AI power demand: how many gigawatts do AI stock prices need?

The numbers

Fact from filings or reports, sourced below · Assumption a Stonkometer default you can change · Our view opinion

415 TWhused by all data centres in 2024, about 1.5% of the world's electricityFact

$18B → $40BNVIDIA revenue per gigawatt of data centre: Hopper, then Vera RubinFact

474 GWof large-load requests in Texas (ERCOT), about 90% data centres, against a 91 GW peakFact

61 monthsmedian wait from grid request to running for US power plants built in 2025Fact

116 GWof GE Vernova gas turbine orders and reservations, mostly sold out to 2030Fact

97 GWof AI data-centre power that today's AI share prices imply by 2030, in our modelAssumption

Forecast	Starting point	Where it goes
IEA, world (Apr 2025)	415 TWh in 2024, about 1.5% of world electricity	about 945 TWh in 2030; 700-1,700 TWh in 2035 across its four scenarios
LBNL, United States (Dec 2024)	176 TWh in 2023, 4.4% of US electricity	325-580 TWh in 2028, 6.7-12% of US electricity
EPRI, United States (May 2024)	about 4% of US electricity in 2023	4.6-9.1% of US electricity in 2030
US utilities, via Grid Strategies (Nov 2025)		about 90 GW of new data-centre peak load by 2030; Grid Strategies thinks about 65 GW is more realistic

These cover all data centres, including email, video and cloud storage; AI is the part that grows fastest. Forecasts made only a year apart differ by two times or more, which tells you how uncertain this is.

From share prices to gigawatts

Three ways to estimate how much power AI will need, against the new power you think can be built. The revenue comes from our valuation models; everything else is an assumption you can change. Assumption

Chips the prices imply. The data-centre revenue that NVIDIA's and Broadcom's share prices need each year, turned into chips, then into gigawatts. Chips stay in service for the years you choose.
Computing the AI labs' revenue implies. OpenAI's and Anthropic's revenue in our models, the share they spend on computing, scaled up to all AI users.
Usage minus efficiency. How fast AI use grows, minus how fast each unit of AI work gets more energy-efficient.

Revenue:

Chips the prices implyComputing the AI labs' revenue impliesUsage minus efficiencyPower availableBought but not powered

To run the chips that today's prices imply, the world needs about 97 GW of AI data-centre power by 2030 and 174 GW by 2036, up from 30 GW today. With the supply you assume, up to 17 GW of chips (about $628B) would be bought but not powered, at the worst point in 2031.

New power needed a year to 203017 GWyou assume 14 GW a year

Electricity in 20361140 TWh2.4% of the world's electricity

In the US in 203049 GWutilities forecast about 90 GW of new data-centre load by 2030

Power bill in 2036$91Bvs $1.66T of chips bought that year

NVIDIA sales at risk, 2027-2036$402B5.3% of the data-centre sales these prices need

Change the assumptions

Chips NVIDIA + Broadcom share of all AI chip spending85% Chip spending per GW of data centre, today$30B Change in chip spending per GW, a year4% Years a chip stays in service5 yrs AI data-centre power running today30 GW

AI labs and usage Share of OpenAI's and Anthropic's revenue spent on computing50% Their share of all AI computing35% Cost to rent 1 GW of AI computing for a year$11B Growth in AI use a year, next 5 years70% Growth in AI use a year, the 5 years after40% Fall in energy per unit of AI work, a year25%

New power for AI a year by 2030 Grid connections7 GW Gas, on site or dedicated5 GW Solar, wind and batteries (firm equivalent)3 GW Nuclear from 2031 (restarts, uprates, small reactors)2 GW Flexibility (cutting use at grid peaks)2 GW

Running costs Average use of the power capacity75% Electricity price$80/MWh Share of AI power in the US50%

How it works: chips bought each year = data-centre revenue / NVIDIA and Broadcom's share; GW added = chips / spending per GW; power needed = chips still in service. Power available = today's plus the new supply you set, ramping from 60% of your 2030 pace in 2027 to 100% by 2030 (nuclear 0.5 GW a year until 2030). Chips above the power line are "bought but not powered". NVIDIA sales at risk = the yearly rise in unpowered chips, valued at that year's spending per GW, times NVIDIA's share: orders that buyers would delay if they cannot plug them in. Our view

Default anchors: NVIDIA said in August 2026 that it earns about $25B per GW with Grace Blackwell and $40B with Vera Rubin, and about $18B with Hopper. All data centres averaged about 47 GW in 2024 (IEA), and we assume about 30 GW of it is AI today. OpenAI said it ended 2025 with 1.9 GW running. Fact

