Consider the circuit. AI infrastructure is no longer merely a collection of servers fitted with faster processors; it is a new infrastructure category governed by power, thermal, optical, mechanical and financial constraints. The emerging “AI factory” combines accelerators, CPUs, memory, networking, optical interconnects, power conversion, grid connections, batteries, generation, cooling, software orchestration and specialized financing 81. NVIDIA remains the central compute-platform beneficiary, but the economic opportunity is broadening well beyond GPUs.
The claims considered here were concentrated between 28 July and 10 August 2026, with most evidence published in early August. Corroboration is generally limited because most claims rely on a single source. Several important conclusions—AI-infrastructure investment, Oracle’s Bloom Energy commitment, Oklo’s nuclear thesis, Teradata’s repositioning and the immersion-cooling opportunity—have two sources and therefore deserve greater weight 45,47,48,62,82. The conclusion for NVIDIA is constructive but not simple: accelerating cluster deployment supports demand for NVIDIA systems, while constraints in power, cooling, networking and finance increase the value of a complete platform and create openings for AMD, custom ASICs and alternative accelerator ecosystems.
From AI Demand to Physical Buildout
The AI factory is a system, not a chip
Next-generation facilities require GPU accelerator clusters, high-performance computing, enterprise storage, high-speed networking, cloud platforms, intelligent power distribution, advanced cooling, cybersecurity and monitoring 56. Servers pair CPUs with specialized accelerators such as GPUs 24, yet CPUs remain essential for orchestration, scheduling, retrieval, data movement, security, preprocessing and general-purpose workloads surrounding the accelerator cluster 38. NVIDIA’s opportunity is consequently larger than discrete GPU sales. Integrated rack-scale systems, networking, software and cluster orchestration become increasingly important as buyers deploy coordinated pools of compute rather than isolated components.
The optical layer is expanding as hyperscalers connect accelerator capacity into large clusters 6. Operators are also addressing rack-level orchestration, fault tolerance and predictive maintenance 42. This is the proper Gestalt of the system: compute performance depends upon the reliable movement of power, data and heat. A fast accelerator attached to a weak power-delivery network is rather like a powerful engine bolted to a bridge of uncertain strength.
The supply chain reflects this physical reality. Before an AI facility can operate, developers must procure steel structures, precast concrete, transformers, high-voltage switchgear, generators, industrial cooling, fiber, copper, liquid-cooling equipment, server racks, GPUs, power-distribution units, batteries and security systems 34. Transformers must arrive before servers can be installed 34, and logistics providers subsequently move power-distribution units and liquid-cooling systems 34. Medium-voltage switchgear, large transformers, copper, cold plates and liquid-cooling piping are among the critical components 66.
This breadth explains why power, cooling and electrical-equipment suppliers can benefit regardless of which accelerator or cloud provider prevails 7. Companies identified as beneficiaries include GE Vernova, Vertiv, Eaton, Quanta Services and Celestica, among other infrastructure providers 14,15,50. The capital requirement is substantial 55. Reported projects include a 10-gigawatt Ohio campus 80, a proposed $5.1 billion facility in Salem, Oregon 17, a 2-gigawatt Chinese campus with one million accelerators 79 and xAI’s Colossus cluster in Memphis, described as the largest AI-training facility by compute capacity and IT power 69.
OpenAI is reportedly leasing a 10-gigawatt project in southern Ohio developed by SoftBank’s energy subsidiary 2, with OpenAI expected to serve as anchor tenant 80,86. The wider Stargate plan involving OpenAI, Oracle and SoftBank has been described as approaching $500 billion of investment 61,84. The proposed Ohio transaction would support leasing of a $500 billion, 10-gigawatt hub developed by SB Energy 35. These figures are reported claims, not independently validated evidence of funded NVIDIA purchase orders. The distinction is material.
Power Availability as a Deployment Constraint
Dedicated generation and private power
The operating model increasingly combines hyperscale campuses, high-density computing, advanced cooling, renewable contracts, storage, onsite generation, flexible load management and potentially nuclear or small modular reactor supply 24. Developers are considering colocating data centers directly at nuclear facilities 74. Reliable nuclear, gas or other dispatchable generation near AI clusters could benefit Constellation Energy, Vistra and Talen Energy 41.
