Nebius Power Strategy Explained: 5GW Contracted Power, Data Center Capacity and the AI Grid Bottleneck

Summary

For an AI cloud, GPUs are only useful after electricity arrives.

That sounds obvious. In 2026, it has become one of the most important constraints in the entire AI infrastructure market.

Nebius now expects to finish 2026 with approximately 5 GW of contracted power, up from more than 4 GW previously. Its Q2 shareholder letter defines contracted power as capacity supported by contracted land and power commitments.

This does not mean Nebius already has 5 GW of operating data centers.

It means the company has secured a much larger future power footprint from which it intends to build.

There is also a second “5 GW” number in the Nebius story: the NVIDIA partnership aims to enable more than 5 GW of NVIDIA systems by the end of 2030.

These are related ambitions, but they should not be treated as the same metric.

Contracted Power Is Not Live Power

This is the first distinction to get right.

A site can have:

• land;
• a power agreement;
• permits;
• construction plans

and still not be ready to run GPUs.

The progression is closer to:

Contract power

→

Build generation/grid connection

→

Construct data center

→

Install electrical/cooling equipment

→

Install GPUs

→

Commission

→

Serve customers

Why Power Has Become a Bottleneck

U.S. electricity demand is growing unusually quickly as data centers, manufacturing and electrification expand simultaneously.

Reuters reported in August that PJM, the largest U.S. power market, fell 6.8 GW short of its targeted capacity level in its 2028/29 auction even as data-center demand surged.

Another Reuters analysis described “time to power” as a central investment issue because generation projects and grid connections are not arriving as quickly as large new loads.

For Nebius, that means securing GPUs without power is not enough.

Why 5 GW Is Such a Large Number

One gigawatt equals:

1,000 megawatts

So 5 GW represents:

5,000 MW

That does not translate into a fixed number of GPUs because power density depends on the architecture, cooling design and supporting infrastructure.

But it shows the physical scale Nebius is targeting.

The NVIDIA Partnership Adds a Second Layer

In March 2026, NVIDIA agreed to invest $2 billion in Nebius and deepen engineering cooperation across AI factories, inference and system design.

The two companies said the partnership could support deployment of more than 5 GW of NVIDIA systems by the end of 2030.

Reuters described the scale as enough electricity to roughly match the needs of more than four million U.S. homes, illustrating just how large the planned infrastructure footprint is.

Why On-Site Generation Is Becoming Attractive

Waiting for a traditional grid connection can take years.

That creates an incentive to put generation closer to the data center.

Nebius has partnered with Bloom Energy on fuel-cell deployments, including a planned first project with 328 MW of installed capacity targeted for operation in 2026. Nebius said the behind-the-meter model can reduce dependence on new transmission infrastructure.

This is essentially a time-to-power strategy.

Why Pennsylvania Matters

In Q1, Nebius disclosed that it had secured up to 1.2 GW of power and land for a new owned AI factory in Pennsylvania.

That illustrates how the company is thinking about capacity:

not in 20 MW increments,

but increasingly in gigawatt-scale campuses.

Power Is Also a Competitive Advantage

Imagine two AI cloud companies with access to the same NVIDIA systems.

Company A can energize a new cluster in six months.

Company B has to wait three years for grid capacity.

Company A can sign and serve customers earlier.

So the competitive moat may be partly:

time-to-power

rather than merely:

time-to-GPU.

Sarah Chen: Power Is Becoming AI Cloud Inventory

Sarah Chen, MEXC senior crypto industry analyst, describes contracted power as a form of future inventory. A retailer secures products to sell later; an AI cloud secures megawatts that can eventually be converted into compute capacity. That analogy is imperfect, but it helps explain why power announcements can move NBIS even before a single new GPU begins billing customers. More of her work is available through her MEXC author profile.

Chen also cautions against valuing all 5 GW as though it were operating today. Between contracted power and revenue sit construction, generation, cooling, hardware procurement, commissioning and customer delivery. The bull case depends on Nebius shortening that conversion timeline. The bear case is that contracted power grows faster than the company's ability to deploy profitable computing infrastructure.

Why Power Efficiency Matters Too

Not all megawatts produce the same amount of AI output.

New hardware architectures can produce more compute per unit of power.

Networking and cooling efficiency also matter.

This means a future 1 GW site can potentially create more useful AI capacity than a 1 GW site built with older technology.

That is one reason Nebius's access to successive NVIDIA architectures matters.

What Investors Should Watch

The best power dashboard separates:

Contracted power

from

energized power

from

installed IT load

from

revenue-generating capacity.

Only the final stages generate customer revenue.

MEXC's broader highlights power and cooling as increasingly important constraints across the sector.

How Power Affects NBISON

NBISON does not represent a claim on a megawatt of electricity.

Power affects the underlying Nebius business.

If the market believes Nebius can secure scarce power earlier than competitors, NBIS may receive a higher valuation.

If sites suffer delays, the opposite can occur.

NBISON then reflects that NBIS exposure.

FAQ

Does Nebius currently operate 5 GW of AI data centers?

No. The Q2 figure refers to expected contracted power by year-end 2026.

What is the NVIDIA 5 GW target?

Nebius and NVIDIA said their partnership could enable deployment of more than 5 GW of NVIDIA systems by the end of 2030.

Are those two 5 GW numbers the same?

No.

Why is on-site generation useful?

It can reduce dependence on slow transmission and grid-interconnection timelines.

How large is the Bloom Energy project?

The first disclosed project targets 328 MW of installed capacity.

Risk Disclaimer

Contracted power is not the same as operating computing capacity. Projects face permitting, construction, generation, transmission, equipment, financing and commissioning risks.