Huang shouts «AGI is here» – but forgets the power crisis behind the scenes

Huang's big claim

Nvidia CEO Jensen Huang has on several occasions in recent months claimed that general artificial intelligence – AGI, meaning machines that match or surpass human cognition across tasks – is in practice already here. According to Yahoo Finance, this is a claim that has attracted a great deal of attention, but which at the same time overshadows a more prosaic problem investors seem to be ignoring: where is all the electricity supposed to come from? (finance.yahoo.com)

While the debate over AGI definitions rages on, the numbers point to a concrete, measurable bottleneck problem. AI models require exponentially more computing power, and computing power requires electricity in quantities the global power grid was not built to deliver quickly enough.

The power wall nobody is talking about loudly

The International Energy Agency (IEA) estimates that the world's data centers used around 415 terawatt-hours (TWh) of electricity in 2024, equivalent to about 1.5 percent of global power consumption. In the IEA's base-case scenario, this doubles to nearly 945 TWh by 2030, with AI-driven computing as the fastest-growing component.

In the US, the picture is even sharper. Goldman Sachs Research estimated in June 2024 that power demand from data centers will increase by 160 percent by 2030, and that data centers will then account for 8 percent of total US electricity consumption – up from around 3 percent in 2022. Lawrence Berkeley National Laboratory found that US data centers drew 4.4 percent of the country's electricity in 2023 (176 TWh), with a forecast of between 6.7 and 12 percent by 2028.

Waiting lists, equipment shortages, and a price-shock auction

The problem is not only how much power is needed, but how long it takes to build the capacity. According to LBNL's «Queued Up» analysis, more than 2,200 gigawatts of proposed new power generation and storage were queued for grid connection in the US at the end of 2024. The median wait time from application to commercial operation has increased from under two years in the 2000s to around five years today. In the PJM grid, which covers 13 states in the Midwest and on the East Coast, projects in high-growth zones for data centers now wait 36 to 48 months.

The equipment side is also failing: delivery times for large high-voltage transformers have increased from a normal 50–80 weeks to as much as 150–210 weeks, or nearly four years.

This had a direct impact on electricity prices. PJM's capacity auction for 2025/2026 cleared in July 2024 at $269.92 per MW-day, an increase of more than 800 percent from $28.92 the year before. Market monitor Monitoring Analytics estimated that accelerated data center forecasts accounted for 63 percent of the price jump – around $9.3 billion in increased capacity costs passed on to consumers in the region.

Accelerated computing is sustainable computing – but only if the power actually exists where you build

Big Tech responds with nuclear power – and meets resistance

The tech giants have tried to bypass the slow grid queues by connecting data centers directly to power plants, known as «behind-the-meter» solutions. This has triggered regulatory resistance.

In March 2024, Amazon Web Services bought a data center campus next to the Susquehanna nuclear power plant from Talen Energy for $650 million, with plans to increase direct nuclear power delivery from 300 to 480 megawatts. The US Federal Energy Regulatory Commission (FERC) voted in November 2024 against the deal 2 to 1, after competitors such as Exelon and American Electric Power warned that taking 480 MW of baseload out of the ordinary grid could weaken grid stability and shift costs onto ordinary electricity customers.

Others have taken a different route. Microsoft entered into a 20-year power agreement with Constellation Energy in September 2024 to reopen Reactor 1 at Three Mile Island, now renamed the Crane Clean Energy Center, with planned startup in 2028. Google signed an agreement that same autumn for 500 MW from small modular reactors from Kairos Power, aiming for delivery from 2030.

Crypto miners become power brokers

An unexpected result of the power shortage is that Bitcoin miners have become attractive infrastructure players. After the halving in April 2024, which cut the block reward from 6.25 to 3.125 BTC, several publicly listed miners began offering their already energy-supplied facilities to AI companies.

Core Scientific entered into a twelve-year hosting agreement with CoreWeave in June 2024 to deliver 200 MW of high-density HPC infrastructure, worth an estimated $3.5 billion in contracted revenue. Companies such as TeraWulf, Iris Energy and Hut 8 have converted parts of their power agreements from pure Bitcoin mining to GPU operations, as AI players have been willing to pay $100–140 per MWh for electricity that is already ready to use – far above what pure mining operations can justify.

Critical voices from within the industry itself

It is not only outside skeptics pointing to the problem. Meta CEO Mark Zuckerberg said in April 2024 on the Dwarkesh podcast that before hitting capital constraints, you will hit energy constraints, and that new power plants and transmission lines require heavy regulatory processes and many years of lead time. OpenAI CEO Sam Altman stated at the World Economic Forum in January 2024 that there is no way forward without a breakthrough – either fusion energy or radically cheaper solar power with storage at massive scale.

Even Jensen Huang has acknowledged the reality: he has stated that future data centers will likely need to be built where surplus energy exists, not where the population lives.

A Norwegian angle: power surplus as a competitive advantage

The same pattern of long grid queues and power shortages is also found in Europe, including Norway. Statnett has for several years reported capacity constraints and waiting lists for new large electricity customers, particularly in central regional areas. Norway's combination of power surplus from hydropower and cold climate makes the country a potentially attractive host nation for energy-intensive data centers – something that has already sparked debate about prioritization between industry, exports, and new AI infrastructure.

What this means for investors

Huang's AGI claim is difficult to verify objectively, since there is no universally accepted definition or test for when AGI has actually been achieved. There is therefore reason to treat the statement as marketing and strategic positioning from a player who benefits from continued AI investment growth, rather than a neutral technical assessment.

What is documented, however, with concrete figures from the IEA, Goldman Sachs, LBNL and PJM's own market monitor, is that power supply – not chip capacity alone – could become the decisive constraint on how fast the AI industry can actually scale in the coming years. For investors in Nvidia, hyperscalers, and power companies, it is this bottleneck, not the debate over AGI definitions, that will likely determine returns going forward.