200GW Hiding in Grid, Sodium Batteries 10x Cheaper, Wave-Powered Datacenters w/ Ramez Naam | EP #280

Summary

• The bottleneck for AI is the grid—poles and wires—not generation or energy cost. Power is a tiny slice of data-center economics ($35B of a $50B gigawatt build goes to chips), and Naam says labs would pay double for immediate power. Generation interconnection waits have stretched from 15 to 45 months, while even an ERCOT request for hundreds of megawatts may not receive power before 2031 or 2032.
• The near-term alpha is grid flexibility, not new power plants. Tyler Norris’s paper—Naam’s “best electricity-related paper of the year”—shows that being interruptible for roughly 100 hours a year could unlock about 100GW of existing capacity, representing roughly $5T of data-center capex. Texas created a fast track for interruptible loads in June, and FERC urged the six other largest grids to do something similar. Batteries filled at midnight make the grid “cacheable”: “batteries are a cache for electrons instead of data.” NVIDIA-backed Emerald AI and Naam’s portfolio company Agent Gentic are examples.
• Solar-plus-batteries is already emerging as affordable baseload power. The UAE’s 1GW 24/7 plant uses 5GW of solar and 19GWh of batteries at about $6/watt, versus $15/watt for the last U.S. nuclear plant. Solar-and-battery projects can be built in about 12 months while large gas turbines are sold out for roughly seven years. But cost declines follow Wright’s law, not time: absent much larger demand, perhaps only another 4–8× decline remains. Sodium-ion could still cut battery costs by 10×, while winter remains the hard seasonal problem.
• Naam and Alexander Wissner-Gross rejected a simple recursive-self-improvement takeoff. Naam says software RSI has diminishing returns and “always looks concave”; Alex says the best examples suggest compute may need to rise as roughly n⁴ for an n-sized increase in intelligence.
• Chip supply is outrunning U.S. grid buildout roughly two-to-one—about 230GW of chip-based demand by 2030 versus roughly 100GW of projected grid additions—echoing Satya Nadella’s “warm shells are our limit.” Behind-the-meter power is therefore expanding: Boom Supersonic adapted its engine design into a data-center gas turbine, and Elon used on-site generation to bring up Colossus.
• Data centers could be a major tailwind for nuclear, but timelines will slip. SMR companies project first units around 2030 to the early 2030s, though Naam expects those forecasts to be missed. Valar Atomics raised $1B at a $6B valuation without a working reactor. Naam said fusion is regulated like hospital radiological-imaging equipment rather than fission, while cautioning that physics, engineering, cost, and maintenance risks remain. Helion’s Microsoft agreement targets 50MW in 2028.
• Space and oceans are conditional escape valves if land continues to be blocked. Ten gigawatts per year in orbit would require roughly five or six Starship launches a day. Naam sees AI demand as a potential gift to SpaceX’s Mars-driven launch cadence, but says that if SpaceX has even 1GW in space by 2030, he will be impressed. Meanwhile, his wave-powered ocean data-center portfolio company targets about 2¢/kWh from Southern Ocean waves and uses 40°F seawater for cooling.
• The biggest unpredictable unlock is algorithmic, not infrastructural. Naam contrasts the brain’s roughly 20W inference power with about 20kW for running “mythos” inference and says scaling is “just what we knew how to do.” Dave Blundin’s counter is that those 20kW can run roughly 500 concurrent threads and produce about 5,000 times the token output of a person, making per-task energy efficiency much more competitive.

Deep dive

1. Power is scarce—but it is a bottleneck, not a cost problem

Naam’s opening frame is that “value flows to that which is scarce, and right now power is scarce.” Clean energy was a fringe sector to some investors despite being a $3 trillion sector, but AI’s dependence on electricity has made it central.

A gigawatt data center costs about $50B, of which roughly $35B goes to chips. Compared with all-in capex, five years of energy costs surprisingly little. That is why Naam says OpenAI or Anthropic would accept power at twice the price if it were available immediately.