What has to be true to get there

With our defaults, the market's view needs about 17 GW of new AI data-centre power a year until 2030. For scale: Amazon added about 3.9 GW of power in all of 2025, Microsoft adds about 1 GW a quarter, and only 525 MW of new or upgraded capacity cleared PJM's latest auction, the largest US grid. Fact

Several sources at once. No single source can add 15-20 GW a year for AI by 2030: gas turbines are sold out, nuclear is slow, solar and wind need storage, and the grid queue takes about five years. Our view
Faster grid connections. In June 2026 the US energy regulator (FERC) ordered grid operators to revise or justify their rules for connecting data centres and other large loads. Texas now charges a fee and requires disclosure of duplicate requests. Fact
Flexible data centres. Accepting short cuts at peak times is the quickest way to unlock existing grid capacity. Fact
Efficiency that does not cut demand. If each unit of AI work gets much cheaper and people do not use more of it, chip sales fall with power. The prices need AI use to keep growing faster than efficiency. Our view
Somebody pays. Power is a small share of an AI data centre's cost, so tech companies can outbid others; that shifts costs to households and other businesses, which brings political pushback. Our view

Where could the power come from?

Source	What it offers	The catch	Cost
Gas turbines and plants	The fastest big source today, on site or on the grid. Gas was 46% of the capacity that won PJM's July 2026 auction.	GE Vernova had 116 GW of gas turbine orders and reservations in July 2026 and is mostly sold out to 2030; prices rose 10-20% in half a year.	$48-109 per MWh (combined cycle)
Nuclear restarts and uprates	Closed plants reopening and existing ones pushed harder, sold to tech companies on 20-year contracts.	Only a handful of plants can restart. Crane (Three Mile Island unit 1, 835 MW, for Microsoft) aims for 2027.	$141-220 per MWh (new build)
Small modular reactors (SMRs)	Factory-built reactors next to data centres, ordered by Google, Amazon and others.	First commercial units are not expected before about 2030; costs are still unknown.	Unknown until the first units run
Solar, wind and batteries	Cheapest new energy and quickest to permit in many places.	Weather-dependent: a data centre running day and night needs batteries, backup or a grid behind it. Federal tax credits were cut back in 2025.	$38-78 per MWh (solar); $50-131 with storage
The grid and transmission	Connecting to existing power and building new lines.	Grid queues: the median power plant built in 2025 waited 61 months from request to running. ERCOT in Texas has 474 GW of large-load requests, about 90% data centres, five times its peak demand.	Shared by all customers
Flexibility	Data centres that cut use for a few hours at grid peaks can connect years earlier.	Duke University found US grids could add 76-126 GW of new load if it accepts cuts in 0.25-1% of hours. Training jobs can pause; chat and search cannot easily.	Cheap, but needs new contracts
Moving the computing	Building where power is spare: the Gulf, Nordics, Canada, and (one day) space.	Export rules, delays sending data far away, and politics. Orbital data centres are a moonshot on the SpaceX page.	Varies

Fact Costs are Lazard's unsubsidised ranges per MWh for new US plants (June 2025). Anthropic estimated in July 2025 that the US AI industry needs at least 50 GW of new power by 2028.

Using less: efficiency and adoption

Each new chip generation does far more AI work per watt, and software tricks (smaller models, better algorithms) cut energy further. But every time AI gets cheaper, people and companies use more of it, the Jevons paradox. Whether total power use rises or falls depends on which grows faster: use or efficiency. Try it with the "Efficiency surprise" and "Demand slows" buttons above. Our view

Adoption matters as much as efficiency. Chat answers use little energy each; AI agents that work for hours, video generation and reasoning models use many times more per task. A shift from chat to agents can multiply demand even if the number of users stops growing. Our view

Stranded GPUs: chips with nowhere to plug in

When chips arrive faster than power, they wait: in warehouses, in half-finished buildings or in racks that are switched off. This is already happening.

