How Nvidia Is Trying to Solve the Data Center Power Bottleneck

Nvidia is tracking “every single gigawatt of land, power and shell around the world, literally everything on the planet,” CEO Jensen Huang said at Goldman Sachs’ annual tech conference earlier this month. “We know where everything is.”

Why track power so intently?

Nvidia sees power availability as a key barrier to getting its AI server chips online as soon as they ship, as we’ve chronicled in this column over and over. While securing power is the first step to any AI data center project, even before getting the land or chips, developers that work with Google, Microsoft, Oracle and others fear a glut of unused chips due to an eventual power shortage.

Nvidia’s recent investments in power-related companies like Cloverleaf, SB Energy and Lancium show it’s using its balance sheet to ensure that available power is directed to data centers that will be filled with its chips and not those of its rivals.

But Nvidia and large users of its chips have also expanded their focus from massive, gigawatt-scale projects that could cost hundreds of billions of dollars, due in part to power availability concerns and unpredictable timelines.

These days, smaller facilities and pockets of power are becoming attractive to investors and data center builders alike.

Going Smaller

Two data center executives said it’s impossible to know when exactly power will come online, in part because of community pushback and regulatory hurdles, so it’s important to diversify and bet on a large number of smaller facilities.

An Nvidia executive told me earlier this month that the company is “working very hard” to solve potential power shortages, and one way to do so is adding “modular” or super-small data centers. Data center developer Crusoe, for instance, said this week it has raised $3.9 billion in fresh funding—including from Nvidia—to help increase production of its tiny, 1 MW, shipping-container–size modular data centers.

Such facilities are easier to turn on faster than big ones and can involve refurbishments of old office buildings or manufacturing facilities that already have power.

Like many of the young cloud providers it supports, Nvidia is spending time looking for places where untapped power already exists, particularly in the Nordic countries and Texas, to help its customers find reliable sites, according to Dion Harris, Nvidia’s senior director of high-performance computing and AI infrastructure.

He said if a customer is looking for power, Nvidia can play matchmaker. It can make an introduction to the power provider or data center developer, since it tracks all the world’s available power.

Sometimes Nvidia itself leases existing facilities to lock them down for customers. In its latest quarterly report, the company disclosed $20 billion in data center leases it had signed and intends to pass off to customers over time.

With concerns about power on the rise, firms are also turning their attention to software and hardware innovations to better utilize existing power, a key theme at last week’s AI Infra Summit in Santa Clara, Calif.

Ian Buck, Nvidia’s vice president of hyperscale and high-performance computing, talked up Nvidia’s new MaxLPS product, a power-management system for AI data centers that includes scheduling software, sensors and power controllers that sit inside server racks to spread out power needs more evenly and avoid concentrated power spikes. MaxLPS could help customers use 40% more graphics processing units with the same amount of power, Buck said.

Last week, Nvidia, Google and Emerald AI also announced an alliance with Anthropic, National Grid and others to advocate for power-flexible data centers that are hooked up to the electric grid. Power flexibility means that if the grid needs more capacity for consumers, as during a heat wave, the AI data center could automatically pause or reschedule noncritical AI data center work. The goal would be to convince states to slow pushback against grid-connected data centers.

Power-saving hardware innovations could be on their way, too. I had several conversations last week with executives betting on less power-hungry optical switches and more power-efficient data center networking. And a 20-year-old consumer solar company, Enphase, says its solid-state transformers can be used for AI data centers as well. That would help reduce power loss in the process of converting medium-voltage alternating current (the standard for utility grids) to the lower-voltage direct current that’s needed in an AI data center.

Memory Dealers

While power is a looming concern, memory hardware shortages have been the story of 2026 and have contributed to rising prices for Nvidia AI gear that depends on memory, as we have reported. Nvidia has locked up a lot of capacity with some $280 billion in commitments to memory and compute chipmakers, per its latest quarterly filing. That’s an amount almost no one else can afford.

But what about everyone else who needs memory? A market where investment dollars are aplenty but supply is limited is fertile ground for resellers and other new financing vehicles.

At the AI Infra event last week, one chip design executive told me he was surprised that several people had approached him there, offering to sell high-bandwidth memory that they had “sitting in their warehouses” and could ship immediately.

A credit executive also told me recently that lenders have been exploring financing vehicles to buy and resell memory capacity for a profit. (Investment firm Coatue Management and chip startup MatX are discussing such a venture, though it’s unclear yet whether or how capacity in that vehicle can be resold.) I’m not sure the three dominant makers of high-end memory, Samsung, SK Hynix and Micron, will be too thrilled if memory capacity becomes a financeable asset that others can profit from.