At Goldman Sachs' annual technology conference earlier this month, NVIDIA's chief executive stated that the company is tracking every gigawatt-scale plot of land, power supply, and data center shell across the globe, adding that the firm is aware of the location of virtually all relevant resources. The reason for such intense scrutiny of power availability? NVIDIA views electricity supply as a fundamental bottleneck, as the speed at which AI server chips can go live after shipment is heavily dependent on power conditions. As previously reported, securing electricity is the first step for any AI data center project, taking priority even over land acquisition or chip procurement. Data center developers collaborating with companies like Google, Microsoft, and Oracle are concerned that future power shortages could result in a vast number of chips sitting idle.
NVIDIA's recent series of land and power deals indicates that the company is leveraging its own balance sheet to ensure that limited electricity resources flow toward data centers deploying its chips, rather than projects powered by competitor products. However, due to power supply constraints and uncertainties around project timelines, NVIDIA and its major chip customers have shifted away from placing all their bets on multi-billion-dollar, gigawatt-scale mega projects. Today, smaller facilities and fragmented power resources are gaining greater appeal among investors and data center developers.
Pivot Toward Smaller-Scale Buildouts
Two data center industry executives noted that due to community opposition and regulatory approval hurdles, there is significant uncertainty about when power resources can actually be connected. As a result, spreading investments across numerous smaller projects has become critical. An NVIDIA executive revealed earlier this month that the company is going all-in to address potential power shortages, with modular and ultra-small data centers being one of the solutions. For instance, data center developer Crusoe announced this week that it has completed a new $3.9 billion funding round, with NVIDIA as an investor, to expand production of containerized micro-modular data centers that use only 1 megawatt of power. These smaller facilities can come online much faster than large projects and can also repurpose existing buildings like old offices or industrial warehouses that already have power access.
NVIDIA's senior director of high-performance computing and AI infrastructure stated that, like many of the emerging cloud providers it supports, the company is scouting for plots of land with existing idle power, with a particular focus on Nordic countries and Texas, to help customers secure reliable locations. He added that NVIDIA can act as a matchmaker if clients need power resources. Leveraging its comprehensive tracking of globally available power, NVIDIA can connect customers with power suppliers or data center developers. At times, NVIDIA also directly leases existing facilities to lock in computing space for its clients. The latest financial disclosures show that NVIDIA has signed data center leases totaling $20 billion, with plans to gradually sublease them to customers in the future.
As power shortages become increasingly acute, companies are also relying on software and hardware innovations to improve the efficiency of existing electricity usage, which was a core theme at last week's AI infrastructure summit in Santa Clara, California. NVIDIA's vice president of hyperscale and high-performance computing introduced the new MaxLPS power management system, which includes scheduling software, sensors, and power controllers inside server racks to smooth out computing power draw and avoid instantaneous power spikes. According to the executive, MaxLPS allows customers to deploy 40% more GPUs under the same power allocation. Last week, NVIDIA, Google, and Emerald AI, together with Anthropic and the U.S. National Grid, formed an alliance to advocate for grid-flexible data centers. Power flexibility refers to the ability of AI data centers to automatically pause or reschedule non-critical computing tasks when the grid faces peak demand events like heatwaves and needs to prioritize residential power supply. The alliance aims to push states to tighten restrictive policies on data centers.
Energy-efficient hardware innovation is also advancing steadily. Conversations with several industry executives last week revealed broad optimism about lower-power optical switches and more energy-efficient data center networking equipment. Enphase, a residential solar company with a 20-year history, stated that its solid-state transformers can also be applied to AI data centers, reducing power losses during the conversion from medium-voltage AC grid power to the low-voltage DC power required by AI server rooms.
Rising Market for Memory Resale
While power is a looming concern over the industry, memory hardware shortages have persisted throughout 2026, which has also pushed up prices for NVIDIA's AI devices that depend on memory. According to the latest financial filings, NVIDIA has committed roughly $280 billion to lock in capacity from memory and computing chip makers, a scale of investment that few other companies can match. But how do other market participants secure memory? When investment capital is abundant but supply is limited, conditions become ripe for secondary resale and new types of financing instruments. At last week's AI infrastructure summit, one chip design executive noted that several people approached him claiming to have immediate inventory of high-bandwidth memory (HBM) in warehouses. A credit industry executive also recently indicated that lenders are exploring new financing tools to acquire memory capacity and resell it for profit. Investment firm Coatue Management and chip startup MatX are advancing a multi-billion-dollar chip financing project, though it remains unclear whether the capacity from that project would be allowed for external resale. If memory capacity becomes an asset for financial speculation, the three leading high-end memory makers, Samsung, SK Hynix, and Micron, would likely not welcome that development.