Most conversations about the future of AI focus on hyperscalers, chips and increasingly capable models. For Michael Ferrari, chief AI officer at Moby, this overlooks a critical part of the equation: the physical infrastructure needed to sustain AI’s growth.
Speaking at Cypher 2026, Ferrari said the technology depends on an end-to-end value chain rooted in the physical world. Its constraints extend beyond energy to supply chains, human resources, water and the environments in which computing infrastructure operates.
He argued that the discussion around energy remains “very superficial”. Ferrari said capital must flow towards “smart investments” rather than speculative bets that follow market trends. Such misallocation could harm ordinary investors, including those whose retirement savings are exposed to these investments.
The physical cost of AI
Using International Energy Agency projections as a reference, Ferrari cautioned that the agency tends to underestimate the magnitude of changes, even when it identifies the right trends.
Using 2025 as a baseline, he said global data-centre electricity consumption could double over five years, while AI-related demand could triple over the same period.
Every token, query and model build carries a physical cost, Ferrari argued. He used copper to illustrate the material footprint behind computing, saying every token has “a litre and a kilogram of copper behind it”.
Investors and hyperscalers increasingly recognise energy constraints, he said. Water shortages and supply-chain bottlenecks have also attracted academic attention, but infrastructure development and capital deployment have yet to match the scale of these challenges.
When the grid runs out
Ferrari cited PJM, the regional transmission organisation serving parts of the eastern United States, where the wait between a permit application and a facility becoming operational can stretch to five to seven years.
Available electricity capacity is also shrinking. Ferrari said the surplus had fallen from about 70 gigawatts in 2022 to 18 gigawatts. If current trends continue, he warned, the system could face a deficit of around 40 gigawatts within five years, with demand outpacing new supply.
That has implications for the exchange-traded funds and investment portfolios financing AI infrastructure. “If they defy the laws of physics, it’s not going to matter,” he said.
Ferrari outlined two potential alternatives to terrestrial infrastructure, beginning with the ocean.
Moving compute into the ocean
Ferrari pointed to Microsoft’s Project Natick, which tested underwater data centres. He said the pilots worked, but Microsoft abandoned the project in 2024 because it could not achieve cost parity. Smaller companies are now pursuing similar approaches.
“In the most recent year, I think we’re seeing a bit more sustainable progress than over the previous four years,” he noted.
One advantage is cooling efficiency. Ferrari cited a power usage effectiveness (PUE) figure of 1.07. PUE measures total facility energy consumption against the energy used by IT equipment; a figure closer to one indicates greater efficiency. He described this level as “almost unheard of in other applications”.
Underwater facilities could also reduce dependence on freshwater, an increasingly important consideration for data centres. Ferrari recalled Chennai’s 2019 “day zero” water crisis, highlighting the difficulty of securing enough water for households, industry and computing infrastructure.
“Things can get really ugly really fast,” he said.
The ocean presents its own challenges, including corrosion, environmental risks and maintenance. Repairing a component on land may require sending a technician to a facility. Underwater, the task becomes considerably more difficult.
Ferrari described small-scale underwater computing as a stopgap and a potential bridge towards moving compute off the planet.
The case for orbit
Ferrari said developments including SpaceX’s planned IPO, Starship deployments and Google’s Project Suncatcher have brought orbital computing closer to investors’ attention.
Orbit offers access to solar energy while potentially avoiding the land, power and water constraints facing terrestrial infrastructure. It could also process data closer to where it is collected, reducing the need to transmit everything back to Earth.
However, space introduces another constraint: heat dissipation.
“Just because you have a cooling agent, which is space, you still have to dissipate that heat,” Ferrari said. Without an atmosphere to carry heat away, orbital systems must radiate it into space.
He identified thermal management and launch capacity as the two major hurdles. Ferrari estimated that the levelised cost of compute in space, measured in dollars per GPU-hour, is currently about four times that of terrestrial computing. Although launch costs per kilogram are falling, they remain prohibitively expensive.
If these barriers can be overcome within one to three years, he said, other parts of the ecosystem could begin falling into place.
A gradual shift in compute
Ferrari said it was difficult to bet against Elon Musk’s engineering ambitions, even if his timelines often prove optimistic. He credited that vision with helping advance electric vehicles and reusable rockets, and expects orbital computing to follow a similar trajectory.
However, he does not expect space-based compute to become commercially viable by 2027 or 2028. A market could begin emerging in the 2030s, with widespread viability potentially closer to 2040.
Describing the transition as a “glide path”, Ferrari noted that it is not an either-or scenario.
Terrestrial data centres will remain essential, particularly for latency-sensitive applications. Underwater facilities could take a modest share of computing demand, while orbital infrastructure develops over a longer horizon.
Ferrari expects land’s share of compute to decline gradually, with space potentially becoming a dominant source over time. He argued the transition will reward patient investors willing to finance infrastructure that takes years to mature.
“The limits are that, at the end of the day, there’s still power, water, materials, and silicon,” he said.
For Ferrari, the ocean is a bridge, while orbit represents the longer-term ambition. Whether that future materialises will depend not just on advances in AI, but on whether the cost of building and operating the infrastructure can make economic sense.
