黄仁勋宣布与SB Energy合作，为OpenAI建AI工厂

Core Highlights

Jensen Huang announced on his personal X account that NVIDIA and SB Energy have reached a partnership to build dedicated AI factories for OpenAI at the PORTS-Pike technology park in Ohio. The essence of this transaction is to bundle "compute" and "electricity" together as a leasable infrastructure commodity, with OpenAI serving as the first anchor tenant locked into the deal.

Rather than simply selling GPUs, NVIDIA is now selling an entire vertically integrated production facility for intelligence. The framing matters: in Huang's telling, the future of AI is not data centers but factories that consume energy to manufacture tokens, and the company intends to own the blueprint for those factories from the silicon up. Treating power and chips as one product also lets NVIDIA price the whole stack instead of competing on component margins alone, a notable shift in how the company describes its own role from component supplier to operator of the plants that turn energy into intelligence.

What It Does or What Happened

In the first phase, the partnership will lock in LPS (large power delivery) capacity at the PORTS-Pike park, reserved exclusively for NVIDIA AI factories. The initial deployment is expected to deliver roughly 4.25 gigawatts of AI-factory capacity. Each system generation will pack about 1.5 million NVIDIA high-performance computing chips, a scale that corresponds to a revenue opportunity estimated between 150 billion and 200 billion dollars.

OpenAI has already committed to deploying roughly 12 gigawatts of NVIDIA compute by 2030, with the option to expand that figure to 16 gigawatts. Taken together, NVIDIA frames the total opportunity created by this single arrangement at around 600 billion dollars. The numbers are striking because they describe not one product but a multi-year, multi-gigawatt build-out pipeline that compounds with every generation and gives both companies a shared incentive to keep scaling. The structure also aligns incentives cleanly: SB Energy is paid to keep the lights on, NVIDIA is paid for the compute, and OpenAI receives reserved capacity without owning the physical construction risk.

Technical Details

An AI factory, in NVIDIA's definition, integrates training and inference clusters together with power delivery and thermal management into a single delivered unit. The scale of nearly 1.5 million HPC chips per generation implies park-scale power consumption and interconnect challenges that go well beyond a conventional server room.

Meeting that demand requires SB Energy to secure and stably deliver the reserved capacity on the grid side, including transmission access, substation build-out, and contingency planning for peak loads. The chip count also means the interconnect fabric and cooling must be engineered as part of the same turnkey package rather than bolted on afterward, which is precisely where SB Energy's utility expertise is meant to fit and where most competing builds tend to stumble. The integration further shortens the time from groundbreaking to first training run, because power and compute are specified together rather than commissioned in separate, sequential projects that rarely line up on the same calendar.

Versus Competitors

Compared with the self-built data center strategies pursued by Microsoft, xAI and several other hyperscalers, NVIDIA has chosen to bind large customers by "selling factories rather than just chips." By integrating power, land and silicon into a single turnkey solution, it strengthens its bargaining position across the entire AI infrastructure chain.

The move also blurs the line between chip vendor and infrastructure developer. Competitors must still negotiate power, real estate and networking separately, while NVIDIA presents a finished facility that a tenant like OpenAI can occupy and scale on a timeline measured in quarters rather than years, removing a whole class of execution risk from the customer's plate and making the offering harder to compare against commodity cloud. The contrast becomes sharper as the scarce input shifts from silicon to electrons, where NVIDIA now controls both ends of the equation.

Industry Impact or Use Cases

For OpenAI, locking in long-term power and compute is effectively buying an expansion ticket far ahead of demand. It removes the single largest external uncertainty in scaling frontier models: whether the electricity to run them will exist where and when it is needed, at a price that keeps training economics sane.

For the broader industry, the compute race is upgrading from "who has more chips" to "who secures the electricity first." Energy is rapidly becoming the scarcest underlying resource of the AI era, and arrangements like PORTS-Pike may set the template for how the next decade of intelligence infrastructure is financed, permitted and built at national scale, with utilities and sovereign funds drawn into what used to be purely a Silicon Valley contest. If the template spreads, the winners of the AI race may be decided less by who designs the best chip and more by who can secure megawatts at scale, a dynamic that favors incumbents with balance sheets large enough to underwrite gigawatts of capacity years in advance.