200GW Hiding in Grid, Sodium Batteries 10x Cheaper, Wave-Powered Datacenters w/ Ramez Naam | EP #280
Ramez Naam joins the Moonshots crew to argue that the real bottleneck for AI is no longer chips or even power generation, it's the grid: interconnection queues have stretched from 15 months to 45 months, and a new Texas data center request today might not get hooked up until 2031-2032. He walks through near-term fixes already underway, behind-the-meter gas turbines, new 'interruptible load' regulations that fast-track flexible demand, and batteries that time-shift power from night to day, alongside the longer arcs of solar-plus-storage economics, a nuclear fission renaissance built on small modular reactors, and a fusion industry racing toward reactors like Helion's 2028 Microsoft power deal. The group also stress-tests Elon Musk's vision of terawatt-scale AI compute in space, running the Starship launch-cadence math and finding it plausible only over 15-20+ years, and closes on an emerging wave-powered ocean-datacenter startup as a fourth path alongside desert solar, space, and fusion/fission. Throughout, Naam pushes back on some AI hype: compute-to-intelligence scaling is at best polynomial with no path to a hardware-free recursive-self-improvement takeoff, and today's models remain far less energy-efficient per unit of learning than the human brain even as they out-produce it on raw output.
AI is power-hungry, but energy is a bottleneck, not a costEnergy▶ 6:57
Power is only a small slice of a data center's all-up cost compared to GPUs, but it's the resource hyperscalers physically can't get enough of fast enough.
Compute-to-intelligence scaling laws and the limits of RSIAI▶ 9:40
Naam argues intelligence scales at best polynomially (not exponentially) with compute, and no valid model of software-only recursive self-improvement produces a takeoff.
The grid interconnection queue is the real bottleneckEnergy▶ 7:52
Generation and load interconnection wait times have exploded nationwide; the poles-and-wires distribution grid, not power plants, is what's actually constraining AI buildout.
Behind-the-meter power generation for data centersEnergy▶ 23:16
With large natural gas turbines sold out for years, hyperscalers and unlikely entrants like a former supersonic-jet startup are building their own on-site gas power.
New Texas rules fast-track grid connection for loads that can be turned off at peak demand; a widely-cited paper finds just 100 hours/year of flexibility can unlock 100GW nationally.
Falling solar and battery costs, plus emerging sodium-ion chemistry, are enabling the first affordable '24/7 guaranteed' solar-plus-battery power plants, like a 1GW project in the UAE.
Nuclear fission renaissance: big reactors and SMRsEnergy▶ 1:03:47
US policy is now backing both fleets of standardized large reactors and a wave of factory-built small modular reactors, with new designs that are passively meltdown-proof.
The fusion race: tokamaks, lasers, and pulsed field-reversed designsEnergy▶ 1:19:18
Over 50 venture-backed fusion startups are pursuing three broad approaches; the triple-product metric shows decades of real, non-hype progress toward net energy gain.
Space-based AI data centers and Starship launch mathSpace▶ 1:40:30
Elon Musk's vision of orbital AI compute is evaluated against actual launch-cadence numbers, finding it a plausible multi-decade hedge against terrestrial grid bottlenecks, not a near-term solution.
Ocean-based, wave-powered data centersEnergy▶ 1:54:20
A portfolio company builds bobby-pin-shaped floating data centers that generate power from ocean waves and get free GPU cooling from cold seawater.
AI energy efficiency versus the human brainAI▶ 1:59:32
A debate over whether frontier AI models are anywhere near as energy-efficient as the human brain, versus how much more raw output they produce per watt.
Predictions made
openRamez Naam: A request today for hundreds of megawatts to power a new data center in ERCOT (Texas), the fastest-moving US grid, will not receive power before 2031-2032.
openRamez Naam: Battery-backed, interruptible-load data center deals that time-shift demand off peak hours will go from an unusual approach today to a widely common industry practice.
“12 months from now this will be a super common approach”
Your call:
openRamez Naam: EV-charging-style demand flexibility and load timeshifting will unlock roughly 100GW of grid capacity for new AI data centers over the next five years.
