NVIDIA and IREN Strike a $2.1 Billion Deal to Build 5 Gigawatts of AI Infrastructure

The amount of electricity needed to power most of a mid-sized American state for an entire year is roughly 5 gigawatts. That's the scale NVIDIA and IREN committed to on May 7, 2026, when they announced a strategic partnership to build up to 5 gigawatts of AI data center infrastructure across IREN's global portfolio.

The financial structure is as notable as the scale. NVIDIA didn't write a check. Instead, IREN issued NVIDIA a five-year warrant to purchase up to 30 million ordinary shares at $70 per share, a potential $2.1 billion investment contingent on regulatory approval. IREN trades on NASDAQ under the ticker IREN, and shares rallied sharply on the news.

What makes this unusual for an infrastructure deal is the alignment it creates. NVIDIA's financial upside is now directly tied to IREN's ability to execute on AI infrastructure at scale. IREN gets access to NVIDIA's architecture standard, its enterprise customer relationships, and an implicit stamp of approval from the dominant AI hardware company. It's a different structure than a customer-vendor deal. It's closer to a shared bet.

What NVIDIA's DSX Architecture Actually Is

DSX is NVIDIA's Vera Rubin DSX AI Factory reference design, and it represents a material shift in how large AI infrastructure is designed.

Traditional data center engineering treats compute, power, cooling, and networking as separate problems. You source servers from one set of vendors, cooling systems from another, grid connections through electrical contractors, and management software from a third category of providers. The result is that integrated facility performance depends on coordination between systems that weren't built to work together.

DSX changes the philosophy. NVIDIA's reference architecture specifies how compute, power, cooling, and networking should be co-designed as a unified system. The design goal is to maximize what the industry calls token per watt, the amount of useful AI output produced per unit of electrical power. A DSX-aligned facility can run the same AI workloads using less electricity than a comparable facility built to uncoordinated specifications.

The architecture has four major integrated components. DSX Max-Q optimizes workload placement within fixed power budgets. DSX Flex connects AI facilities to the power grid and enables dynamic adjustment of consumption in coordination with onsite generation. DSX Air creates high-fidelity digital models of the facility's GPU clusters and network infrastructure before construction begins. DSX Exchange integrates compute, network, energy, power, and cooling signals between IT and operational systems so the facility can respond to real-time demand changes.

The most distinctive element is DSX's digital twin capability. Before a server rack is installed, the entire facility can be simulated in NVIDIA's Omniverse platform. Operators model the GPU clusters, test cooling scenarios, simulate electrical loads, and identify bottlenecks before construction decisions become physical and irreversible. For a facility representing billions of dollars in capital expenditure, the ability to validate design choices in simulation is a meaningful risk reduction tool.

NVIDIA's DSX framework has broad industry support from Cadence, Siemens, Schneider Electric, PTC, Procore, Eaton, Vertiv, and GE Vernova, among others. That breadth matters because it means equipment from these vendors ships with SimReady models and validated specifications that integrate into DSX without custom engineering work. For IREN, adopting DSX as its infrastructure standard isn't a branding choice. It's a commitment to building facilities the way the AI accelerator market's dominant supplier specifies they should be built, with the full validation framework and vendor ecosystem behind it.

The flagship deployment for this partnership is IREN's Sweetwater campus in West Texas, spanning two sites across approximately 2,200 acres. Sweetwater 1 hit a major milestone on May 1, 2026, exactly one week before the NVIDIA partnership was announced. IREN completed energization of the site's 1,400-megawatt bulk substation, a 345kV to 138kV interconnection that gives the facility access to grid-scale power. The substation represents years of infrastructure preparation: IREN secured 2.75 gigawatts of grid connection agreements in West Texas in early 2025, then executed substation construction while simultaneously building out data center facilities.

When fully operational, the Sweetwater campus is designed to support more than 700,000 liquid-cooled GPUs. Sweetwater 2, the second site with an additional 500 acres, is expected to reach energization in 2028. A direct fiber loop between the two sites ensures low-latency connectivity across the full campus.

West Texas is chosen for practical reasons that apply specifically to AI infrastructure at scale. The region has substantial wind and solar generation capacity, which keeps power costs competitive and helps operators meet sustainability commitments that cloud and AI customers increasingly require. Land is available at industrial scale. The ERCOT grid, which covers most of Texas, has significant generation capacity even as industrial demand grows.