Microsoft (Nov 2025). CEO Satya Nadella said the biggest issue is not a glut of computing but power, and that Microsoft had chips "sitting in inventory" it could not plug in. Fact
Amazon (Oct 2026). Business Insider reported that AWS takes ownership of NVIDIA GB300 racks before its sites can power them, with about $5B of racks in US warehouses in August and about $3B more a month expected. Amazon called staging hardware near sites best practice. Fact
Santa Clara (2025). Two finished data centres of nearly 100 MW in NVIDIA's home town were reported idle because the city's grid could not yet supply them. Fact
Capacity sold out. Amazon says AWS power stays short through 2027 and reserves blocks of hundreds of megawatts to 1 GW well in advance; Microsoft says it remains capacity constrained. Fact

What it means for NVIDIA

Sales are booked when chips ship

NVIDIA records revenue when it delivers, so racks waiting in a warehouse still count. The risk shows up later, as smaller or delayed orders once buyers hold too much unplugged hardware. Our view

Power, not money, caps the market

If only so many gigawatts can be connected each year, NVIDIA's growth is capped by gigawatts times its revenue per gigawatt. That is why it pitches revenue per GW: $18B for Hopper, $25B for Grace Blackwell, $40B for Vera Rubin. Fact

Best chips win when power is scarce

With a fixed power budget, a data centre earns the most with the chips that do the most work per watt. That supports prices for NVIDIA's newest chips, but makes older ones worth less sooner. Our view

NVIDIA is now in the power business

NVIDIA guarantees up to $108.5B of partners' data-centre obligations, $105B of it for SB Energy's sites, and signs leases it plans to pass to others. If power projects slip, those guarantees tie NVIDIA to its customers' problems. Fact

Signals to watch: NVIDIA's receivables and inventory, customers' "construction in progress" (buildings not yet running), cloud capacity comments on earnings calls, gas turbine and transformer lead times, PJM and ERCOT auction results, and rental prices for AI chips.

Who pays for the power?

For a tech company, power is cheap next to chips: with our defaults a year of electricity for 1 GW costs about $1B, against about $30B of chips. So AI buyers can outbid everyone else for scarce power. Assumption

PJM's auction for mid-2028 to mid-2029 cleared at its price cap of $325 per MW-day and still came up 6.8 GW short of its reliability target; without the cap it would have cleared at about $555. The market monitor attributes $6.3B of the $16.4B cost, 38%, to data centres. Those costs reach household bills in 13 states. Fact

Lexicon: power words

Watt, megawatt (MW), gigawatt (GW): Power: how much electricity something draws at one moment. A GW is a billion watts, about one large power plant. AI campuses are now planned at 1-5 GW each.
Watt-hour, terawatt-hour (TWh): Energy: power used over time. 1 GW running all year is 8.76 TWh. The US uses about 4,000-4,500 TWh a year.
IT load vs facility power: IT load is what the chips and servers draw; facility power adds cooling and losses. The ratio is called PUE, about 1.1-1.4 for new AI sites.
Interconnection queue: The waiting list to connect a power plant or a big customer to the grid. Most requests in it are never built.
Capacity auction: Where grid operators such as PJM pay power plants to be available at peak times. High prices signal a shortage, and the cost lands on everyone's bills.
Behind the meter: Power made on site (often gas turbines) and used directly, without waiting for a grid connection.
Stranded or dark GPUs: AI chips that have been bought and delivered but are not running, usually because the building has no power or cooling yet.
Perf per watt, tokens per watt: How much AI work a chip does for each unit of power. When power is the bottleneck, this decides how much revenue a data centre can earn.
Jevons paradox: When something gets more efficient and cheaper, people often use so much more of it that total consumption rises.

Questions people ask

How much electricity does AI use?

The IEA estimates all data centres used about 415 TWh in 2024, about 1.5% of the world's electricity, and that AI servers are the fastest-growing part. Our model assumes AI data centres run about 30 GW today, roughly 200 TWh a year at the usage we assume.

How much power will AI need by 2030?

Forecasts range widely. The IEA's base case has all data centres at about 945 TWh in 2030; US utilities forecast about 90 GW of new data-centre load by 2030. Our model turns today's AI share prices into about 97 GW of AI data-centre power worldwide by 2030.

What are stranded GPUs?

AI chips that are bought but not running, because the data centre has no power or cooling yet. Microsoft's CEO said in November 2025 that chips were "sitting in inventory" that he could not plug in, and in October 2026 Business Insider reported Amazon was storing billions of dollars of NVIDIA racks in warehouses waiting for power.

Is power a risk for NVIDIA?

Power limits how many chips customers can install each year, whatever their budgets. If chips pile up unplugged, orders get delayed. It also pushes buyers towards the chips with the most work per watt, which so far favours NVIDIA's newest products. See the stranded chips section above.

Value the companies behind the numbers: NVIDIA · Broadcom · OpenAI · Anthropic · CoreWeave · Oracle · Microsoft · Amazon · Alphabet · Is AI a bubble?

Sources

Not investment advice. Stonkometer is an educational calculator, not investment advice. Valuations depend entirely on the assumptions you choose. We are not a broker or adviser, and nothing here is a recommendation to buy or sell any security. Do your own research and consider a licensed adviser. Full disclaimer