“a power purchase agreement from Microsoft to provide 50 megawatts of power... in 2028”
Your call:
openRamez Naam: Nvidia's CUDA software moat will be effectively broken as AI-assisted code recompilation lets workloads run efficiently on rival chips like AMD and Cerebras.
“I think the CUDA moat is broken this year and next year”
Your call:
Numbers that matter
Building a 1GW data center costs ~$50B, of which ~$35B is chipsillustrates that power is cheap relative to GPU capex even though it's the bottleneck
Generation interconnection wait times grew from 15 months (20 years ago) to ~45 months todayUS grid queue delays for new power plants
ERCOT (Texas) grid peaks at ~80GW; over 200GW of load has been submitted to its interconnection queuemost submissions are speculative, but shows demand-side queue overwhelm
A hundreds-of-megawatts data center request in ERCOT today won't get power before 2031-2032even in the least-regulated, fastest-moving US grid
Projected AI chip power demand is ~200-275GW (call it 230GW) vs. projected US grid buildout of only ~100GW by 2030the gap between chip manufacturing pace and grid buildout pace
Eric Schmidt's public estimate: ~96-100GW of additional US power needed by 2030compared against Naam's chip-based 230GW estimate
Total data center power draw is nearly double the raw GPU draw once IT equipment and cooling are includeda common miss in power-demand forecasts
400MW-class natural gas turbines are sold out for about 7 yearsdriving the rush to behind-the-meter power generation
The 200GW grid capacity gap represents ~$10 trillion of AI capex, against an estimated $7 trillion of total AI capex over the next 5 yearsquantifying the value of unlocking grid capacity
Being flexible just 100 hours/year (4 days, 1% downtime) could unlock 100GW of grid capacity, worth ~$5 trillion in data center capexTyler Norris's grid-flexibility research finding
US grid demand averages ~500GW, ranging from ~400GW (winter nights) to ~600GW (summer afternoons, mostly AC)illustrates the daily/seasonal load swing that flexibility can exploit
Texas's June 2026 regulation cuts interruptible-load grid connection time from 5-7 years to 12-18 monthsnew fast-track interconnection rule for flexible loads
Solar panel cost fell from $100/watt in 1975 to about $0.08/watt today for Chinese panelsmore than a 1000x price decline over five decades
Battery prices have dropped 14x since 2010; sodium-ion could cut costs by another 10xbattery cost trajectory enabling solar-plus-storage at scale
UAE's 1GW '24/7' solar+battery plant uses 5GW solar plus 19GWh of batteries at ~$6/watt capexvs. $15/watt for the last US nuclear plant and $4/watt for China's cheapest nuclear plants
Worth digging into
🕳️ Tyler Norris's grid-flexibility research ('100 hours unlocks 100GW')
A single academic paper reframes the AI-power crisis as solvable via demand flexibility rather than new generation, with a claimed $5 trillion capex unlock, which could reshape how data centers get sited and financed.
🕳️ Texas's June 2026 interruptible-load interconnection rule and FERC's letter to the other six grids
If adopted nationally, this could be the single biggest near-term unlock for AI data-center power, cutting connection times from years to months.
🕳️ Helion's 50MW fusion power purchase agreement with Microsoft, targeted for 2028
It's the most aggressive fusion commercialization timeline discussed, and fusion's unique 'default-off' regulatory treatment (like medical imaging equipment) is a structural advantage over fission that's underappreciated.
🕳️ The unnamed wave-powered ocean data center startup (Peter Thiel-backed latest round)
Claims ~2 cents/kWh power plus free GPU cooling from cold seawater, and is framed as 'space-based solar, but in the ocean' — a genuinely novel compute-siting bet that isn't widely tracked.
🕳️ Commonwealth Fusion Systems' compact high-temperature-superconductor magnets versus ITER
The claimed ~600MW break-even point (versus ITER's 5GW/$40B) hinges entirely on the new magnet technology scaling as promised, making it the single biggest lever in the 'fusion is a when, not an if' thesis.
🕳️ Elon Musk's 100GW/year orbital-compute target and its launch-cadence math
Naam independently sanity-checked the numbers (roughly 30,000 launches/year, one every 15 minutes) and remains skeptical versus Musk's public framing, making it a clean test case for tracking hype versus reality.