Oklo’s investment thesis is explicitly tied to reliable electricity demand from AI and hyperscale facilities 62, with its Aurora Powerhouses intended to supply those customers 62. Oklo is developing advanced nuclear facilities for long-term AI demand, although its concentration creates customer and end-market dependency risk 62. Vistra’s Helix Digital Infrastructure and broader Helix strategy likewise target AI-related electricity demand 60.
Oracle illustrates the movement toward private, dedicated power. It has contracted for up to 2.8 gigawatts of Bloom Energy capacity 48,72, intended for its cloud and AI services 72. Bloom reportedly delivered power to an Oracle data center within 55 days 37, and nearly half of Oracle’s fuel-cell projects were underway 72. Oracle has separately identified natural-gas fuel cells as part of its data-center power strategy 37.
The New Albany, Ohio project combines AI-computing facilities with dedicated generation or distribution infrastructure 54 and reflects the increasing use of privately developed and operated power infrastructure 54. California Resources Corporation is also pursuing hyperscalers seeking reliable power, although its strategy depends on both hyperscaler demand and successful data-center development 67.
Grid interconnection and site control
Grid availability, favorable regulation and rapid interconnection may become location-level competitive advantages 5. ERCOT’s delayed transmission study could affect project timing and viability 21. The opportunity therefore extends beyond generation to transmission lines, substations, grid upgrades, power distribution and cooling 14.
Former cryptocurrency-mining platforms may possess advantages in power procurement, interconnection, site control and high-density electrical infrastructure 89. These characteristics help explain conversions and greenfield strategies at companies such as Hut 8 and Galaxy Digital 1,46. Brownfield redevelopment can shorten deployment timelines, but the electrical and thermal systems must still be examined rather than assumed serviceable. Is this truly negligible, or have we missed a coupling between the old site and the new load?
Power Architecture and Thermal Management
Liquid cooling is now part of the electrical design
Direct liquid cooling is being adopted for AI data-center infrastructure 77. AI facilities require liquid cooling, high-density racks, extensive fiber, specialized batteries and dedicated power systems 34. Immersion cooling is described as the fastest-growing end-use segment of the global immersion-cooling market, with colocation operators able to sell it as a premium service for high-density AI workloads 82.
Vertiv’s primary AI exposure is critical power and liquid cooling, supplying power and thermal-management systems inside increasingly dense data halls 14. At gigawatt scale, efficient liquid cooling requires site selection, power architecture and thermal-system design to be coordinated from the earliest planning stages 66. Customized power, cooling, cabling, structural and spatial features substantially increase construction costs relative to conventional air-cooled facilities 87. Heat is not a footnote to the power budget; it is the power budget’s physical consequence.
The move toward 800VDC
Traditional AC facilities perform multiple conversion steps before electricity reaches DC computing equipment 37. An 800VDC architecture can reduce power requirements, copper usage and transmission losses 85. Adoption depends on new construction, retrofit economics, hyperscaler standards, utility availability and rising accelerator power 44. Such systems require more sophisticated conversion, higher-voltage handling, improved efficiency and thermal management 65, while improved power conversion can reduce electricity losses and heat 53.
A forecast that low-voltage transformer demand could decline from 55,000 MVA to 30,000 MVA between 2027 and 2030 suggests potential mix risk for legacy equipment 85. It is, however, an isolated forecast and should not be generalized across the entire electrical-infrastructure market. The proper Ansatz is to examine each voltage class, site design and retrofit population separately.
The implications are competitive as well as mechanical. Customers deploying AMD’s MI450 and Helios architecture must prepare compatible power-delivery and cooling systems 83. Oracle is reportedly planning a Helios supercluster using AMD MI450 GPUs and EPYC Venice CPUs 63, and the planned Helios supercluster is explicitly based on AMD technology 63. These claims do not establish broad substitution away from NVIDIA. They do show that hyperscalers will architect around multiple accelerator platforms when power, availability, cost or workload requirements justify it.
Transitions across 400V and 800V, solid-state transformers, liquid versus air cooling, Ethernet versus proprietary fabrics and optical technologies could redistribute share or obsolete products 5. NVIDIA’s system-level position will therefore depend increasingly on compatibility, integration and performance per watt, not merely on peak accelerator performance.