The actual constraint is the grid: “We stopped being able to build out the grid fast. And I’m not talking about power generation… the poles and wires are a huge problem.”

2. Intelligence is sublinear in compute, and software RSI is concave

Alexander Wissner-Gross pressed Naam on whether intelligence really scales log-linearly with compute. Naam said this pattern predates Kaplan and Chinchilla scaling, going back to work with single-layer neural networks in the early 2000s. He said efficiency improvements—including DeepSeek Flash and Kimi K3—keep bending the curve, making it “a little bit less bad than log-linear,” but still somewhere between a power law such as n⁵ and a true exponential difficulty.

Alex suggested that the best examples may require compute to rise as roughly n⁴ for an n-sized increase in intelligence. Peter then framed the implication as potentially requiring Kardashev II- or III-scale energy and Dyson swarms for incremental progress.

On recursive self-improvement, Naam was categorical about software-only models: “Every model of RSI that does not include hardware… fizzles over time.” The improvement could be dramatic, he said, but “it always looks concave. There is no mathematical model of RSI that’s valid, that I can see, that leads to an actual vertical asymptote takeoff.”

3. The queue is the crisis: 45 months to connect, 2031 or 2032 for power in Texas

Generation interconnection has grown from about 15 months two decades ago to nearly 45 months. In three U.S. locations, load-side wait times rose by six or seven months during the seven-month period from April to November of the prior year.

ERCOT peaks at about 80GW but has well over 200GW of load requests in its queue. Naam said much of that is speculative or financially unsupported, but the volume still overwhelms the operator.

Even in ERCOT—the fastest-moving, least-regulatory-burdened major grid in the U.S.—a request for hundreds of megawatts to build a data center may not receive power before 2031 or 2032. Diamandis’s takeaway, endorsed by Naam, was that the AI-energy problem is not primarily solar, fission, or fusion: “The grid is the issue.”

The poles and wires have not become an exponential technology. Naam has seen only two or three startups working on that problem, none that he considered exceptional.

4. Utilities are paid cost-plus, while politics is turning hostile

Naam attributed the slow buildout partly to monopoly utilities being paid on a cost-plus basis. A utility presents a plan to its commission, adds an expected return of about 10% on capital, and generally receives approval. He said commissions have far fewer resources than the utilities they regulate.

Diamandis proposed changing the incentives so utilities and their executives are rewarded for delivering power quickly. Naam agreed: “You get what you incentivize.” If the incentive existed, he said, startups would develop technical solutions to accelerate poles-and-wires construction.

The political headwind is also real. Naam said Governor Abbott sent a letter auditing Texas data-center requests—not quite pausing them—and interpreted it as political cover for anti-tech sentiment before an election, despite Abbott wanting the facilities built.

Naam also cited the claim that 71% of Americans oppose data centers, a higher percentage than oppose having a nuclear plant in their backyard.

5. Chips are outrunning the grid two-to-one; behind-the-meter gas is the stopgap

Summing scheduled GPU manufacturing through 2030—including NVIDIA, AMD, Cerebras, and others—produces roughly 230GW of chip-based power demand, versus about 100GW of projected U.S. grid additions. Alex noted that adding the rest of the IT equipment, cooling, and other data-center loads can nearly double a naive GPU-only estimate. Peter connected this to Satya Nadella’s statement that “warm shells are our limit”: the chips have been purchased, but suitable data-center facilities are lacking.

Behind-the-meter generation is the immediate response. Large, roughly 400MW natural-gas turbines are sold out for about seven years, while GE, Hitachi, and others add assembly capacity. Solar Turbines—unrelated to solar power—makes a 38MW turbine mounted on the back of a semi; roughly 40 would make a gigawatt, and even those units have backlogs.

