ChatGPT Is Terraforming Earth

Your ChatGPT subscription is subsidising the terraforming of Earth.

Not metaphorically. China’s solar overbuild is crashing electricity prices toward zero. AI demand absorbs the surplus. When power is essentially free, we manufacture fuel from thin air instead of drilling for it.

Solar has become a deflationary technology, following the same exponential cost curves that made computing ubiquitous. In 2024, China installed more solar capacity in a single year than the United States has installed in its entire history.¹ The world added roughly 600 GW last year, equivalent to installing Australia’s entire grid every few months.

None of this was planned. More than $130 billion poured into China’s solar industry, driven by what industry leaders have called “blind investments and disorderly financial lending.”² China now holds over 80% of global manufacturing capacity across the entire supply chain, with utilisation rates hovering around 50-60%. The industry is producing far more than the market can absorb. Solar hardware costs have collapsed to near $0.10 per watt and continue falling. We have left the era of energy scarcity and entered the era of manufactured energy.

Into this glut arrives the hungriest electricity buyer in history.

Grid connections take seven years while AI scaling laws operate on months. Tech giants are building dedicated solar installations behind the meter, bypassing the glacial pace of grid development. AI data centre demand is projected to double by 2030, potentially reaching a Japan-sized energy footprint. But unlike households or hospitals, AI loads are flexible. Workloads can shift across hours or days. Compute can relocate to wherever energy is cheapest. Training runs are price-sensitive and will chase arbitrage. This makes AI the perfect absorber of surplus daytime solar. Solar is becoming the world’s cheapest input precisely as AI becomes the world’s hungriest buyer.

To run a data centre 24/7 on solar power, you must overbuild capacity by four to five times.³ For roughly eight hours each day during peak sun, these facilities generate gigawatts of excess electricity that the batteries cannot hold and the servers do not need.

Turning air and water into natural gas is thermodynamically possible but economically absurd at normal electricity prices. Capture CO2 from the atmosphere, electrolyse water into hydrogen, combine them to produce methane. The chemistry has been understood for over a century, but it requires so much electricity that it only makes sense when power costs around $10 per megawatt hour.

That “waste” energy from AI solar farms is exactly that free electricity.

Terraform Industries has built container-sized units that do precisely this.⁴ Instead of curtailing that midday spike, we feed it into machines that convert sunlight and air into pipeline-grade natural gas. Input: sunlight plus air. Output: methane. Carbon becomes a circular industrial input rather than a pollutant extracted from underground. The synthesis plants soak up whatever the batteries cannot store.

Once electricity becomes essentially free during daylight hours, we stop drilling for fossil fuels and start manufacturing them from air. Desalination becomes trivial. Synthetic ammonia production gets cheap, shifting the cost structure of global food. Plastics become carbon-neutral.

Energy independence stops being about who controls oil reserves and starts being about who has land and sunlight. The Middle East of the future is Nevada, Australia, and Namibia.

This reframes the entire conversation about AI power usage. The real question is whether we can generate electricity cleanly. The answer is becoming yes, in such abundance that we will need to invent new uses for the surplus.

For most leaders, this feels distant. Most of us are not building hyperscale data centres. But the second-order effects matter.

Watch energy costs in cloud pricing. As electricity becomes the dominant cost in AI inference, providers will increasingly price by consumption. The ones who figure out solar arbitrage will offer cheaper inference. This will show up in your bills before you understand why.

Watch local AI. When running models locally becomes cheap enough, the economics of cloud versus on-premise shift. The line is moving. Keep an eye on where it lands.

Prepare for abundance thinking. The interesting question is shifting from “how do we afford AI?” to “what becomes possible when AI is essentially free?” LLM costs per token have dropped 100x in two years.⁵ Planning for scarcity when abundance is arriving is the wrong bet. The companies that win will be the ones who figure out what to build when the constraint disappears.

China’s industrial policy is creating the oversupply. AI is arriving to absorb it. Synthesis technology makes the excess useful. The path to addressing climate runs through extreme abundance, and your AI usage is part of what makes it viable.