IREN's institutional background is cryptocurrency mining, a business with nearly identical infrastructure requirements: access to cheap power at scale, continuous high-intensity GPU operations, and facilities engineered for uninterrupted compute loads. The company pivoted to AI cloud infrastructure as that market matured. The Sweetwater campus is the largest single expression of that transition. For NVIDIA, it provides a proven, grid-connected deployment site for Vera Rubin GPUs at a scale that few independent data center operators can currently match. Having a 1.4GW facility that just completed its power infrastructure, operated by a company now financially aligned with NVIDIA's success, is a faster path to deploying DSX-aligned capacity than development from scratch.

The Warrant Deal and IREN's Software Play

A warrant gives NVIDIA the right to buy IREN shares at $70 per share over five years. If IREN executes well, shares appreciate and NVIDIA's warrants gain value. If IREN underperforms, NVIDIA doesn't have capital locked in a losing equity position. For NVIDIA, the structure provides financial upside without requiring upfront cash. For IREN, issuing warrants avoids immediate dilution. Existing shareholders are only diluted if NVIDIA exercises, which requires IREN's performance to be strong enough that the $70 strike price is attractive. The $70 per share exercise price and the 30 million share cap give the warrants a maximum investment value of $2.1 billion. At exercise, that would make NVIDIA a significant IREN shareholder.

The financial alignment this creates is more interesting than the dollar figure alone. When NVIDIA holds warrants in an infrastructure operator, it has a direct financial incentive to help that operator succeed. Steering DSX adoption toward IREN, prioritizing Sweetwater for early Vera Rubin deployments, and supporting IREN's enterprise customer acquisition all become NVIDIA's interests as much as IREN's. This is structurally different from NVIDIA's typical customer relationships. When a data center operator buys NVIDIA hardware, NVIDIA's revenue is captured at sale. After that, NVIDIA's ongoing incentives around that operator's long-term success are primarily reputational. The warrant structure extends NVIDIA's financial interest in IREN's performance for five years, spanning at least two product generations of NVIDIA AI accelerators.

The partnership is contingent on regulatory approval, a condition both companies presumably anticipated given NVIDIA's dominant market position. Any arrangement in which NVIDIA takes a financial stake in infrastructure operators will receive scrutiny from competition authorities. The approval timeline is not disclosed, but the public announcement signals both companies' confidence in the regulatory outcome.

Two days before the NVIDIA partnership announcement, IREN completed the acquisition of Mirantis for $625 million. Mirantis is a Kubernetes and OpenStack infrastructure company with a history in managed cloud services that shifted toward enterprise container infrastructure. The acquisition matters for the NVIDIA partnership because DSX is a full-stack architecture. NVIDIA's reference design specifies not just hardware and power, but the software systems that manage how workloads are distributed across the facility. Running 700,000 GPUs efficiently requires sophisticated cluster orchestration, workload scheduling, and resource management that pure hardware operators typically outsource.

IREN now has in-house capability at exactly that software layer. For enterprise customers evaluating AI infrastructure options, the combination of gigawatt-scale power and land, NVIDIA-aligned GPU hardware, and enterprise Kubernetes orchestration makes IREN a fuller-stack option than a pure hardware hosting provider. The Mirantis engineering team also brings cloud infrastructure operational experience that extends IREN's capabilities beyond AI training workloads into the managed infrastructure services that enterprise customers frequently need alongside raw compute capacity. The timing of both announcements within 48 hours suggests the two deals were coordinated, with IREN moving to close the software gap before the NVIDIA partnership made that gap visible to the market.

AI Infrastructure Market Context and Implications

The NVIDIA-IREN partnership makes sense as one node in a capital deployment cycle that has no recent precedent. Microsoft announced plans to spend $80 billion on AI data center infrastructure during 2025. Meta committed more than $65 billion to infrastructure in 2026. Google's capital expenditure for AI infrastructure is running at approximately $75 billion for the year. The combined capital flowing into AI compute infrastructure from major hyperscalers is larger than any prior sustained technology infrastructure investment in the industry's history.

The result is persistent demand across the full supply chain: GPU manufacturers, power equipment suppliers, fiber and networking companies, land and real estate operators, and electrical infrastructure builders. SiTime, which produces precision timing semiconductors critical for data center operations, reported 88% revenue growth in its latest quarter, driven primarily by AI infrastructure demand. Akamai's shares rose 26% this week after the company announced a $1.8 billion AI infrastructure deal. These numbers confirm that the AI infrastructure buildout is producing demand signals across every layer of the supply chain, not just at the GPU and software level.

