OpenAI and Nvidia Are Building 10 Gigawatts of AI Data Centers — Here Is Why That Number Is Staggering

Introduction: A Number That Changes How You Think About AI

There is a number at the center of the biggest corporate infrastructure deal in technology history, and most of the people who read it do not fully understand what it means. That number is ten gigawatts. It appears in a letter of intent signed on September 22, 2025 by OpenAI and Nvidia — a document that formalized a plan to build the largest AI computing infrastructure ever attempted, backed by up to $100 billion in Nvidia investment. Nvidia CEO Jensen Huang called it "the biggest AI infrastructure project in history." OpenAI CEO Sam Altman said there is "no partner but Nvidia that can do this at this kind of scale, at this kind of speed."

The first phase is already coming online in the second half of 2026. And the energy required to run it will reshape power grids, electricity bills, and the global race for computing dominance in ways that reach far beyond any technology company's balance sheet.

What Ten Gigawatts Actually Means in Plain Language

A watt is a unit of power. A gigawatt is one billion watts. Ten gigawatts is ten billion watts of continuous electrical power being consumed at all times — not generated, not stored, consumed. To understand what that means in human terms, a few comparisons are necessary.

A typical nuclear power plant generates approximately one gigawatt of electricity. That single gigawatt is enough to power roughly one million average American homes continuously. OpenAI and Nvidia's combined infrastructure target is ten times that figure: enough electricity to power ten million homes, running around the clock, every day of the year. To put it another way, the entire data center power consumption of the United States currently stands at approximately 41 gigawatts — meaning this single partnership between two companies is targeting a capacity equivalent to roughly one quarter of all existing US data center power demand. Nvidia CEO Jensen Huang told CNBC that the 10 gigawatt project is equivalent to between four million and five million graphics processing units — approximately as many GPUs as Nvidia ships in an entire year, and twice what it shipped the year before.

The cost of building that infrastructure is proportionally enormous. Huang told investors that building one gigawatt of data center capacity costs between $50 billion and $60 billion, of which roughly $35 billion covers Nvidia chips and systems. At that rate, ten gigawatts represents somewhere between $500 billion and $600 billion in total construction cost — a figure larger than the GDP of most countries. Nvidia's $100 billion investment commitment is substantial precisely because it represents the chip company paying for a significant portion of its own customer's buildout, a structure that one analyst described plainly as "mutually beneficial circular vendor financing": Nvidia pledges money to OpenAI in exchange for OpenAI buying Nvidia systems.

The Letter of Intent and What It Actually Commits Both Companies To

It is worth being precise about what the September 2025 announcement was, and what it was not. The document signed by Altman and Huang was a letter of intent — a non-binding statement of purpose, not a signed contract. As of December 2025, Nvidia CFO Colette Kress confirmed that the companies had not yet completed a definitive agreement. OpenAI's current GPU purchases continue to flow through cloud partners Microsoft and Oracle while final terms are negotiated.

That distinction matters for investors and for anyone tracking deployment timelines, but it does not change the direction of what is being built. OpenAI will work with Nvidia as its preferred strategic compute and networking partner for its AI factory growth plans. The two companies have committed to co-optimizing their respective roadmaps — OpenAI's model and infrastructure software alongside Nvidia's hardware and software — in a technical partnership that goes beyond a simple procurement relationship. OpenAI is also developing its own AI chips internally, uses Google's TPUs and AMD's GPUs, and is expected to spend approximately $350 billion on cloud services primarily with Oracle, with additional capacity across Microsoft and Google. The Nvidia deal sits on top of all of that, representing the dedicated compute layer for what both companies have described as the path to superintelligence.

OpenAI is separately in advanced negotiations to lease a 10 gigawatt data center campus on Department of Energy land in Pike County, Ohio, with an estimated buildout cost of at least $500 billion. Nvidia would use its balance sheet to guarantee both OpenAI's lease payments and SB Energy's project financing — its largest-ever infrastructure guarantee. SB Energy, majority-owned by SoftBank, is developing the site with at least 9.2 gigawatts of natural gas-powered generation and billions in transmission upgrades. The first 800-megawatt phase of that facility is expected to begin operations in 2028, with full buildout taking at least a decade.

What the Vera Rubin Platform Is and Why It Was Chosen

The first gigawatt of Nvidia systems deployed under this partnership will run on Nvidia's Vera Rubin platform — the company's next-generation chip architecture designed specifically for the kind of massive-scale AI workloads that OpenAI requires. The name combines "Vera" — after Vera Rubin, the astronomer who provided the first strong evidence for dark matter — with "Rubin," honoring her further.

The platform entered production in August 2025 and began shipping in late 2025. Each Vera Rubin R100 GPU draws approximately 2,300 watts of thermal design power — nearly double that of its Blackwell predecessor. The Vera Rubin NVL144 configuration is expected to offer 3.6 exaflops of FP4 performance and 1.2 exaflops of FP8 performance. An exaflop is one quintillion floating-point operations per second. These are numbers that require scientific notation to write out fully, which is itself a signal of how far AI computing has moved beyond the hardware assumptions of just five years ago.

