When OpenAI and Nvidia announced a strategic partnership involving 10 gigawatts of GPU-driven infrastructure, the public headlines mostly focused on chips, scale, and ambition. But lurking behind that technical grandeur is a fundamental constraint: access to electricity. Power isn’t a side concern — it’s the linchpin without which data centers cannot run, GPUs cannot compute, and AI models cannot scale.
Utilities and energy infrastructure are already under strain, and adding tens of gigawatts of demand essentially building several city-sized loads creates profound challenges. The current U.S. grid was not designed for dense, localized AI power loads. Renewable generation, transmission capacity, permitting, and resilience all become critical risk vectors.
In this article, I’ll walk through:
-
What the Nvidia-OpenAI deal involves (scale, commitments, targets)
-
Why electricity supply and grid constraints are the single biggest bottleneck
-
How OpenAI and partners are confronting energy, infrastructure, and regulatory hurdles
-
What risks remain, in worst-case and midrange scenarios
-
What lessons this offers for the AI industry more broadly
The Nvidia OpenAI Deal: Ambition Meets Infrastructure
What Was Announced
In September 2025, Nvidia and OpenAI signed a letter of intent: Nvidia will invest up to $100 billion tied to deployment of 10 gigawatts of AI compute infrastructure. The partnership is structured so that Nvidia invests incrementally as each gigawatt of compute (and accompanying data center/power capacity) is brought online. The first gigawatt is expected to be deployed in the second half of 2026.
OpenAI will build out data center sites and infrastructure in collaboration with partners (including Oracle, SoftBank) under the “Stargate” initiative, targeting deployment of large AI data centers across U.S. sites. The total roadmap includes 7 gigawatts underway, with a goal of reaching 10 gigawatts over time.
Nvidia’s systems (GPUs, networking, servers) are central to this deployment. OpenAI will use those systems to train and serve next-generation models, making compute infrastructure a foundational asset.
The Scale of the Challenge
Ten gigawatts of infrastructure is enormous. To put it in perspective:
-
It’s roughly the equivalent load of a large metropolitan area at peak usage (e.g. New York City in summer).
-
Utilities across the U.S. are already reporting that data center growth is stressing capacity; adding 10 GW would push many grids into new demand tiers.
-
Nationally, analysts estimate that utilities will need about 60 GW of new power capacity by the end of the decade to serve new data center demand a number that dwarfs typical projections.
In short: the infrastructure (generation, transmission, substations, cooling, redundancy) must expand just to absorb the demand, before any computation even begins.
Why Power Is the Biggest Bottleneck
It’s tempting to think compute (chips, servers) is the scarce resource. But in this case, many of those components are supply-managed, scalable, or can be deferred. Electricity, on the other hand, is physical, regulated, regional, and slow to expand. Here’s why it becomes the bottleneck:
1. Grid Capacity, Transmission & Local Constraints
Power grids operate with limited headroom. Even where generation exists, local substations and transmission lines may not be sized to handle spikes in load. Many grid systems operate near capacity under high summer loads. A massive new data center cluster demands upgraded lines, transformers, switchgear, and often new transmission corridors all of which require planning, permitting, environmental review, and years of build time.
2. Generation Availability & Energy Mix
To supply gigawatts sustainably, you need stable generation sources often baseload (nuclear, gas, large hydro) or reliable dispatchables (gas turbines, energy storage) plus renewables. In many U.S. regions, renewable capacity is growing, but intermittency and grid integration challenges persist. Without firm generation, data centers risk power scrubbing, outages, or relying on expensive peaker plants.
3. Permitting, Zoning & Regulatory Delays
Power plants, transmission lines, substations, and substations require regulatory approvals and permitting, which often take years. Environmental impact studies, community objections, land acquisition, and coordination with agencies can slow progress.
4. Heat & Cooling & Power Efficiency Overhead
Running AI hardware involves cooling, power conversion loss (power usage effectiveness, PUE), redundancy, backup systems. A portion of the input electricity doesn’t go to computation it is dissipated as heat, overhead infrastructure, or overhead loss. The more efficient you are, the less overhead, but you never reach zero.
5. Peak Demand, Redundancy & Resilience
Data centers need to handle peak loads, fallback, redundancy. They can’t run at 100% load constantly without degradation risk. To ensure resilience, they need surplus capacity, backup generators, uninterruptible power, which further increases power provisioning demands.
6. Local Grid Stability & Reliability Issues
Adding large AI loads can destabilize local grids: voltage drop, frequency fluctuations, brownouts. The utility must manage more complex load profiles, sometimes requiring reactive power compensation, more robust control systems, or limiting co-loads during peak hours.
7. Energy Cost Volatility & Contracts
Electricity pricing varies regionally, time of day, and contract structure. Gigawatt loads exposed to volatile wholesale power markets may incur huge costs during peak demand periods unless the operator secures long-term power purchase agreements (PPAs) or on-site generation.
What OpenAI and Partners Are Doing to Address the Power Crisis
To overcome this immense barrier, OpenAI, Nvidia, and their partners are deploying a multi-pronged strategy. Some approaches are already in motion; others are speculative.
1. Building Data Centers in Power-Friendly Locations
OpenAI is selecting sites with favorable power availability and grid access. The Stargate project includes data centers in Texas, New Mexico, and other regions where energy supply and infrastructure are relatively robust. Locations are chosen based on available land, local incentives, and utility cooperation.