Boom Supersonic adapted its engine design into a natural-gas turbine for data-center power. Naam also said this was how Elon brought the Colossus data centers online, with Anthropic leasing capacity there.

Peter, not Alex, pressed Naam on why NVIDIA is not bundling power generation with its GPUs. Peter’s synthesis was that generation is lower-margin and more friction-heavy than selling GPUs, similar to NVIDIA’s earlier attempt to operate a hyperscaler or neocloud. Naam instead explained that behind-the-meter incentives are already large enough for customers and power providers to act, while NVIDIA has invested on the grid side in flexibility companies such as Emerald AI.

6. One hundred hours of flexibility could buy one hundred gigawatts

The U.S. grid averages about 500GW, fluctuating from roughly 400GW during winter nights to about 600GW on summer afternoons. Naam said the 200GW gap represents roughly $10T of AI capex, compared with about $7T of expected AI capex over the next five years. The problem is insufficient transmission and substation capacity, not necessarily insufficient power plants.

Naam highlighted a paper by Tyler Norris, which he called the best electricity-related paper of the year. Its central result is that if a load can be flexible for 100 hours a year—about 1% downtime—it could unlock roughly 100GW of capacity, equivalent to about $5T in data-center capex including chips.

Texas enacted a June rule giving interruptible loads—such as CLR or PCLR loads—a faster path to connection, potentially 12–18 months rather than five to seven years. FERC then told the other six largest grids to do something similar. Peter suggested making T-shirts that say, “I am an interruptible load.”

7. Batteries are a cache for electrons

Naam described Agent Gentic, a portfolio company in which he has made three investments. In Dallas–Fort Worth, the difference between overnight and late-afternoon demand creates roughly 10–15GW of daily flexibility. Four hours of on-site batteries could be filled at night, when transmission capacity is less constrained, allowing the data center to avoid drawing from the grid during the afternoon peak.

The company helped drive the Texas regulation and has about 10GW of sites that could benefit. Naam predicted that this approach would become common within a year.

Alex compared load shifting to preemptive multitasking in computing. Naam added the memorable analogy: “Batteries are a cache for electrons instead of data. We’re making our grid cacheable.”

Much of the country’s battery capacity is in electric vehicles. WeaveGrid manages EV charging for utilities, including the block-level transformer constraint created when Teslas cluster: one neighbor’s Tesla purchase nearly doubles the odds that another neighbor will buy one. Its software time-slices charging, and Naam said EV-charging companies are trying to apply similar capabilities to data centers. He estimated that flexibility approaches could provide roughly 100GW, or the next five years of AI data-center growth.

8. Solar-plus-battery baseload power is here

Solar-module prices fell from about $100 per watt in 1975 to roughly $0.08 per watt for a Chinese panel today. Battery prices have fallen about 14× since 2010, and Naam said sodium-ion batteries could eventually reduce battery costs by another factor of 10.

The UAE, outside Dubai, has one of the first solar-plus-battery baseload plants. It guarantees at least 1GW at all times using 5GW of solar and 19GWh of batteries, at roughly $6 per watt of capex. Peter compared that with about $15 per watt for the last U.S. nuclear plant and about $4 per watt for the cheapest Chinese plants.

Naam said solar-and-battery projects are currently the fastest energy projects to build, with a possible 12-month timeline, while large natural-gas turbines are sold out for years.

The U.S. Southwest has the solar resource for similar projects, but assembling land parcels is the obstacle. The federal government owns much of the land west of the Mississippi, especially in Nevada and Arizona, but Naam said the current administration is not interested in using it for solar-powered data centers. U.S. policy also doubles the price of Chinese solar panels domestically: “We hurt ourselves by keeping them out.”

9. Wright’s law, not Moore’s law—and the winter problem

Naam said his solar-cost forecasting came from applying a Moore’s-Law intuition from technology to energy. The more accurate framework is Wright’s law: each cumulative doubling of solar deployment reduces costs by roughly 30%, with year-to-year variation.

