Growing power constraints are prompting technology firms to explore orbital data centres, but consultancy firm Wood Mackenzie estimates they would cost more than three times as much as equivalent facilities on Earth.
The concept is being driven by a growing challenge facing the technology sector. The rapid expansion of AI is increasing power requirements at a time when access to grid capacity is becoming more constrained.
A recent study showed that the ChatGPT-4 model consumes 0.14 kWh each time it generates 100 words. This is roughly the same amount of energy used to boil one litre of water in a kettle.
According to research from Business Energy UK combined with findings from The Washington Post, the large language model (LLM) consumes an estimated 39.98 million kWh of electricity each day.
In practice, that is enough electricity to charge around eight million smartphones or power the Empire State Building for more than 500 days.
While ChatGPT represents just one AI application, it offers a glimpse of the growing energy demands associated with large-scale AI systems.
Viewed over a full year, the LLM’s electricity consumption exceeds the annual power consumption of 117 countries, highlighting the scale of energy required by today’s AI models.
Those requirements are being multiplied across thousands of AI workloads running in data centres worldwide, driving a sharp increase in electricity demand from the sector.
Global data centre power demand currently stands at 460 TWh in 2026, equivalent to around half of Japan’s total power generation. Wood Mackenzie forecasts that figure will reach 1,280 TWh by 2030 and 3,700 TWh by 2040, a 704% increase from current levels.
This need for more power is met with a series of challenges, especially in the US where data centres are on course to account for almost half of electricity demand growth between now and 2030.
Built to train xAI’s Grok chatbot, the Colossus data centre spans sites in Tennessee and Mississippi and houses more than 200,000 Nvidia GPUs ©Steve Jones Flight by Southwings for the Southern Environmental Law Center
A of energy sources will be tapped to meet data centres’ rising electricity needs, though renewables and natural gas are expected to be the leading sources.
A key challenge, however, will be securing that power quickly enough to keep pace with demand.
Grid connections in the country can take up to seven years and gas turbine equipment faces long wait times through 2030.
In water-stressed regions, limited supplies are creating competition for cooling systems, while higher labour and material costs are pushing up construction expenses.
“The constraints on terrestrial data centres are genuine, and they are not going away quickly,” said Robert Liew, Research Director at Wood Mackenzie.
There are also challenges associated with specialty gases used in the electrical power industry such as – a gas used in switchgear, most of which the US has historically imported from China and Russia.
In theory, a single gigawatt-capacity data centre in orbit would cost an estimated $170bn, more than three times the cost of an equivalent facility on Earth, with around 60% of the total attributable to launch and satellite costs.
The report notes that such projects will only become viable if the historical trend of declining launch costs continues.
“Putting a data centre in orbit still costs at least three times as much as building one on the ground,” said Liew. “That gap does not close without sustained and dramatic progress on launch costs.”
For now, that progress appears to be continuing. Global orbital launch attempts reached 324 in 2025, a 25% increase over 2024, with commercial operators conducting 70% of those missions.
The trend could accelerate further as launch costs continue to fall. Current-generation reusable rockets have reduced costs by around 90% compared with the 1960s, when it cost roughly $100,000 per kg to send the Apollo Saturn V into space.
A large share of recent launches has been driven by SpaceX, which deployed more than 3,000 satellites across 123 missions in 2025 for Starlink, its satellite internet ‘megaconstellation’.
Creating new demand
Air Products supplied more than 700,000 gallons of liquid hydrogen for NASA’s Artemis II mission ©NASA/Keegan Barber
While orbital data centres remain largely theoretical, the infrastructure needed to support a growing space economy is already being built. From rocket manufacturing and testing to launch operations and satellite deployment, that expansion is creating new sources of for industrial gases.
Earlier this year, Linde said aerospace activity linked to space vehicle production, testing and launch accounted for around half of the company’s 5% manufacturing growth in the US during the quarter.
CFO Matt White revealed that the company plans to isolate aerospace as a separate end market when it “consistently exceeds” 5% or more of global sales, a threshold that would equate to roughly $1.7bn per year.
Other industrial gas suppliers are also positioned to benefit from increased launch activity. Air Products recently supplied more than 700,000 gallons of liquid hydrogen for NASA’s Space Launch System rocket supporting the Artemis II programme.
Space activity and AI-driven semiconductor demand are also creating for specialty gases such as helium and xenon, according to Messer CEO Bernd Eulitz.
“For every $10bn to $15bn a global chip company invests into a chip fab, a gas company will invest roughly $1bn,” he said at gasworld’s Specialty Gas Summit.
Despite growing interest in orbital facilities, investment remains firmly grounded. Wood Mackenzie forecasts around 395 GW of new terrestrial data centre capacity will be added between 2026 and 2040 under its base-case scenario.
Meeting that demand will require around $9 trillion of investment, underlining the scale of the opportunity available to conventional data centre developers even as space-based alternatives continue to attract attention.
“We forecast $9 trillion of terrestrial data centre investment between now and 2040,” said Liew. “Orbital data centres are a serious long-term proposition, but right now they remain a bet on the cost curve.”
