2026-08-15

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

Energy investor and futurist Ramez Naam joins the Moonshots panel for a deep dive on the power bottleneck strangling AI growth: the grid, not generation, is the constraint, with interconnection queues stretching to 45 months and a ~200GW gap between chip-driven demand and planned US grid buildout. Naam walks through near-term fixes (behind-the-meter gas turbines, battery time-shifting, Texas's new interruptible-load regulations) and longer-horizon bets (solar-plus-battery mega-plants, sodium-ion batteries, small modular reactors, fusion, space- and ocean-based data centers). The panel presses him on Nvidia's CUDA moat, the economics of vertically integrated power-and-compute plays like Elon Musk's, and whether super intelligence will ever be resource-unbounded. Naam's throughline: power is cheap relative to chips, the poles-and-wires bottleneck is underinvested and fixable with the right incentives, and no single energy source wins outright — solar, nuclear, fusion, space, and ocean power all need to scale in parallel.

▶ Watch on YouTube

Topics

AI is power hungry, not power-cost hungry Energy ▶ 6:04

Naam argues energy is cheap relative to GPU capex (chips are $35B of a $50B gigawatt data center), so the real issue is that energy is the bottleneck constraining how much AI compute can be built, not its cost.

Grid interconnection queue is the real chokepoint Energy ▶ 7:56

Generation interconnection wait times have grown from 15 months 20 years ago to nearly 45 months today; in Texas (ERCOT), the fastest-moving grid in the US, a new large-load data center request won't get power before 2031-2032.

Politics of data center buildout Energy ▶ 17:42

Rising local and voter opposition to data centers, including Texas Governor Abbott ordering an audit of data center power requests, is described as political cover rather than genuine opposition to AI infrastructure.

Behind-the-meter power generation Energy ▶ 23:19

With grid connections taking years, data center developers are building their own natural gas turbine power on-site; large turbines are sold out for ~7 years, pushing companies like Boom Supersonic to pivot jet-engine expertise into gas turbines for data centers.

Grid flexibility and interruptible loads unlock hidden capacity Energy ▶ 28:24

New Texas regulation (June 2026) fast-tracks grid connection for interruptible loads that can shed power at peak demand; FERC has told six other major US grids to adopt similar rules. Being flexible just 100 hours/year can unlock 100GW of grid capacity worth roughly $5 trillion in data center capex.

Batteries as a cache for electrons Energy ▶ 33:58

Startups like Agentic Infrastructure are pairing data centers with on-site batteries charged at night to avoid drawing on congested daytime transmission lines, likened to preemptive multitasking or cache prefill in computing.

Solar-plus-battery data centers and sodium-ion batteries Energy ▶ 38:44

Solar cost has fallen 1000x since 1975; battery prices are down 14x since 2010 and sodium-ion could cut costs another 10x. The first gigawatt-scale 24/7 solar-plus-battery plants (UAE, Chile) are now cost-competitive with nuclear.

Where to site solar-powered AI data centers Energy ▶ 44:19

Naam argues the US southwest, Australia, Mexico, and Chile have world-class solar resources; the main US barrier is land-parcel permitting, not cost, and countries must guarantee IP/data protection to attract hyperscalers.

The winter problem for solar Energy ▶ 52:18

Solar's real weakness isn't nighttime (batteries solve that) but winter, when regions like London get 1/16th the insolation of summer; electrified heating could double winter electricity demand, making nuclear, fusion, and geothermal essential complements to solar in higher latitudes.

Nuclear fission: large reactors and SMRs Energy ▶ 63:47

Fission is cheap only where built repeatedly (China, France); the US is financing fleets of AP-1000-style large reactors while a wave of small modular reactor startups (Oklo, X-Energy, Natrium, Radiant) bet that factory-built, standardized designs will cut costs fastest. New Gen 3+/4 designs are passive-safe.

Fusion's three technical families and accelerating timelines Energy ▶ 79:18

Over 50 venture-backed fusion startups now exist, split into tokamaks (Commonwealth Fusion), lasers (NIF), and pulsed/field-reversed designs (Helion, Avalanche Fusion). Fusion is regulated more leniently than fission because a malfunction simply stops the reaction rather than causing a meltdown.

Space-based and ocean-based AI compute Space ▶ 100:34

Space compute becomes cost-competitive only once launch costs fall 4-10x; Elon Musk's ~100GW/year orbital compute target would require ~30,000 Starship launches/year. A wave-powered ocean data center (Naam portfolio company) targets 2 cents/kWh using Southern Ocean waves and free seawater cooling.

AI vs. brain energy efficiency and the next real unlock AI ▶ 119:32

The human brain runs inference on ~20 watts versus ~20 kilowatts for an equivalent GPU rack, but that rack produces far more tokens per second than a person; Naam argues the biggest unpredictable future unlock is algorithmic/architectural efficiency, not more power.

Predictions made

open Ramez Naam: Scheduled GPU manufacturing volume (Nvidia, AMD, Cerebras) through 2030 implies roughly 230GW of chip-driven US AI power demand, while projected US grid buildout will only add about 100GW.

· due: 2030 · ▶ watch

“US grid buildout is projected at roughly 100 gigawatt... there's 230 gigawatts of demand there.”

Your call:

open Ramez Naam: Battery-backed, off-peak-charging data centers (currently a novel approach pioneered by companies like Agentic Infrastructure) will become a very common industry practice.

· due: 2027 · ▶ watch

“12 months from now this will be a super common approach.”