Networking, Optics and Memory Bottlenecks
GlobalFoundries’ co-packaged-optics project could improve bandwidth and energy efficiency in AI clusters, hyperscale data centers and HPC installations 36. Co-packaged optics, low-loss materials, thermodynamic computing and improved memory could allow compute capacity to expand without electricity and communication requirements rising proportionally with processing power 36. Optical replacement of conventional data-center wiring is another developing technology 79, and companies throughout the optical-infrastructure supply chain are benefiting from AI-data-center demand 88.
Amkor’s planned Arizona packaging facility is intended to serve AI, HPC, communications, automotive and industrial applications 32. Cohu’s AI/HPC test customers include fabless companies and hyperscalers developing proprietary GPUs, custom ASICs and network processors 40. The implication is plain: the limiting element may be bandwidth, packaging or test capacity rather than the accelerator die itself.
Memory and system balance matter alongside accelerator counts. AI data centers may use fewer accelerators per unit of useful work while requiring substantially more memory around each accelerator 23. Cloud accelerator programs make data-center accelerators an important reference point in evaluating GPU-class design assets 22. Intel unveiled its next-generation Jaguar Shores accelerator on 7 August 2026 58, while retaining an established Xeon server CPU position and benefiting from agentic-orchestration demand 28. DigitalOcean maintains a multi-accelerator strategy and had preallocated AI capacity 7.
The direction of travel is toward hardware diversity, cluster management, workload scheduling and software orchestration rather than single-chip optimization alone 73. This favors vendors that can make heterogeneous clusters behave as one dependable machine.
Demand, Financing and Counterparty Risk
Oracle, OpenAI and the conversion problem
Oracle is investing billions in AI infrastructure, data centers, cloud capacity and energy systems 34. Its relevant business includes OCI and the design, construction and operation of large-scale facilities for AI training and inference 61. OCI targets database-driven AI, advanced analytics and mission-critical workloads; it offers NVIDIA GPU instances and supports enterprise priorities such as security, sovereignty, data integrity and auditability 29. Its GPU portfolio has included H100, A100, A10, V100 and P100 instances 29, while additional GPU services are intended to support scalable enterprise workloads 30. Oracle’s customer base is primarily corporate, and the company is attempting to move from enterprise software and cloud applications toward hyperscale infrastructure 4.
Oracle also supports OpenAI’s cloud capacity and the Stargate initiative 8. OpenAI is reportedly party to a five-year, approximately $300 billion cloud agreement linked to Stargate 8. Oracle announced a large OpenAI-related agreement in late 2025 involving $300 billion of remaining performance obligations 26, although reporting also described a purported $300 billion contract or offer more cautiously 4. The claims establish substantial reported commercial linkage, but the exact contractual status, timing, economics and conversion into delivered GPU demand remain uncertain.
Oracle is adding data centers alongside the operational Abilene campus 61 and may be attempting to build faster than competitors to meet compute demand 4. Yet the investment case must distinguish announced capacity from energized capacity, installed racks, GPU acceptance and recurring workload utilization. A nameplate gigawatt is not a revenue line.
Financing structures can delay otherwise durable demand
The Ohio proposal involves banks and investors financing construction, SB Energy building the site, OpenAI leasing computing capacity and NVIDIA supplying accelerators 86. AI facilities can use private credit, bank debt, equipment finance, GPU-backed lending, vendor financing, sale-and-leasebacks and hybrid real-estate/infrastructure structures 87. Hyperscaler tenants may provide contractual support 89.
The financing system has limits. Banks may have limited capacity to finance additional Oracle-, OpenAI- or lease-related exposure 59, and single-counterparty limits may restrict expansion 59. Banks have attempted to reduce exposure through discounted loan sales and risk transfers 59. Data-center leases connected to the OpenAI–Oracle structure may consequently come under stress 59.
These risks need not diminish long-term AI demand, but they can delay construction, defer accelerator purchases and increase working-capital or credit risk across the supply chain. The principal financial risk is therefore not necessarily a collapse in interest; it is a mismatch between announced capacity and funded, operational demand.