Underneath the capital numbers is a real constraint that limits how fast even well-capitalized players can move: the combination of power-secured, grid-connected land at gigawatt scale is genuinely scarce. Finding locations with both the electrical capacity and the physical land for gigawatt-scale AI facilities takes years, involves complex grid interconnection negotiations, and requires operational expertise that most organizations don't maintain internally. IREN's position in West Texas, with 2.75 gigawatts of secured grid connections and an energized 1.4GW substation, addresses exactly that bottleneck. The scarcity of grid-connected AI infrastructure capacity at gigawatt scale is what makes IREN's assets valuable to NVIDIA, not just the operator relationship.

For a deeper look at how compute access, cloud choice, serving design, and capacity constraints are shaping AI deployment decisions today, the AIntelligenceHub AI Infrastructure guide covers the technical and strategic dimensions of the current buildout.

For organizations evaluating AI infrastructure options in 2026, the NVIDIA-IREN partnership adds a new operator to a field dominated by hyperscalers. IREN positions itself between two existing models: on-demand hyperscaler GPU capacity with no capital commitment, and self-owned infrastructure requiring hardware purchase and operations. IREN's model is reserved GPU capacity in dedicated clusters, at data center scale, with software-managed orchestration now handled in-house through the Mirantis acquisition. For organizations running sustained, intensive AI training workloads, that model has a specific economic argument. Hyperscaler GPU pricing for large dedicated clusters has been elevated as demand outstrips available capacity. Companies that need reliable access to large GPU clusters, without the capital expense of hardware ownership, sometimes find that reserved capacity from infrastructure-specialized operators offers more cost predictability than on-demand hyperscaler rates.

The DSX alignment matters for enterprise teams because it means IREN's facilities will be built to the same architecture that NVIDIA and its software ecosystem partners have validated. Workloads optimized for NVIDIA hardware and the DSX power and thermal specifications will run on IREN infrastructure the way they were designed to run, without the operational variability that comes from facilities built to non-standard configurations. NVIDIA's financial stake also functions as an informal quality signal. When NVIDIA holds warrants in an infrastructure operator, it's expressing confidence in that operator's ability to build and maintain DSX-compliant facilities at the quality level NVIDIA's enterprise customers expect. In a market where buyers struggle to evaluate infrastructure quality before committing, that signal carries real informational weight.

The IREN deal fits a pattern NVIDIA has been building across the AI hardware ecosystem: creating financial and architectural alignment around the full AI infrastructure stack, not just the chip layer. By establishing DSX as the reference architecture and taking financial positions in operators who adopt it, NVIDIA creates a cycle where broader DSX adoption increases the value of DSX compliance, which attracts more operators to the standard. The connection to NVIDIA's enterprise software partnerships is worth noting. ServiceNow and NVIDIA's Project Arc established an enterprise agent governance framework built on NVIDIA hardware. The IREN partnership builds the physical infrastructure layer beneath that governance stack. NVIDIA is filling in the AI stack from enterprise software applications down to physical power and land, one strategic relationship at a time.

Jensen Huang, NVIDIA's CEO, put the strategic rationale directly in the partnership announcement: "AI factories are becoming foundational infrastructure for the global economy. Deploying these systems at scale requires deep integration across the full stack." That framing matters. Huang isn't describing a data center deal. He's describing an infrastructure category that NVIDIA intends to architect and, through financial instruments like this warrant structure, to partially own.

The Sweetwater 1 energization on May 1 was the immediate operational prerequisite for the partnership's credibility. With the 1.4GW substation live, the first phase of DSX-aligned deployment can begin. The five-year warrant term gives NVIDIA flexibility to exercise as IREN executes on capacity additions. Sweetwater 2, expected to energize in 2028, adds the next major tranche of Texas capacity. IREN's European and APAC locations extend geographic reach for customers who need compute infrastructure outside the United States.

The combined scale of what both companies have committed to is worth stating plainly: $2.1 billion in potential NVIDIA investment, up to 5 gigawatts of capacity across three continents, over a five-year buildout timeline. By any prior standard for data center infrastructure deals, those numbers would have been extraordinary. By the standards of the AI infrastructure buildout currently underway, they're a large but not unprecedented expression of where the market is heading.

What the NVIDIA-IREN deal signals most clearly is that AI infrastructure is entering a consolidation phase. The operators who succeed at gigawatt scale will be those who combine secured power and land, architectural alignment with the dominant hardware supplier, in-house software capability, and direct customer relationships with the AI companies that define the compute workloads. IREN, with the Mirantis acquisition and the NVIDIA warrant deal executed within 48 hours of each other, has made a clear argument that it intends to be one of those operators. Whether the Sweetwater campus delivers on that argument is the operational question the next several years will answer.