Musk Warns of a 15-Gigawatt AI Power Shortfall by 2027

Elon Musk’s main warning to G20 ministers on September 1 was that electricity—not just advanced chips—could become the next hard limit on artificial intelligence. He cited an industry consensus that AI chips could face at least a 15-gigawatt power shortfall in 2027, while chip production grows roughly 40–50% a year and power generation outside China grows only about 10–20%. 1611

Musk paired that warning with an expansive economic forecast: he said digital AI could increase global economic output by roughly 20% to 30%, or about $20 trillion to $30 trillion annually. That is Musk’s estimate, rather than an independent forecast from the International Energy Agency. 18

Why electricity is becoming AI’s bottleneck

The problem is not simply the amount of electricity generated worldwide. AI data centers need enormous, continuous and reliable supplies of power in specific locations. They also compete for substations, transformers, transmission capacity, permits and grid connections.

AI facilities can be planned and constructed faster than the infrastructure needed to energize them. In many regions, connecting a new facility to the grid can take four to 10 years, while AI data centers are typically planned and built within two to three years. 30 That timing gap means servers may be ready before the electricity supply is.

Location adds another constraint. Data centers tend to cluster where they can access affordable electricity, land, water or cooling resources, communications networks and favorable policies. Concentrated demand can therefore create acute local grid pressure even when the global electricity market appears to have sufficient capacity. 1821

The demand is large, but the local impact can be larger

The IEA’s base case projects that global data-center electricity consumption will more than double to around 945 terawatt-hours (TWh) by 2030, representing just under 3% of global electricity use. From 2024 to 2030, data-center consumption is expected to grow by about 15% annually, with AI the most important driver. 2223

The global percentage can obscure the challenge faced by individual utilities and regions. A small number of hyperscale campuses can require power on a utility scale, creating long interconnection queues and delaying other projects. The IEA estimates that around 20% of planned data-center projects could face delays if grid risks are not addressed. 23

Companies are trying to secure power directly

The response is increasingly moving beyond buying processors and leasing data-center capacity. Technology companies and their partners are trying to control more of the infrastructure that turns chips into usable compute.

SpaceX: making turbine components in-house

Musk has said SpaceX plans to manufacture key gas-turbine components, including blades and vanes. The move is intended to address equipment bottlenecks and accelerate the deployment of dispatchable power for AI infrastructure. 3839

That strategy reflects the broader shift from a chip-centered supply chain to a system that also depends on turbines, fuel, substations, transmission and site-level generation. Owning or producing parts of that chain may reduce some delays, but it does not eliminate permitting, environmental or fuel-supply constraints.

Microsoft and Chevron: dedicated generation

Microsoft and Chevron have agreed to develop a co-located natural-gas-fired facility to supply a Microsoft data-center campus in West Texas under a 20-year arrangement. Reuters reported that Chevron was also exploring similar data-center power deals elsewhere in the United States. 3335

Dedicated generation can avoid some grid-connection delays and provide firm power close to the computing load. It also brings trade-offs: natural-gas generation creates emissions and ties data-center expansion to fuel infrastructure and regulatory decisions.

NVIDIA and the wider infrastructure race

NVIDIA’s role illustrates why the issue extends beyond chip manufacturing. GPUs have commercial value only when data centers have the power and grid access needed to operate them. As a result, the AI buildout is increasingly linked to investment in electricity generation, data-center construction and interconnection capacity—not just accelerator supply.

Musk’s policy prescription: build toward energy abundance

Musk urged governments to treat power infrastructure as strategic industrial capacity. His argument was that countries should build generation, transmission and data centers quickly enough to capture the productivity gains he expects from AI.

The policy message fit the broader U.S. position at the ministerial. Reuters reported that U.S. officials pressed G20 members to take a light-touch approach to AI regulation and avoid creating new international oversight bodies. 50 CNBC also reported Musk’s call for governments to make new activity “default legal, not default illegal.” 55

That approach frames regulation and infrastructure as part of the same competitiveness debate: rules that slow construction may reduce near-term deployment, while rules that ignore practical harms may leave workers, communities and energy systems carrying more of the cost.

Why the G20 discussion mattered

The Chapel Hill meeting brought AI infrastructure into a wider contest over economic growth, technology standards and strategic advantage. Musk appeared virtually on the first day, while NVIDIA CEO Jensen Huang and OpenAI CEO Sam Altman were among the prominent participants in the two-day ministerial. 495261

Electricity was therefore not discussed as an ordinary utility issue. It was presented as a potential determinant of which countries can build AI capacity, attract data centers and capture the economic benefits of the technology. Musk’s 15-gigawatt figure remains a projection, but the underlying concern is supported by independent energy analysis: data-center demand is rising rapidly, grids are already constrained in some regions, and the infrastructure needed to connect new facilities often takes longer than the facilities themselves to build. 222330

The practical question for governments is no longer only how many chips the AI industry can produce. It is whether power systems can deliver reliable electricity to those chips quickly enough—and who should pay for the generation, transmission and environmental costs required to do it.