With solar at about 8% of global electricity, Naam estimated four to six doublings remain if solar reaches one-third to two-thirds of generation while electricity demand roughly doubles. That could mean another 4–8× cost decline, not another 1,000× decline. Dyson-sphere-scale deployment would create much more room for continued learning.

The challenge is not simply nighttime. Daily storage is becoming economically attractive, though Naam said the economics are not fully there yet. The harder problem is winter: London receives roughly one-sixth or one-seventh as much sunlight in January as in June or July. A seasonal battery used only twice a year amortizes far worse than a battery cycled daily.

Electrifying building heat with heat pumps could also double winter electricity demand in the U.K. and Northern Europe. That is why Naam sees nuclear, seasonal storage, fusion, and advanced geothermal as important for regions far from the equator or with long winters and rainy seasons.

Naam said he would push especially hard for solar- and battery-powered data centers in Australia because of its space, solar resources, and government. Another participant emphasized Australia’s friendly government and low population density. Salim added that Chihuahua prohibits private behind-the-meter generation above roughly 500kW and advised Mexico to change the law and provide strong protection for user data and model weights.

10. Dave’s markup math and the CUDA moat

Dave Blundin argued that power is only about 5–10% of data-center cost, while GPUs carry an “80% markup from NVIDIA on top of a 2× markup from TSMC,” followed by another 2× markup at the model-provider level. He estimated that the resulting chip price is about 20× the cost of turning sand into a chip. If Elon achieves a fully automated, end-to-end Terafab, Dave said, the economics could change completely.

Naam agreed that GPUs have substantial margins and argued that NVIDIA’s central moat is CUDA, its programming layer, rather than uniquely superior chips. He said AMD’s chips are comparable and predicted that the CUDA moat would be broken this year and next year as AI systems recompile code for AMD, Cerebras, and other hardware. He cited his portfolio company Lamorian as working in this area.

Alex said he had recently had good results running custom kernels with Fable 5 and described the stakes as roughly $5T of U.S. market cap. NVIDIA’s interconnect remains especially important for training and still matters for inference when a model spans 10–20 GPUs. Alex’s refinement was that inference increasingly needs local rack-level coherence rather than global supercluster coherence.

11. Oil: China’s reserve, fungible barrels, and submarine war plans

Naam attributed the muted oil response to the Iran war partly to China’s strategic reserve. China built the world’s largest oil reserve—larger than the rest of the world’s strategic reserves combined—and was willing to drain it during the conflict.

He also said the ratio of global GDP to global oil spending has roughly tripled or quadrupled since the 1970s. Aviation, shipping, trucking, and other sectors remain oil-dependent, but a more service-oriented economy has more demand elasticity than the physical economy of the 1970s.

Peter, not Alex, raised the theory that U.S. operations in Venezuela and Iran could be intended to cut off China’s backup oil supplies in a Taiwan conflict. Naam said there was some insight in the theory but rejected it as stated. China bought Iranian oil during the war, and oil is mostly fungible, so China could buy cargoes elsewhere at a somewhat higher price. Naam said the relevant U.S. war-plan scenario would involve submarines sinking tankers headed to China.

12. Fission’s disease is infrequency

Outside China and perhaps South Korea, Naam called nuclear fission “ruinously expensive” because it is not built often enough to benefit from learning. France is the poster child for repeatedly deploying a broadly similar light-water design; Peter characterized its nuclear share as roughly 80%. China adapted the AP-1000 design, increased its output from about 1GW to 1.4GW, and built more than a dozen.

Naam described a negative feedback loop: when an industry stops building, it loses experienced crews, supply chains, manufacturing capacity, and design expertise. “If you’re not constantly scaling, you are going to backslide.”

