TLDR: AI is eating electricity faster than we can build the grid. The bottleneck today isn’t cost; it’s poles, wires, and permitting. Solar plus batteries is the fastest path to power today, but gas turbines are sold out 7 years ahead and nuclear’s comeback rides on SMRs and fusion. The race to a terawatt is the defining infrastructure story of the AI age. Here’s the full picture from my conversation with Ramez Naam.

Last week on Moonshots, I sat down with Ramez Naam, a dear friend of nearly 20 years and my go-to authority on all things energy. Ramez is a founding faculty member of Singularity University, partner at Planetary VC, author of the Nexus trilogy, and one of the top five forecasters of solar costs in the world. He gave us an epic masterclass on what he calls ‘the innermost loop’: energy. Here is what you need to know that Ramez taught us…

1. AI IS POWER HUNGRY (BUT POWER ISN’T THE COST)

A 1-gigawatt data center costs roughly $50 billion to build. $35 billion of that is chips. The 5-year energy cost? It’s trivial next to the capex. This flips the usual intuition on its head.

If you tell OpenAI or Anthropic, ‘We can give you power at twice the cost tomorrow,’ they’ll take it. Energy is the BOTTLENECK, not the cost line. The AI labs are racing to build the biggest models, and every month of delay costs them more than a decade of cheap power would save.

Energy is the BOTTLENECK, not the cost line.

Chart 1: The gap between projected GPU power demand and grid buildout. Source: Ramez Naam, Moonshots podcast.

Data center electricity could reach 20% of US consumption by 2035, according to EPRI. Globally, we’re looking at 1,050 TWh by 2026, more than the entire country of Japan uses today. And that’s the conservative scenario.

Chart 2: US data center electricity as a share of total consumption. Source: EPRI, IEA.

2. THE GRID IS THE BOTTLENECK TODAY

Interconnection queues went from 15 months twenty years ago to 45 months today. In ERCOT (Texas), the fastest grid in the US, a new request for hundreds of megawatts means ‘good luck getting power before 2031-2032.’

Here’s the uncomfortable truth: poles and wires have NOT become an exponential technology. Everything else in tech doubles and halves on a predictable curve. Transmission lines don’t.

“Poles and wires have NOT become an exponential technology.”

71% of Americans oppose data centers in their area, a higher number than oppose nuclear plants in their backyard. The NIMBY problem is real, and it’s slowing everything down.

But there’s a solution emerging: interruptible loads. Battery-buffered data centers that charge off-peak and can be curtailed when the grid is stressed. ERCOT passed this rule in June 2026. FERC sent letters to 6 other grids to adopt it. Ramez’s company Agentic has 10GW of qualifying sites. Interruptible status cuts interconnect time from 5-7 years to 12-18 months.

3. SOLAR + BATTERIES: THE FASTEST PATH

Solar went from $100 per watt in 1975 to 8 cents per watt today. That’s more than a 1,000x decline. But here’s the nuance Ramez drove home: the cost decline follows Wright’s Law — 30% cheaper for every doubling of cumulative scale… NOT Moore’s Law. With 4-6 doublings left in the pipeline, we’re looking at 4-8x cheaper, not another 1,000x. Still enormous, but let’s be precise.

Batteries are down 14x since 2010. Sodium-ion could cut another 10x. The cost curves are still running.

The UAE built a 1GW 24/7 solar-plus-battery plant: 5GW of solar and 19GWh of batteries at $6 per watt capex. Compare that to the last US nuclear plant at $15/W, or the cheapest Chinese nuclear at $4/W. Solar plus battery is also the FASTEST energy project you can build, about 12 months from groundbreaking to electrons.

Chart 3: Capital cost per watt across energy sources. Source: Ramez Naam Moonshots podcast, IEA.

But here’s the real problem… WINTER. London gets one-sixth the insolation in January compared to July. Batteries solve day/night. They don’t solve seasons. That’s where the solar-plus-battery story hits its ceiling, and where nuclear and geothermal pick up the slack.

4. NATURAL GAS: THE BRIDGE IS SOLD OUT

Large 400MW gas turbines are sold out 7 years ahead. GE Hitachi is adding assembly lines to keep up. Even 38MW truck-mounted turbines have backlogs. Boom Supersonic pivoted their jet engine design into gas turbines for data centers. When a supersonic aviation company starts building power plants, you know the demand signal is intense.

