The gigawatt is the floor, not the size of the panel field. On Moonshots with Peter Diamandis, Ramez Naam described a solar-and-battery project in the United Arab Emirates designed to deliver at least one gigawatt of electricity around the clock. To support that output, he said, it needs roughly five gigawatts of solar panels and 19 gigawatt-hours of batteries.
That combination anchors his argument for a different geography of AI: instead of moving electricity long distances to places that already need it, build new data centers beside abundant solar resources. Unlike an existing city, a computing facility that has not yet been built can choose its location.
But Naam's enthusiasm comes with a seasonal limit. A battery that earns its keep every night is a different proposition from one that saves summer energy for winter.
What cheap panels do—and do not—buy
Naam began with the extraordinary decline in panel prices: from roughly $100 per watt in 1975 to about eight cents for Chinese panels in the prices he cited. That is a component price, not the cost of dependable electricity delivered to a data center.
The distinction matters because the computing equipment is expensive. Naam argues that running those chips only when the sun shines would waste too much of the investment. Solar needs storage, or another source of power, to keep them working after sunset.
In the UAE example, the larger panel field would generate electricity for immediate use while charging batteries for later. Gigawatts measure how quickly power can be supplied; gigawatt-hours measure how much energy the batteries can hold. Naam put construction cost at roughly $6 per watt of the intended continuous output. That is an upfront investment figure, not a price for each unit of electricity ultimately delivered, and the project's intended output should not be confused with an operating record.
He sees speed as another advantage. With gas turbines facing years-long waits, he said a solar-and-battery project can be built in 12 months. His later discussion of land and regulatory obstacles makes clear why that is a construction opportunity, not a timetable every site can meet.
Battery costs are central to his forecast. He cited a fourteenfold decline since 2010 and argued that sodium-ion technology, using an element much more abundant than lithium, could eventually help drive another tenfold reduction. That is an expectation about future costs—not an announced tenfold commercial breakthrough.
Another thousandfold decline takes more than time
A panelist asked what it would take for solar to repeat its thousandfold price decline. Naam's answer turned on a distinction easily lost in charts that plot prices against years: the learning curve follows cumulative deployment, not the calendar.
Under this relationship, known as Wright's law, each doubling of the total amount deployed brings another percentage reduction in cost. The industry must keep getting larger to repeat the gain.
In his 2020 solar-cost analysis, Naam fitted seven price and cost series covering global averages, the United States, India and China, principally over 2009–2020. He estimated declines of 30–40% per doubling of cumulative global deployment and used 30% for his projections. Learning in equipment production was global, he argued, rather than confined to the country installing the panels.
There were limits within those results. Cheaper financing and thinner developer margins helped lower auction prices, but neither could keep falling indefinitely. The projections also concerned new solar generation; storage and wider electricity-system costs were separate.
On the podcast, Naam sketched a world in which solar supplies a much larger share of electricity and total demand doubles. He suggested that further deployment could bring a fourfold or perhaps eightfold cost reduction—not another thousandfold within that scenario.
The problem is not just night. It is January.
London makes the seasonal problem concrete. Naam said it receives roughly a sixth or a seventh as much sunlight in January as in June or July. Covering that gap means more than installing enough batteries to last until sunrise: it requires much more generation, long-duration storage, other sources of electricity, or some combination.
His battery arithmetic explains why. A battery charged and discharged daily can spread its upfront cost over about 365 cycles a year, or thousands over a decade. Storage used only once or twice a year to bridge seasons has far fewer opportunities to recover its cost. This is the capital-cost part of the calculation; electricity used to charge it, energy losses and operating costs also matter.
Naam sees the economics of day-to-night storage moving in the right direction. Seasonal storage needs a different approach, with developers exploring options including compressed air and converting electricity into storable fuels.
Winter also brings extra demand. Heat pumps use electricity to heat buildings that may previously have burned gas. Naam argued that a broad shift to electric heating could roughly double winter electricity demand in the UK, just when solar output is weakest. That is why he sees a role for seasonal storage and other power sources, including nuclear and advanced geothermal, in places with long winters or rainy seasons.
The Royal Society's 2023 storage report explored this problem for Great Britain—not for a universal solar-only system. It modelled electricity supply and demand hour by hour using weather from 1980–2016, with a principal assumption of 570 terawatt-hours of annual demand in 2050.
A generation mix of roughly 80% wind and 20% solar reduced average seasonal imbalance, but substantial variation between years remained. The report found that wind and solar backed by large storage could meet demand, with underground hydrogen particularly suited to infrequently used, long-duration reserves. Batteries served different needs, including rapid response.
Nor did adding inflexible generation automatically make the modelled system cheaper: its price had to beat the system's existing average. The results depended on future costs, demand and weather assumptions, and the authors recommended contingency beyond the historical weather sample.
Move the load, not the electrons
Naam's siting argument starts with the demand still to come. Existing populations cannot readily relocate to follow cheap power; new AI facilities have more freedom.
Australia stood out in his shortlist because of its space, strong solar resources and a government he regarded as friendly to the investment. He also pointed to Chile and Mexico. The Gulf has excellent sunlight, but it does not have the same concentrated advantage in solar that some countries have in oil.
In the American Southwest, Naam said, assembling land parcels can be a bigger obstacle than the technology. He proposed opening suitable federal land in sunny states such as Nevada and Arizona, while avoiding areas of high natural value.
His account of a visit to Chihuahua, Mexico, brought the legal constraints into view. He said restrictions on private generation for on-site use frustrated the case he was making for solar-powered AI facilities. His proposed remedy was to enable that generation and provide strong protection for data and intellectual property.
AI developers worry not only about user data being leaked or seized, he said, but also about losing their model weights—the numerical settings learned during training that enable a model to work. Cheap electricity is not enough if a company is uncomfortable placing those assets under a country's laws.
For Australia, Chile and the Gulf, the opportunity he describes is therefore to export computing services rather than electricity. Saudi Arabia need not build power lines to America to sell something made with its sunshine. Naam's advice to energy-rich countries used to be to find industrial uses for their electricity. Now it is to “make data, make intelligence”—with the land, permissions and legal protections that would persuade AI companies to build there.