Your call:

open Ramez Naam: Grid flexibility approaches, including EV-charging-style load shaping, will supply about 100GW of the next five years of AI data center power growth.

· due: 2031 · ▶ watch

“I think that's 100 gawatts... that's the next five years of AI data center growth.”

Your call:

open Ramez Naam: Nvidia's CUDA software moat will be effectively broken as AI-assisted code recompilation lets developers efficiently target AMD and other non-Nvidia chips.

· due: 2027 · ▶ watch

“I think the CUDA moat is broken this year and next year.”

Your call:

open Ramez Naam: No company will actually have a commercial small modular reactor operating, despite industry projections targeting the early 2030s.

· due: 2031 · ▶ watch

“I don't expect anyone to actually have a commercial small modular reactor in 2030, 2031.”

Your call:

open Ramez Naam: Helion will deliver on its power purchase agreement to provide Microsoft with 50 megawatts of fusion power.

· due: 2028 · ▶ watch

“The most aggressive timeline is Helion. They have a power purchase agreement from Microsoft to provide 50 megawatts of power... in 2028.”

Your call:

open Ramez Naam: Around 2030, SpaceX aims for 100 gigawatts/year of orbital AI compute capacity, roughly equal to all of Earth's total compute today, requiring about 30,000 Starship launches per year (one roughly every 15 minutes).

· due: 2030 · ▶ watch

“His target was 100 gawatts per year of compute in orbit initially, equivalent to the entire compute today.”

Your call:

open Ramez Naam: SpaceX will not get more than roughly a single gigawatt of AI compute into space, which would already be an impressive outcome relative to Musk's much larger stated ambitions.

· due: 2030 · ▶ watch

“I don't think we're going to have... by 2030 if SpaceX has a single gigawatt in space, I will be very impressed.”

Your call:

Numbers that matter

$50B gigawatt data center, ~$35B of that is chips Illustrates that energy cost is a small fraction of AI data center capex compared to GPUs.
Interconnection queue time grew from 15 months (20 years ago) to ~45 months today Measures how much slower it has become to connect new generation to the US grid.
ERCOT (Texas) peaks at ~80GW but has over 200GW of (mostly speculative) load requests in its demand queue Shows the gap between grid capacity and speculative AI data center demand submissions.
Chip-driven AI power demand ~230GW by 2030 vs. projected US grid buildout of ~100GW Naam's estimate of the demand-supply gap based on scheduled GPU manufacturing.
Large behind-the-meter natural gas turbines sold out for ~7 years Explains why data center operators are turning to smaller mobile turbines and other workarounds.
~200GW grid capacity gap ≈ $10 trillion of AI capex potential; ~$7 trillion total AI capex expected over next 5 years Frames the financial scale of the grid bottleneck.
Being flexible 100 hours/year (1% downtime) unlocks 100GW of grid capacity ≈ $5 trillion in data center capex Tyler Norris's '100-by-100' finding on flexible interconnection.
Battery prices down 14x since 2010; sodium-ion could cut costs another 10x Trajectory of battery cost declines enabling solar-plus-storage data centers.
UAE 1GW 24/7 solar+battery plant: 5GW solar + 19GWh batteries at ~$6/watt capex, vs. $15/watt for the last US nuclear plant and ~$4/watt for the cheapest Chinese nuclear builds Cost comparison showing solar-plus-battery now competitive with nuclear.
Earth receives about 8,000x more solar energy in a year than humanity consumes as a species Illustrates the scale of latent, currently unusable energy abundance.
ITER-class tokamak needed ~5GW scale and $40B+ cost; Commonwealth Fusion scaled break-even down to ~600MW Shows how compact magnet technology shrinks the minimum viable fusion reactor size and cost.
71% of Americans oppose nearby data centers, higher than opposition to a nearby nuclear plant Quick stat on public sentiment toward AI infrastructure vs. nuclear power.
Getting 1GW of AI compute into orbit requires ~6x SpaceX's best annual launch year and 2x its cumulative launch history to date Illustrates how far current launch cadence is from space-based AI compute at scale.
Human brain: ~20 watts; equivalent-output GPU rack (12 GPUs): ~20 kilowatts, producing ~5,000x more tokens/output than a person Naam and Dave Blundin compare human brain vs. AI inference energy efficiency and output.

Worth digging into

🕳️ Tyler Norris's '100-by-100' flexible-load research paper

Naam calls it the best electricity paper of the year — it quantifies precisely how much data-center flexibility unlocks grid capacity without building any new poles and wires.

🕳️ Agentic Infrastructure and Texas's new interruptible-load regulation

First mover exploiting Texas's fast-track rule for interruptible loads with 10GW of qualifying sites; FERC has told six other major US grids to adopt something similar, which could reshape data center siting nationally.

🕳️ Commonwealth Fusion, Helion, and Avalanche Fusion's competing bets

Three very different technical approaches (tokamak, pulsed/field-reversed, ultra-compact) with wildly different timelines and risk profiles, including Helion's 2028 Microsoft power purchase agreement.

🕳️ Wave-powered ocean data centers

Naam's portfolio company (backed by Peter Thiel's most recent round) targets 2 cents/kWh using Southern Ocean wave energy near Antarctica plus free seawater GPU cooling that reportedly reduces chip failures.

🕳️ Sodium-ion batteries reaching grid and data-center scale

Could cut battery costs another 10x beyond the 14x drop since 2010, which would meaningfully change the economics of solar-plus-battery AI data centers and seasonal storage.