Gas is the behind-the-meter solution happening right now. If you need power for your AI workload and you can’t wait for the grid, you buy a turbine, pipe in gas, and generate on-site. It works. But it’s a stopgap. It’s the bridge — and the bridge is sold out.

5. NUCLEAR: FISSION, SMRs, AND THE COMEBACK

The cheapest nuclear lever is currently stopping the shutdown of existing ones. Germany’s mistake was closing working reactors. France gets 80% of its electricity from nuclear power and exports the rest to Europe. That’s the model that works. Extend life. Restart plants like Three Mile Island. That’s step one.

Here’s the rule for new nuclear: the first unit of any new model runs over time and over budget. You need 3-5 builds to sort the design. SMRs? Companies say 2030. Ramez expects those to slip. As he puts it: ‘Construction is a dirty word. Manufacturing gets cheaper.’ Fully factory-built small reactors — Aalo, Radiant, are the most promising because they take construction out of the field and put it on an assembly line.

“AI data centers might be the best thing that’s ever happened to the nuclear industry.”

AI companies need firm, carbon-free, 24/7 power. Nuclear delivers exactly that. Microsoft is already signing PPAs with Three Mile Island. The demand side is locked and loaded. The supply side is the question, and SMRs and factory-built reactors are the most credible answer.

Chart 4: Expected timeline for each energy source reaching scale. Source: Ramez Naam, IEA, industry projections.

6. FUSION: NO LONGER 50 YEARS AWAY

The NRC regulated fusion like hospital imaging machines, NOT like fission reactors. That distinction matters enormously. Fission is default-on; if you lose control, you have a meltdown. Fusion is default-off; if you lose containment, it just stops. Huge regulatory unlock.

Helion is the most aggressive; however, they have a Microsoft PPA for 50MW in 2028. Ramez finds that timeline plausible, not guaranteed. Commonwealth Fusion is MIT’s nuclear engineering department privatized. They shrink the ITER concept from 5GW and $40 billion down to roughly 600MW. That’s the kind of order-of-magnitude cost reduction that changes the economics entirely.

Every startup exaggerates timelines.

Here’s the paradox: fission, which we know how to do, is a 2031-2032 thing. Fusion, which we’re still learning, has a 2028 target. The triple product temperature times pressure times duration — has been on a steady march upward since 1956. Fusion was creeping up on us and people weren’t paying attention. Now the private capital is flowing, the regulation is sorted, and the timelines are measured in years, not decades.

7. THE FIVE PATHS TO A TERAWATT

Ramez laid out his framework for getting to a terawatt of new clean power: equatorial deserts with solar plus batteries, nuclear (fission or fusion), space-based compute, ocean data centers, and geothermal, the dark horse.

Ocean: Pantellassa is building floating data centers in the Southern Ocean, wave-powered at 2 cents per kWh. Peter Thiel led the last funding round. The ocean is a massive heat sink and an infinite power source if you can engineer for salt and storms.

Space: 1GW of orbital compute would require roughly 6x SpaceX’s best launch year. Launch permitting is the ceiling — about 200 Starship launches per year max. The physics work. The regulatory and launch economics don’t yet. But give it a decade.

Geothermal: Fervo is doing enhanced geothermal today. Quaise is using plasma beams to drill ultra-deep, unlocking geothermal energy anywhere on Earth, not just in Iceland and California. If the drilling tech works, geothermal becomes baseload power available everywhere. That’s a moonshot worth watching.

WHAT THIS MEANS FOR YOU

Entrepreneurs: The poles-and-wires problem is unsolved. Startups that speed up grid construction — faster permitting, modular transmission, automated installation — are desperately needed. This is a multi-billion-dollar opportunity hiding in plain sight.

Investors: Energy infrastructure is the picks-and-shovels of the AI age. Gas turbines, battery buffering, solar-plus-battery projects, and nuclear restarts are where the real money flows. The AI gold rush needs pickaxes.

Executives: Lock in your power supply now. Interruptible load status is the fastest path to connection. If you’re building a data center and you haven’t talked to your utility about interruptible status, you’re already behind.

Policymakers: Incentivize speed of power delivery, not cost-plus capex. You get what you incentivize. Reward grids for connecting new load fast, not for building gold-plated infrastructure on 7-year timelines.

To a future of abundant energy,

Peter