Oracle’s cloud offering is reported to lag AWS, Azure and Google Cloud 3, while competition also includes Microsoft, Amazon, Google, NVIDIA and other providers 4. Oracle may mitigate underutilization risk by repurposing facilities 4, and previous conversions of Bitcoin-mining facilities into AI data centers provide precedent 4. Repurposing is not frictionless: facilities built for older rack densities may require substantial modification for new AI systems 57, and long-term value depends on the ability to absorb future processors and accelerators 57. Facilities designed for flexible power, cooling, modularity and successive hardware generations should command more durable economic value than narrowly optimized sites 57.
AI-Optimized Capacity Versus Conventional Colocation
AI-optimized facilities share hyperscale sites’ large power requirements and campus configurations, but differ in design, revenue model, counterparty profile, technology exposure and financing constraints 87. They sell processing capacity rather than simply leasing physical space 87, and an operator may own both the facility and the GPUs within it 87. This arrangement creates greater upside when utilization and pricing are strong, but also exposes operators to GPU depreciation, technology obsolescence, tenant concentration and financing risk.
Applied Digital generally leases buildings, power and cooling infrastructure rather than purchasing GPUs 33. It has leased 1,420 MW of capacity, of which 175 MW was operational 10, and has moderate demand exposure to powered shells and dedicated capacity 7. The distinction matters for NVIDIA because GPU ownership can accelerate procurement while concentrating risk at the operator level.
On-premises or colocated AI hardware may offer lower long-run cost for steady, heavy, latency-sensitive or data-sensitive workloads, but customers bear power, cooling, maintenance and depreciation costs 9. Distributed locations can support regional demand, lower latency, customer proximity and resiliency 7. Remote greenfield campuses near generation assets are particularly suited to concentrated training 37, whereas latency-sensitive inference, enterprise AI, sovereign workloads and data-residency-sensitive applications may require facilities near users and network interconnection points 37. Oracle’s infrastructure connects on-chain execution with external data used in real-world decisions, illustrating how workloads may extend beyond conventional cloud compute 68.
The broader opportunity spans hyperscale deployment, colocation, hosting, AI/HPC campuses and specialized platforms 19,79. EMCOR’s growth is concentrated in AI data centers, networking, water and wastewater, healthcare, institutional work and reshoring infrastructure 39. BorgWarner is pivoting toward data-center and industrial markets, targeting turbine generators, batteries and inverters 51. IREN’s model centers on securing power and converting it into high-density AI capacity 66. Identified AI-infrastructure beneficiaries include Oracle, CoreWeave, Vertiv, Eaton and Constellation Energy 70. Sector rotation has moved toward the physical buildout layer, particularly cooling and photonics 71, supporting a broader basket approach rather than exclusive focus on accelerator designers.
Software, Sovereignty and Longer-Duration Concepts
Teradata is repositioning from a legacy on-premises data-warehouse vendor into a hybrid and sovereign AI platform 47, although its AI pivot may remain aspirational 47. OVHcloud emphasizes energy-efficient facilities 29, while cloud and infrastructure strategies increasingly stress sovereign, private and secure environments. Customers requiring proprietary model weights, continuous fine-tuning, reinforcement learning, evaluation, low-latency inference, classified operations or isolation from third-party providers may need private clouds, secured colocation, customer facilities or dedicated public-cloud instances 43. These requirements can support NVIDIA systems outside the largest hyperscalers, but may also favor local or proprietary accelerators.
AI-enabled operating systems and forecasting tools are presented as methods to improve operations, reduce costs, thermal instability and reliability 25,56. Expected efficiency gains could reduce permanent data-center staffing over time 20, while technology obsolescence and efficiency could alter staffing and infrastructure needs at individual facilities 20. Traditional data centers remain primarily CPU-centric and focused on enterprise applications, databases and cloud services 81; AI infrastructure is more accelerator-intensive and thermally constrained. NVIDIA’s software and systems strategy should benefit if it makes heterogeneous clusters easier to manage, but the execution requirement extends well beyond silicon performance.
Orbital data centers are an emerging but speculative extension of the theme. The concept uses space-based computing and solar energy to reduce conventional cooling requirements 11 and could reduce latency or bandwidth needs for data generated in space 12. Proposed architectures depend on commercial off-the-shelf accelerators and radiation-tolerant FPGA components 16. The efficiency and cooling benefits remain theoretical rather than demonstrated economics 78, and orbital systems create launch, manufacturing, thermal, power and debris challenges 52. Data centers are expanding into orbital concepts 49, but orbital computing is best treated as a long-duration option rather than a current NVIDIA revenue driver 13,49.