2. Private & On-Site Generation
To reduce reliance on the public grid, private generation such as natural gas turbines, fuel cells, or even micro-grids can buffer load. This reduces load spikes on the grid and can enable seamless operation during grid constraints. In past AI infrastructure efforts (e.g. “Stargate” precedents), some operators have used on-site generation to support peak loads.
3. Strategic Power Purchase Agreements (PPAs) & Long-Term Contracts
Securing long-term PPAs allows data centers to lock in electricity prices and supply. It also gives utilities certainty to invest in upgrades. In many cases, data center builders partner with utility commissions or power providers prior to construction. Some large cloud providers are doing this already.
4. Energy Efficiency, Power Capping, and Optimization
Optimizing GPU utilization, reducing idle time, using workload scheduling, and employing dynamic power capping techniques can reduce the total energy draw. Research (e.g. “Sustainable Supercomputing for AI: GPU Power Capping at HPC Scale”) shows that appropriate capping strategies can reduce energy and heat while maintaining performance.
Also, empirical studies show that an 8-GPU H100 node’s peak draw is ~8.4 kW (versus manufacturer rating of 10.2 kW) under efficient conditions. These margins matter at scale.
5. Phased Deployment & Incremental Growth
Instead of turning all 10 GW on at once, deployment can be phased, allowing grid upgrades, generation scaling, and lessons learned. Each gigawatt ramp can act as a benchmark. Nvidia’s investment model reflects that approach: investing as phases are deployed.
6. Grid Collaboration & Utility Partnerships
OpenAI and partners must engage utilities, regulators, and grid operators early, to co-plan upgrades, balancing, and demand management. Joint planning reduces risk of surprise constraints.
7. Energy Storage & Demand Shifting
Battery systems, pumped hydro, thermal storage, or other storage methods may buffer intermittent demand or shift load to off-peak hours. This can reduce peak stress. Also, models can be scheduled for lower load periods, smoothing demand curves.
Remaining Risks & Possible Failure Modes
Despite mitigation strategies, risks remain that could derail or slow deployment.
Grid Bottlenecks & Transmission Delays
Even if generation exists, without adequate transmission the power can’t reach the data center. Upgrading transmission lines often runs into permitting and land use constraints.
Permitting Delays & Community Pushback
Residents may resist new generation plants, substations, lines, or noise. Regulatory delays can slow progress by years.
Renewable Integration & Variability
If supply relies heavily on renewable generation, intermittency and variability complicate meeting constant loads. Without sufficient storage or firm backup, reliability suffers.
Cost Overruns & Price Uncertainty
Power costs may escalate. If wholesale prices spike, even locked contracts may get renegotiated. Operational budgets may be strained.
Dependence on Fossil Generation
To ensure reliability, many data centers may lean on gas turbines or diesel backup—raising emissions, regulatory, and public relations issues.
Scaling Efficiency Limits
As infrastructure grows, gains in marginal efficiency may decline. Overhead, cooling, cooling inefficiencies, and diminishing returns may reduce performance per watt.
Environmental & Sustainability Concerns
Massive power draw may strain emissions targets or sustainability goals. The optics of huge energy usage may provoke regulatory scrutiny or backlash.
Scenarios: Best Case, Mid Case, Worst Case
-
Best Case: OpenAI and Nvidia successfully phase in 10 GW over several years with grid upgrades, private generation, storage, and efficiency improvements. Deployment proceeds on schedule, compute scaling continues, rivals scramble to catch up.
-
Mid Case: Deployment is delayed months to years. Some sites suffer partial capacity (e.g. 60–80% of target). GPU utilization is throttled due to power constraints. Compute growth slows; rivals gain time.
-
Worst Case: Power limitations stall large parts of the infrastructure. OpenAI must scale back ambition, shift workloads to cloud providers, or reduce model scale. The deal risks being underbuilt, producing stranded assets.
Lessons for the AI Industry at Large
The constraints OpenAI faces are not unique. Any company attempting to build AI “factories” must contend with energy.
-
Electricity is the ultimate limiting resource — before chips, before networking, before cooling — for massive AI deployments.
-
Site selection is as important as server design — good compute hardware on bad power is worthless.
-
Energy collaboration must be baked in at the start — infrastructure, grid, utility, permitting, power design should be core, not afterthought.
-
Modular deployment wins — deploying in phases and learning helps manage risk.
-
Hybrid energy and storage are essential — relying solely on grid power is increasingly risky.
-
Efficiency matters at scale — small improvements in power draw, GPU utilization, cooling design, software scheduling multiply at gigawatt scale.
-
Sustainability will be under scrutiny — as AI grows, environmental impact, emissions, renewable use, regulatory pressure will intensify.
Power Is the Control Lever Behind AI Scale
OpenAI’s $100 billion Nvidia deal sets a bold stage. But in the unfolding narrative, the real drama may not be about how fast models train, how many GPUs spin — but whether the lights stay on. Access to power, grid capacity, energy contracts, and infrastructure risk are the hidden enablers behind every layer of AI scaling.
If OpenAI (and Nvidia) succeed in surmounting the power challenge, that will mark a pivotal moment. But if they fail — or stumble — it will be a stark reminder: in computing, the true bottleneck isn’t always tech; sometimes it’s the kilowatts behind it.