The choice of Vera Rubin for the first phase of the deployment tells you something important about the timeline. The first gigawatt of Nvidia systems "will generate their first tokens" — in OpenAI's phrase — in the second half of 2026. That means the most powerful commercial AI computing cluster ever built will be operational before this year ends, running on hardware that Nvidia only began producing in 2025. The pace of that compression — from announced partnership to operational gigawatt-scale infrastructure — is without precedent in the history of technology infrastructure.

The Energy Crisis Nobody Is Mentioning in the Press Releases

OpenAI and Nvidia's press releases are precise about gigawatts of compute. They are noticeably less specific about where ten gigawatts of electricity actually comes from, and what happens to the grids and communities that supply it.

AI data centers already consume more than 1,000 terawatt-hours of electricity globally per year as of 2026. Total US data center consumption of 41 gigawatts rivals the combined generating capacity of all US nuclear power plants. PJM, the grid operator serving the mid-Atlantic and Midwest — including Ohio, where OpenAI's largest campus is planned — projects a 6 gigawatt shortfall in generating capacity by 2027. Capacity market clearing prices in PJM for the 2026–2027 delivery year increased to $329 per megawatt, over ten times higher than the price two years prior, with rapid data center growth identified as a major contributing factor.

AEP Ohio, the utility that would serve any data center in Pike County, has already paused all new data center interconnections due to insufficient power infrastructure. Communities across Georgia, Indiana, Missouri, and Washington are pushing back against proposed facilities, demanding that technology companies fund their own power plants and transmission upgrades rather than shifting costs to local ratepayers. The average American household electricity bill could rise between $15 and $25 per month due to grid upgrade costs driven by AI data center demand — costs that are already beginning to appear in rate filings before a single Vera Rubin GPU has been switched on at the Ohio campus.

Faced with the reality that renewable energy alone cannot scale fast enough to meet AI's power appetite, technology companies including Microsoft, Google, and Amazon have signed nuclear power purchase agreements and are investing in small modular reactors. Microsoft restarted the Three Mile Island Unit 1 reactor — now renamed after a branding agreement with Constellation Energy — specifically to supply its data centers. The Ohio campus's planned power supply includes at least 9.2 gigawatts of natural gas generation. The AI industry's energy strategy, stated plainly, is to build its own power plants rather than wait for public utilities to expand.

Each ChatGPT query already uses approximately ten times more power than a standard Google search — equivalent to running an LED light bulb for twenty minutes. At the scale OpenAI is targeting, the aggregate energy consumption of its AI services will be measurable at the level of national electricity grids.

What the Infrastructure Arms Race Means for Ordinary Technology Users

For someone who uses ChatGPT, Copilot, or any of the AI tools built on top of OpenAI's models, the ten gigawatt buildout has several direct implications — most of them positive in the short term, with important caveats for the medium term.

More compute means more capable models. Every significant leap in AI capability over the past five years has been associated with a corresponding increase in the computing power used to train the underlying model. GPT-4 was more capable than GPT-3 in large part because it was trained on more hardware, for longer. The ten gigawatt infrastructure is explicitly described, in the OpenAI-Nvidia announcement, as what is needed to "reach the next level of scale" and to "empower every individual and business" with capabilities that are not currently possible. If that projection holds, the AI tools ordinary users interact with in 2027 and 2028 will be meaningfully more capable than the ones available today, and that change will be a direct consequence of the infrastructure being built right now.

The negative implication is cost socialization. The $500 billion to $600 billion in construction costs, the nuclear power plants, the grid upgrades, and the transmission investments all have to be paid for somewhere. Some of those costs appear in subscription prices for AI services — a dynamic already visible in GitHub Copilot's shift to token-based billing discussed earlier this week. Some appear in electricity rates for households near data center clusters. Some are absorbed by investors who fund the construction in exchange for future returns. None of them disappear. The ten gigawatt number is not just a measure of computing ambition. It is a measure of financial and physical commitment that will be distributed across the economy in ways that are only beginning to become visible.

OpenAI President Greg Brockman described the scale of what the partnership represents by comparing it to a moment from the company's founding: "This is a billion times more computational power than that initial server." Jensen Huang hand-delivered that first Nvidia DGX system to OpenAI's office in 2016. The partnership announced nine years later targets a billion times more capacity. The pace of that scaling is the central fact of the AI era — and ten gigawatts is what it looks like when you write it in units of electricity.

Conclusion: The Physical Consequence of Digital Ambition

The OpenAI-Nvidia partnership is not primarily a software story. It is an infrastructure story — one about land, steel, cooling systems, power cables, natural gas plants, nuclear reactors, and the communities that sit next to them. The ten gigawatt target is not a number that exists only on a slide deck. It will exist as physical facilities drawing power from real grids in real states, with real consequences for electricity prices, water consumption, carbon emissions, and local economies.

The first gigawatt comes online before the end of 2026. The full ten gigawatts, if the partnership reaches its stated goal, will take years beyond that. What will be built in the meantime is not just the most powerful AI computing infrastructure in history. It is the physical foundation on which the next generation of AI models — and every application built on top of them — will actually run.

The number is ten gigawatts. Now you know what it means.