grid Resilience Specification addressing the behavior of large AI installations during grid disturbances.

The specification establishes expectations for facilities to remain connected, respond automatically to disturbances and recover toward normal operating load rapidly.

These initiatives reflect the growing importance of treating the AI factory as an integrated electrical and computational system.

Buck made the broader objective explicit:

"AI factories need to be part of the grid, not just consumers but contributors."

He described a future in which computing workloads could respond dynamically to electrical-system conditions, shifting or reducing demand when necessary rather than behaving exclusively as inflexible loads.

NVIDIA also highlighted Visionbay's planned K-1 AI factory architecture, which incorporates 800-volt DC power distribution and supporting power and thermal technologies from Vertiv.

Higher-voltage distribution promises reductions in copper requirements and conversion stages while supporting increasingly dense computing installations.

According to Buck, 800VDC approaches can reduce copper requirements by approximately 45% relative to certain conventional alternatives.

The significance extends beyond individual racks. As AI deployments approach hundreds of megawatts and eventually gigawatt scale, electrical distribution becomes a central constraint on facility design, deployment speed and operating efficiency.

Managing Power as a Computing Resource

NVIDIA also introduced another important dimension of AI factory optimization: using software to extract greater computing productivity from existing electrical infrastructure.

Buck discussed DSX MaxLPS , an approach to dynamic power management that allocates electrical capacity according to actual workload requirements rather than assuming every processor will continuously operate at its maximum rated draw.

The premise is straightforward. A data center may have a fixed utility allocation, but its computing equipment does not necessarily consume that full allocation at every moment.

That difference creates opportunities to accommodate additional workloads or computing systems, provided power use can be managed within electrical limits.

In a previously published joint evaluation with Nscale at its Iceland facility, NVIDIA demonstrated an increase from 140 to 192 managed GPUs within the same provisioned 264.4-kilowatt power budget.

The test produced approximately 49% higher aggregate inference throughput under the evaluated workload configuration.

Although the results are specific to the test environment and should not be assumed to apply uniformly across deployments, they point toward an important emerging strategy.

AI factory operators may be able to increase the productivity of constrained electrical capacity through coordinated workload scheduling, dynamic power management and more granular infrastructure controls.

That introduces another dimension to the industry's familiar pursuit of improved power usage effectiveness.

The objective is not merely to reduce the energy consumed by supporting infrastructure. It is to increase useful computing output within the electrical capacity already available.

Optical Networking and an Open Transport Ecosystem

NVIDIA also detailed advances in its Spectrum-X networking architecture, including co-packaged optics and multiplane Ethernet switching.

The company said its new Spectrum-X Photonics Multiplane switches will improve energy efficiency and fault response while reducing the complexity of large AI fabrics.

Buck highlighted hardware-assisted fault detection capable of identifying network problems in approximately 2.6 milliseconds.

NVIDIA also announced expanded networking collaboration with Meta, including Spectrum-6 technology for the hyperscaler's new liquid-cooled IceCube switching platform.

Separately, NVIDIA and Google announced work to support Google's Falcon transport technology on future NVIDIA ConnectX-10 SuperNICs and BlueField DPUs.

These developments reinforce the expanding importance of common networking interfaces, even as major suppliers compete through different silicon, switching architectures and software implementations.

The industry's increasingly complex AI networks cannot afford to be constrained by completely isolated hardware ecosystems.

The New Infrastructure Equation

The opening keynotes at OCP 2026 demonstrated how thoroughly AI is changing the engineering assumptions behind data center development.

The shift toward agentic workloads is creating new demand for CPU infrastructure alongside accelerating investment in GPUs and purpose-built silicon.

Gigawatt-scale projects are changing expectations for construction schedules, cooling, network architecture and operational reliability.

And electrical infrastructure, once largely treated as a supporting requirement, is becoming inseparable from computing-system design and software orchestration.

Yet the most revealing feature of the summit may be the degree to which competing technology companies are acknowledging a shared engineering problem.

AMD, NVIDIA, Arm, Broadcom, Meta and the broader OCP community continue to compete vigorously in processors, networking technology, software and integrated systems.

At the same time, the physical and operational complexity of AI infrastructure increasingly demands common interfaces, reusable designs and collaborative specifications.

That tension—between proprietary computing advantage and the need for open infrastructure—now sits at the center of the AI factory buildout.

For data center developers, operators and the suppliers supporting them, the implications extend far beyond the next generation of server hardware.

As AI computing expands from individual racks into geographically distributed, power-intensive infrastructure, the industry's ability to deploy and operate capacity increasingly depends on how effectively the entire system is engineered together.

The next frontier of open computing is no longer confined to the server rack. It extends from the silicon to the utility grid.