Nvidia 与 AMD 将认购 Meta 共 40 亿美元股份,并扩大芯片合作
Nvidia, AMD to buy $4B in Meta shares as part of expanded chip partnerships - SiliconANGLE
Meta 宣布 Nvidia 和 AMD 计划在其最高 950 亿美元的股票发售中认购合计 40 亿美元股份,Broadcom 也参与其中,三家芯片厂商合计投资 50 亿美元。Nvidia 出资 30 亿美元,Meta 将扩大使用其 Rubin GPU、Vera CPU、Spectrum-X 以太网产品和 BlueField-4,并借助 Nvidia 优化 Muse 大模型与消费级 AI 智能体的性能与成本。AMD 出资 10 亿美元,Meta 已在其数据中心部署首批基于 Instinct MI450 的 Helios 机架并正在测试,后续还将采用基于 AMD Instinct 架构的定制 AI 加速器。
推荐理由
AI 评分 71
Meta 获 Nvidia、AMD、Broadcom 合计 50 亿美元入股,同时扩大 Rubin、Helios 等芯片采购,呈现算力与资本绑定的合作结构。
主题
正文
Meta Platforms Inc. today announced that Nvidia Corp. and Advanced Micro Devices Inc. plan to buy $4 billion worth of its shares.
The chipmakers will purchase the equity as part of an up to $95 billion stock sale that the Facebook parent intends to kick off in the near future. The fundraising initiative will also include the participation of Broadcom Inc. Meta didn’t specify how much the company is investing, but it stated on Monday that Nvidia, AMD and Broadcom together plan to invest $5 billion. That suggests the networking chip giant will buy $1 billion in shares.
Nvidia is investing $3 billion as part of a newly expanded partnership that also includes other components. Notably, Meta plans to broaden its use of the chipmaker’s Rubin graphics processing units and Vera central processing units. Companies often deploy the two chips together in Nvidia-designed DGX Vera Rubin NVL72 appliances. Each system includes 72 GPUs and 36 CPUs.
Meta also plans to buy more Nvidia networking gear. It will expand its use of the Spectrum-X Ethernet product family, which includes SuperNICs and switches. SuperNICs are chips that connect servers to a data center’s network, while switches coordinate how traffic flows through that network.
GPU clusters’ traffic patterns tend to change rapidly. As a result, connectivity bottlenecks sometimes emerge faster than the network can respond to them. Nvidia says that its Spectrum-X Ethernet product series addresses the challenge by checking GPU traffic for anomalies at sub-microsecond time intervals. A microsecond corresponds to one millionth of a second.
Another focus of Meta’s planned network investment is Nvidia’s BlueField-4 chip family. The accelerators in the series are designed to speed common networking, storage and cybersecurity tasks.
Nvidia will help Meta improve the performance and cost efficiency of its Muse large language model family. The effort will extend to the Facebook parent’s eponymous consumer AI agent.
Additionally, Meta plans to explore “collaboration opportunities” related to Nvidia’s recently introduced OpenShell Reference Architecture. It’s a collection of technical resources that makes it easier to run the chipmaker’s OpenShell toolkit. The toolkit runs AI agents in isolated sandboxes designed to mitigate cybersecurity risks.
“Nvidia and Meta are expanding our partnership to optimize the computing infrastructure end to end for capable, efficient and secure AI agents, so Muse can empower billions of people to achieve more and pursue greater ambitions,” said Nvidia Chief Executive Jensen Huang.
AMD, which will invest $1 billion in Meta, today revealed that the Facebook parent has adopted its Helios rack design. Helios contains 72 GPUs much like the DGX Vera Rubin NVL72. It’s based on AMD’s latest Instinct MI450 graphics card architecture, Epyc CPUs and Pensando data processing units, which are the chipmaker’s answer to the BlueField-4.
It says Meta has deployed an “initial” set of Helios racks in its data centers and is currently testing them. The social media giant plans to deploy additional systems in the near future.
Further down the line, Meta will adopt a custom AI accelerator based on AMD’s Instinct graphics card architecture. The chipmaker, for its part, will use the technical data collected through the partnership to improve its future chip and rack designs.
“Together, we are building and deploying infrastructure at gigawatt scale and shaping future generations of silicon, software and systems to enable Meta’s most ambitious superintelligence models and services,” said AMD CEO Lisa Su.
Image: Nvidia