Securing the Infrastructure of Intelligence

NVIDIA is moving deeper into the boring, expensive parts of AI. The company said it is partnering with SB Energy to secure LPS capacity at the PORTS-Pike Technology Campus in Portsmouth, Ohio, so NVIDIA compute can live there. OpenAI will be the tenant. That sounds less flashy than a new model release, but it may matter more in the long run.

The pitch is simple: AI factories need more than chips. They need packaging, memory, networking, land, power and shell, plus the kind of long-term site control that lets multiple generations of hardware roll through the same buildings. NVIDIA says it has already used its scale and supply-chain relationships to secure semiconductor resources, and is now applying the same approach to LPS capacity for its own AI factories.

At PORTS-Pike, OpenAI is expected to build and operate a world-class AI factory using NVIDIA’s DSX platform, with GPUs, CPUs, networking and infrastructure software. The initial deployment is expected to deliver 4.25 gigawatts of capacity. NVIDIA says each generation of systems deployed there could amount to about 1.5 million GPUs, or about $150 billion to $200 billion in NVIDIA revenue. The site is built for the long haul, with room for multiple upgrade cycles over 20 years.

NVIDIA is also explicit about how the deal works. Its support covers defined portions of lease and power payments, plus a residual-value commitment. It is not taking on the full site cost or all of the tenant’s obligations. The guarantee starts in phases as data centers come online between 2028 and 2030, and NVIDIA says its exposure falls as OpenAI makes lease payments and capacity is delivered.

There’s a broader bet underneath all this. NVIDIA says OpenAI has committed to roughly 12 gigawatts of NVIDIA compute through 2030, with the possibility of stretching to 16 gigawatts if NVIDIA extends the PORTS-Pike arrangement beyond the first 4.25 gigawatts. At those levels, NVIDIA pegs the opportunity at roughly $600 billion of compute through 2030. That is a lot of hardware, but it is also a reminder that the real bottleneck in frontier AI is increasingly physical, not just software.