The physical deployment of AI also carries policy and societal implications. Central Ohio facilities are described as infrastructure supporting AI, automated management, surveillance and defense applications 75. A center initially approved for routine computing could later support AI, surveillance, government, military, intelligence or cyber operations 64. A secure inference data center is intended for emergency hosting of strategically important or dangerous AI capabilities 76. These claims do not directly change NVIDIA’s near-term financial outlook, but they underscore regulatory, export-control, security and reputational considerations surrounding a foundational supplier.
Implications for NVIDIA
The cluster supports a structurally positive thesis. AI deployment is moving from model experimentation toward large-scale, capital-intensive infrastructure, and each additional cluster requires accelerators, CPUs, memory, networking, power-management hardware, cooling and software. Hyperscalers refresh accelerator platforms frequently and purchase systems in large clusters 31. This favors vendors able to deliver complete, validated architectures rather than stand-alone chips. NVIDIA’s established ecosystem remains strategically important because the bottleneck is increasingly the coordinated operation of the entire cluster.
The strongest investment implication is that NVIDIA should be analyzed as the anchor of an AI-infrastructure stack, while recognizing that much incremental spending accrues to adjacent suppliers. Critical power, liquid cooling, optical connectivity, transformers, substations, batteries, fuel cells and dispatchable generation are becoming prerequisites for GPU monetization. Vertiv, Eaton, Quanta, GE Vernova, Bloom Energy, nuclear developers and specialist contractors may therefore provide more direct exposure to physical constraints than some software-oriented AI beneficiaries. The claim that power, cooling and electrical equipment are required regardless of the accelerator winner is particularly important for portfolio construction 7.
Competitive risk is increasing. AMD’s Helios and MI450 roadmap, Intel’s Jaguar Shores, proprietary GPUs and custom ASICs from hyperscalers, and multi-accelerator procurement strategies all indicate that customers are seeking hardware diversity 7,40,58,63. NVIDIA’s defense is strongest when software compatibility, networking, systems integration and performance per watt outweigh chip-level price competition. It is more exposed where customers can standardize workloads across alternative accelerators. The move toward 800VDC, liquid cooling, optical interconnects and higher memory density means that system compatibility and time-to-deployment will influence purchasing decisions as much as benchmark performance.
NVIDIA also benefits from geographic and architectural diversity. Training can be concentrated in remote campuses near reliable generation, while inference and sovereign workloads require distributed, connected sites 37. Brownfield redevelopment, mining-site conversions and privately powered campuses can shorten deployment timelines, whereas delayed transmission studies, permitting and utility availability can defer projects 18,21,27. The Ohio project’s redevelopment of a former uranium-enrichment and industrial site illustrates how legacy energy assets may be repurposed for AI infrastructure 18,27. This flexibility expands the addressable market but makes project timing and local regulation central variables.
The practical measure of progress is therefore not the announced gigawatt figure. Investors should track the conversion of projects into powered, cooled and utilized clusters; monitor alternative-accelerator adoption and customer-specific architectures; and assess whether NVIDIA continues to capture system-level economics as AI facilities evolve into flexible, multi-generation platforms. The evidence is predominantly single-source and forward-looking. The infrastructure thesis is robust at the thematic level, but individual projects and counterparties require considerably more skepticism.
Practical conclusions
- AI spending is broadening into complete, high-density infrastructure. NVIDIA remains the central accelerator platform, while power, cooling, optics, memory and grid suppliers are increasingly material beneficiaries 7,34,56.
- Reported Stargate and Oracle–OpenAI commitments imply substantial potential GPU demand, but financing constraints, counterparty concentration and stressed lease structures may delay conversion from announced capacity to revenue 8,35,59.
- AMD, Intel, custom ASICs and multi-accelerator strategies are credible competitive signals. NVIDIA’s defense increasingly depends on systems integration, networking, software orchestration and performance per watt rather than GPU availability alone 40,58,63,73.
- Investors should prioritize energized megawatts, installed accelerator clusters, cooling readiness, interconnection progress and utilization over headline project size, while favoring infrastructure able to accommodate multiple hardware generations 21,57.