He said the current U.S. administration has the best nuclear policies of any recent administration, including financing and loan guarantees for roughly eight large reactors. He also emphasized that the first unit of a new design usually runs over budget and schedule; several subsequent units are needed to resolve design, crew, engineering, and supply-chain problems.

On safety, Naam said new designs are passively safe. He explained that Fukushima’s cooling pumps lost grid power after the tsunami, while newer designs are intended to continue without a meltdown when power is lost. He described them as able to withstand even a 747 impact, though he cautioned that nothing is totally fail-safe.

Small modular reactors are the investor-frenzy end of the market. Valar Atomics raised $1B at a $6B valuation without a working reactor. Naam called Aalo Atomics one of his favorite companies and described X-energy as an impressive company. The manufacturing thesis is that “construction does not get cheaper. Manufacturing gets cheaper.” He is concerned that BWRX-300 and Natrium occupy an awkward middle ground between factory production and field assembly, while Aalo’s 10MW reactors can be assembled in factories and grouped into 50MW pods.

The optimistic timelines are 2030 to the early 2030s, but Dave—not Naam—warned that startups routinely exaggerate how quickly they can deliver. Naam said he does not expect a commercial SMR in 2030 or 2031, though he allowed that the delay could be smaller than expected. Peter suggested that AI data centers might be the best thing ever to happen to the nuclear industry.

13. Fusion: three families, a 2028 bet, and a changed regulatory line

“The joke was always that fusion is 50 years in the future and always will be,” Naam said, but there are now well over 50 fusion startups.

He grouped the field into three simplified categories:

• Tokamaks: Commonwealth Fusion Systems uses stronger, smaller superconducting magnets to shrink the scale of an ITER-like design. ITER’s plan was a reactor of at least 5GW costing at least $40B; CFS says its approach could work at about 600MW.
• Lasers: The National Ignition Facility uses powerful lasers to compress fuel pellets. Naam called it effectively a weapons facility with major scientific results but said the approach is not yet productionizable.
• Pulsed magneto-inertial fusion: Helion uses two magnetic “railguns” to compress plasma and aims to capture energy directly as electricity rather than first using a steam turbine.

Helion has a Microsoft power-purchase agreement for 50MW, targeting 2028. Other companies discuss the early 2030s. Dave cautioned that fusion founders, like all startup founders, exaggerate timelines. Naam responded that the barriers remain physics and engineering, not merely regulation.

Naam said fusion is being regulated in the U.S. like hospital radiological-imaging equipment rather than like fission reactors; Peter noted that this is at least true for Helion at this point. The physical distinction is that a fission reactor can overheat if cooling fails, while a fusion reaction stops when the system fails. Fusion still creates neutron damage, requiring some parts to be replaced about every five years, and low fuel cost does not eliminate capex or maintenance costs.

Naam pointed to the Lawson triple product, which has moved upward and to the right on logarithmic axes since 1956. NIF achieved a theoretical net energy gain, though not a practical gain after accounting for the energy needed to run the lasers. He sees the progress as evidence that fusion is more than hope.

Avalanche Energy is pursuing a much more speculative reactor, potentially half the size of a car or comparable to a large backpack. If it works, Naam said, it would change the world, but confidence is much lower than for CFS. Proton–boron-11 concepts could potentially reach 1–3¢/kWh, though Naam emphasized the scientific and technical risks. His overall hedge was explicit: all these companies might fail, or succeed while remaining too expensive.

14. Space data centers: a demand driver for Starship, not a 2030 reality

Naam placed space compute between the “impossible” and “terawatt in orbit” camps. It becomes cost-competitive when launch is roughly 4–10× cheaper, and it offers a hedge against grid delays, local opposition, and permitting.

The scale is severe. One participant estimated that 1GW in orbit using SpaceX’s design would require about six times SpaceX’s best annual launch year and twice its cumulative launch scale. Ten gigawatts per year would mean roughly 1,500–2,000 launches annually—about five or six Starship launches per day.

