OpenAI’s $400 Billion Quest: Examining the Feasibility of Its AI Data Center Expansion

The rapid expansion of artificial intelligence continues to dominate headlines, often accompanied by audacious claims of unprecedented growth and monumental infrastructure demands. Yet, a closer look at the financial and logistical requirements reveals a potentially unsustainable trajectory, particularly for leading AI firms like OpenAI. Recent announcements, such as the partnership between Broadcom and OpenAI for another 10 gigawatts (GW) of custom chips and capacity by 2029, are often reported without critical scrutiny, despite their staggering implications.

The True Cost of AI Infrastructure: A $50 Billion per Gigawatt Reality

Building substantial data center capacity is an incredibly expensive and time-consuming endeavor. While estimates vary, the consensus among industry experts points to an escalating cost. Jensen Huang, CEO of NVIDIA, suggests computing hardware alone can cost $50 billion per gigawatt, excluding buildings and power infrastructure. Barclays Bank places the total cost between $50 billion and $60 billion. After accounting for land, power infrastructure, networking, and the sheer volume of advanced GPUs (like NVIDIA’s Blackwell), a conservative, rounded estimate for constructing a single gigawatt of data center capacity stands at approximately $50 billion. Crucially, such projects typically require at least two and a half years to complete.

OpenAI’s Ambitious Capacity Promises Under Scrutiny

OpenAI has announced plans for an astounding 33 GW of capacity across collaborations with AMD, NVIDIA, Broadcom, and its own “Stargate” data centers. However, some of these claims warrant closer examination. For instance, a site in Lordstown, Ohio, initially linked to Stargate, has been clarified by SoftBank as “not a full-blown data center” but rather a facility for “storage containers that will hold the infrastructure for AI and data storage.” Such discrepancies raise questions about the true nature and scale of OpenAI’s planned deployments.

Unrealistic Timelines: The 2026 Deployment Challenge

OpenAI’s near-term commitments present an even more immediate logistical puzzle. The company has outlined several ambitious goals for the second half of 2026:

• OpenAI and Broadcom aim to tape out, complete, and manufacture enough AI inference chips to fill a 1 GW data center. This proposed data center currently lacks a publicly known location or any commenced construction. Building a 1 GW facility, including securing 1.2 GW to 1.3 GW of power capacity (allowing for cooling systems and transmission losses), typically requires years of lead time.
• AMD and OpenAI plan to begin the “first 1 gigawatt deployment of AMD Instinct MI450 GPUs.” This, too, is slated for an unnamed data center location, which would have needed construction and power procurement to begin at least a year ago to meet the late 2026 deadline.
• NVIDIA and OpenAI intend to deploy the first gigawatt of NVIDIA’s Vera Rubin GPU systems as part of their reported $100 billion deal. Similar to the other projects, this necessitates a currently unnamed data center location, with construction needing to have commenced well in advance of the projected deployment timeline.

These timelines appear highly aggressive, bordering on impossible, given the typical lead times for such massive infrastructure projects. The required funding for these initial data centers alone would exceed $100 billion, with a significant portion needed upfront.

The Global Financial System Cannot Afford OpenAI’s Vision

The financial scale of OpenAI’s ambitions is truly staggering. Calculations suggest that to fulfill its various commitments, OpenAI requires approximately $400 billion within the next 12 months. This figure includes:

• At least $50 billion for Broadcom’s 1 GW data center.
• An additional $200 billion for further data center capacity to reach 10 GW by 2029.
• At least $50 billion for NVIDIA’s 1 GW data center.
• $40 billion for its 2026 compute contracts.
• At least $50 billion for AMD’s chips and 1 GW data center.
• Additional costs for consumer device development and ARM CPU licensing.

This $400 billion demand surpasses the entire global venture capital raised in 2024 ($368 billion) and dwarfs the lifetime funding of companies like Uber. When factoring in operational costs, including sales and marketing (reportedly $2 billion in the first half of 2025 alone) and salaries, the $400 billion estimate quickly becomes a conservative minimum.

For these deals to materialize by the end of 2026, substantial tranches of this capital would be required by early 2026. The notion that OpenAI can simply “raise debt” to cover these sums within such compressed timelines, while simultaneously paying for compute contracts (Oracle, CoreWeave, Microsoft, Google) and converting its non-profit structure to a for-profit entity by October 2026 (to avoid $20 billion in SoftBank funding becoming debt), places an unprecedented burden on the global financial system.

OpenAI’s long-term goal of 250 GW of capacity by 2033, estimated to cost $10 trillion (one-third of the US economy’s output in 2024), seems utterly detached from reality. To put this in perspective, Goldman Sachs estimated global data center capacity at around 55 GW in February – OpenAI aims to add five times that capacity, by itself, in just eight years.

Discrepancies and Underperformance

Concerns also arise from inconsistencies in OpenAI’s operational capacity claims. Sam Altman reportedly stated OpenAI began the year with 230 MW of capacity and aims to exit 2025 with over 2 GW. If OpenAI currently lacks its own capacity, this would imply an acquisition or construction of 1.7 GW in a short span, equivalent to all operational data centers in the UK last year, without public disclosure. Sources like Stargate Abilene, despite projections, currently only have 200 MW of the 1.5+ GW needed, further highlighting the gap between stated goals and reality.

Financially, OpenAI’s revenue projections of $13 billion for the current year seem challenging, given it was at approximately $5.3 billion by the end of August. Meanwhile, spending on research and development has been immense. The Information reported $6.7 billion in the first half of 2025 alone, with much of the 2024 R&D budget allocated to experimental runs and unreleased models, rather than actual user-facing model training. Furthermore, recent releases like GPT 4.5 and GPT 5 have been described as expensive and underwhelming, while Sora 2 faced plagiarism concerns requiring adjustments.

The fundamental question remains: what specific, reliable utility does ChatGPT offer that justifies trillions in data center investment? While AI promises much, the “it’s going to be really good” argument, or even the aspiration for Artificial General Intelligence (AGI) by 2030, does not concretely address the immediate and staggering infrastructure requirements.

The Oracle Deal: A Case Study in Unreality

The reported $300 billion, five-year deal between OpenAI and Oracle serves as a stark illustration of the current challenges. Oracle would need 4.5 GW of IT load capacity to provide the promised compute. Despite Oracle CEO Greg Magouyrk’s confidence, OpenAI’s current financial trajectory suggests it cannot sustain $60 billion payments annually. More critically, Oracle’s own infrastructure falls far short of this capacity. Stargate Abilene, a key component, is already behind schedule and severely underpowered. Oracle’s only other significant data center project, a 1.4 GW plot in Shackelford, Texas, has only just begun construction and is projected to have a single building operational by late 2026 – a fraction of what’s needed.

A Confidence Game with Real Consequences

The confluence of unrealistic timelines, astronomical financial demands, and discrepancies in stated capacity paints a concerning picture. OpenAI’s rapid expansion plans, while lauded by many, appear to operate on a scale that the global financial and industrial systems are currently incapable of supporting. This situation raises serious questions about market manipulation and the potential for a “confidence game” that, if unchecked, could lead to significant financial instability and harm to retail investors who are captivated by the promises of exponential growth. A more grounded and realistic assessment of the actual capabilities and infrastructure required for true AI progress is urgently needed.