英伟达财报与AWS扩GPU部署推高数据中心建设与电力约束

“I’m running out of hyperbole and adjectives.”

That was analyst Steven Dickens, CEO and principal analyst at HyperFrame Research, reacting to Nvidia’s latest results, which showed the chipmaker’s data center revenue climbing 117% year over year to $89 billion.

The growth is translating into another wave of infrastructure demand. Amazon Web Services plans to deploy an additional 2 million Nvidia GPUs across its global infrastructure in 2027 and 2028, following customer demand that exceeded earlier expectations.

For data center operators, the challenge is turning that compute demand into facilities that can actually be built, powered, cooled, and brought online. The number of GPUs ordered does not translate directly into an equivalent electrical load, but it does indicate greater demand for space, power, cooling, and high-performance networking.

“The physical buildout now has to keep pace with a technology deployment curve that can move much faster than generation, transmission, and interconnection infrastructure,” said Neil Osnato, founder of Persistence Analytics Group.

Nvidia reported $96.2 billion in total revenue for the quarter ended July 26, up 106% from a year earlier and 18% from the previous quarter. Data center revenue rose 18% sequentially to $89 billion. The company expects another jump in the current quarter, forecasting total revenue of $108 billion, excluding any data center compute revenue from China.

AWS Doubles Down on Nvidia GPUs

AWS offered a concrete example behind Nvidia’s results, saying on Wednesday that it plans to deploy 2 million additional Blackwell Ultra, Rubin, and Rubin Ultra GPUs across its global infrastructure in 2027 and 2028. The systems will support workloads that include agentic AI, scientific computing, enterprise automation, and “physical AI.”

AWS had already planned to add more than 1 million Nvidia GPUs beginning in 2026. The companies said customer demand has since exceeded those expectations.

Nvidia CEO Jensen Huang said the expanded AWS relationship reflects a market where “demand is running ahead of every forecast.”

AWS and Nvidia did not disclose the capacity or power requirements associated with the deployment.

The announcement comes as cloud providers and AI infrastructure companies pursue larger, denser computing systems that require facilities capable of delivering substantial power, advanced cooling, and high-performance networking.

Power Constraints Complicate GPU Deployments

The 2 million GPUs cannot be converted into a simple megawatt figure. Osnato said electricity requirements will depend on the GPU mix, utilization, cooling architecture, power density, deployment schedule, geography, and supporting equipment. He urged utilities and grid planners to distinguish between “announced compute capacity, executable data center capacity, and dependable electrical load.”

“They are related, but they are not the same thing,” Osnato said.

Power is already becoming a binding constraint in many markets, he added. Developers can order chips faster than substations can be built, transmission can be upgraded, generation can be interconnected, or large-load service can be validated.

“The limiting factor is increasingly shifting from access to compute hardware toward access to executable electrical infrastructure,” Osnato said. That creates a gap between the computing capacity companies announce and the amount of data center load that can actually be energized. The distinction matters because Nvidia’s sales strongly signal demand for computing, while the timing and location of the resulting electrical load remain uncertain.

AI Infrastructure Demand Broadens Beyond Hyperscalers

Nvidia’s results also show the AI infrastructure market broadening beyond the largest cloud providers. The company reported $48.7 billion in second-quarter data center revenue from hyperscale customers, up 13% from the previous quarter and 102% year over year. Its AI Clouds, Industrial & Enterprise (ACIE) business generated $40.3 billion, up 25% sequentially and 138% year over year, reflecting demand from AI-native companies, enterprises, and sovereign customers, as well as hyperscalers using AI clouds, Nvidia said.

“We’re in a rampant buildout phase from enterprise, small regional cloud provider, [neocloud providers], big-name cloud, hyperscale,” Dickens said, noting that Nvidia’s customer base now spans the largest cloud providers as well as smaller cloud companies and enterprises.

The broader customer base also changes the power-demand picture. Hyperscaler demand tends to concentrate into very large campuses and major utility interconnections. Enterprise, sovereign, industrial, and specialized AI deployments can create a more distributed demand profile, with smaller individual loads spread across more utility territories.

“The grid may not just be dealing with a handful of enormous 1 GW campuses,” Osnato said. “It may also be dealing with hundreds or thousands of smaller AI loads competing for capacity at different points in the system.” That makes forecasting harder for utilities and grid planners, who must account for how much load will materialize, where it will appear, when it will become operational, and how reliably it will persist.

Nvidia’s Vera CPU Targets Agentic AI at Scale

Nvidia is also expanding the computing infrastructure around its GPUs. On Thursday, the company said AWS had received its first Vera CPU server and Vera Rubin GPU, following a joint plan to bring Vera-based infrastructure to AWS. Vera is Nvidia’s first custom CPU, designed for agentic AI workloads, in which CPUs handle orchestration, tool calls, data movement, analytics, and other tasks related to GPU-accelerated computing. Nvidia said Vera has 88 custom cores and 1.2 TB/s of memory bandwidth, delivering up to 1.8x faster per-core performance on agentic AI workloads.

Oracle Cloud Infrastructure (OCI) plans to deploy hundreds of thousands of Vera CPUs beginning in 2026, according to Nvidia. OCI is the first cloud provider to deploy Vera at hyperscale, the company said.

Vera also serves as the host processor in Nvidia’s Vera Rubin NVL72 systems, connecting to Rubin GPUs through the NVLink-C2C interconnect.

Nvidia Backs Data Center Builds with SB Energy

Nvidia is moving deeper into the infrastructure needed to deploy its systems. The company said this month it would provide credit support for land, power, and shell construction at SB Energy’s planned PORTS-Pike Technology Campus in Ohio. The project is planned for 8 IT GW of capacity for OpenAI, with an initial 4.25 IT GW secured by Nvidia and an option for the remaining 3.75 IT GW.

OpenAI will lease the facility from SB Energy, which will build, own, and operate the campus under a 20-year agreement. Nvidia also agreed to invest $1.5 billion in SB Energy. The Ohio project gives Nvidia a direct role in securing AI computing facilities, extending its involvement beyond supplying processors.

Nvidia said its fiscal 2028 revenue outlook calls for 70% growth, though the company described that outlook as supply-constrained. For data center developers, utilities, and regulators, the question is how much of that demand can become actual electrical load.

“Strong chip demand is evidence of strong compute demand,” Osnato said. “It is not, by itself, proof that every megawatt being planned around that demand will arrive on schedule or persist for the life of the infrastructure built to serve it.”

Osnato said infrastructure planners increasingly need to distinguish among represented, executable, and durable demand.

About the Author

Senior News Writer, Data Center Knowledge

Shane Snider is Senior News Writer at Data Center Knowledge, covering AI infrastructure, hyperscale data centers, cloud platforms, and the power and energy systems driving modern compute expansion. His reporting focuses on the operational, economic, and environmental forces reshaping digital infrastructure, including AI factories, utility constraints, liquid cooling, renewable energy procurement, and next-generation data center architectures. He has won recent Azbee awards for news series and government reporting. Based in Raleigh, North Carolina, Snider covers how hyperscalers, utilities, chipmakers, and infrastructure providers are responding to the rapid rise of AI workloads and global compute demand.

You can reach Shane at [email protected] or on LinkedIn.