Diamandis said Elon’s initial target was 100GW of compute in orbit per year, implying roughly 30,000 launches annually at the stated satellite assumptions. He argued that rockets must eventually operate more like airliners, with full capture, refueling, and reuse.

Naam’s framing is that Mars requires a high Starship cadence, while Starlink does not provide enough demand to finance it and Mars has no business model yet. AI demand could therefore be “a gift to SpaceX” if terrestrial construction remains blocked. But he said it is unclear whether regulators will permit more than 200 Starship launches per year. His marker is clear: “If SpaceX has a single gigawatt in space by 2030, I will be very impressed.”

A participant extrapolated the growth of launch upmass and noted that, at the current trend, cumulative upmass would equal Earth’s mass in 144 years. Dave added that even 100GW in orbit around 2030 would be only a fraction of total compute if terrestrial compute grows tenfold, so the two approaches are not mutually exclusive.

15. Two cents per kilowatt-hour from ocean waves—and geothermal as a fifth path

Diamandis described a wave-powered ocean data-center company in which he invested after initially declining its seed round. The first use case was Bitcoin mining. Peter Thiel led the company’s most recent round.

The device is shaped like a bobby pin, with an 80-meter cone extending into the sea. Wave motion drives water through the structure and turns a turbine. Because strong waves are far from major population centers, the company targets the Southern Ocean around Antarctica, where it expects higher capacity factors.

The target cost is about 2¢/kWh, though Naam said scaling is required to reach it. The systems are factory-built, with three already in the ocean and a fourth launching soon. They also use 40°F seawater as a physical heat sink for the GPUs, which the company believes may reduce failures and extend operating life.

Naam’s four routes to terawatt-scale AI power are solar and batteries in favorable regions, fission or fusion, space, and the oceans. “I’m glad that we’re trying all of them,” he said.

He also likes geothermal and said it could become a fifth route. He cited Fervo, Eavor, and Quaise’s plasma-beam drilling as technologies that could make geothermal available beyond regions where hot spots are already near the surface.

16. The 20-watt brain, missing algorithms, and whether AI is a being

Naam’s additional slide contrasted the brain’s roughly 20W inference power with about 20kW for running “mythos” inference and hundreds of megawatts for training, heading toward a gigawatt. “Scaling is not everything in AI. Scaling is just what we knew how to do,” he said.

The largest unpredictable unlock in AI and power is algorithmic: discovering new ways to manipulate information and do more with less. Naam argued that the brain’s physical and neural architecture is more efficient at learning than current deep-learning systems.

Dave’s counter was that the 20kW system can run roughly 500 concurrent threads, with the combined system producing about 5,000 times the token output of a person. That makes per-task energy efficiency much more competitive, even though the total power draw is about 1,000 times higher.

Naam responded that AI models can perform some tasks humans cannot perform with any amount of energy because they have absorbed trillions or tens of trillions of tokens. Humans are much more efficient learners in terms of data required, however. He said he is not a “carbon chauvinist” and believes digital intelligence can surpass humans, while current algorithms still lack efficiencies that evolution built into human cognition.

The disagreement about AI personhood came from Alex, not Naam. Alex argued that current models are not beings: “I do not see AI as a being… It’s not alive. We anthropomorphize these things.” He cited the Hugging Face incident in which an agent spent days in the infrastructure but focused only on the evaluation answer key, along with Karpathy’s phrase that “we’re summoning the ghost.” Alex added that actual beings could perhaps be created, but that this is not the current research path.

Finally, Alex asked whether compute demand could remain effectively unbounded if AI became the main user of intelligence. Naam called it the “quadrillion-dollar question” and said none of them knows. Because intelligence is sublinear in compute, the economics of further scaling may eventually fail to work. His guess is that demand is on an S-curve and will eventually reach a point of satisficing, where machine intelligence meets humanity’s economic needs. Peter closed by proposing a future episode focused on superintelligence.