OpenAI's Brad Lightcap: Memory Shortages and Power Constraints Threaten AI Infrastructure

OpenAI chief operating officer Brad Lightcap has named the shortage of memory chips and limited energy capacity in the US as potential bottlenecks that could slow the scaling of AI infrastructure for years to come.

Two physical limits

The race to build new data centers and deploy ever larger models is already running into hard physical constraints, and Lightcap is the first to name both publicly.

The first is the shortage of memory chips, above all HBM (High Bandwidth Memory). This is specialized memory with extremely high bandwidth, without which modern AI accelerators such as NVIDIA H100, Blackwell and AMD Instinct cannot run models at the required speed. Only three companies in the world make HBM today: Samsung, SK Hynix and Micron. Demand from OpenAI, Google, Microsoft and Meta has already outpaced production capacity. Supply queues have stretched to 12–18 months, while building new factory lines takes 2–3 years.

The second factor is the state of US energy infrastructure. American data centers already consume about 4% of the country's electricity, roughly as much as the entire state of California. Goldman Sachs analysts forecast that the load on grids will triple by 2030. Connecting new large facilities to the grid takes from 3 to 7 years: outdated equipment, an acute shortage of industrial transformers and multi-level regulatory approvals create queues that will not clear anytime soon.

Details of the shortage

Each constraint has its own specific nature:

• HBM chips — production is growing, but much more slowly than demand from AI accelerator makers
• Industrial transformers — lead times have grown from one year to two or three because of a global surge in orders
• Sites near power nodes — in key states there are literally no suitable plots left with the required connection capacity
• Electrical engineers and maintenance specialists — an acute labor shortage slows construction as much as the lack of equipment
• ISO/RTO queues — grid operators are swamped with data center applications and take years to review them

Taken together, these factors mean that even with money and political will, quickly building a large data center in the US is extremely difficult today.

Stargate and the reality of resources

In early 2025, OpenAI, together with SoftBank and Oracle, announced Stargate, an investment program of up to $500 billion to deploy AI infrastructure in the US. The first phase planned to invest $100 billion. Against the backdrop of these ambitions, the COO's statement carries particular weight: even an unprecedented financial plan does not lift the limits of the physical world.

If the HBM shortage persists, memory prices will keep rising, and the operating costs of all AI platforms will rise with them. If energy capacity is not brought online on time, some Stargate sites will be delayed by years. Lightcap is, in effect, publicly admitting that OpenAI's growth is now limited not only by algorithms and training data but also by hard physical resources: silicon and electrons.

What this means

The words of the COO of the world's largest AI company are a signal to the entire industry. Physical supply chains have become as strategic a resource as models and datasets. Memory makers and energy companies get strong arguments for accelerating capital investment. For businesses that depend on AI services, it is an important reminder: the availability and cost of tools will be determined not only by progress in the models themselves, but also by whether there is enough physical infrastructure to run them.

*Meta is recognized as an extremist organization and banned in Russia.