A USA ⇄ INDIA DECK
Energy
meets Compute
Compute is expensive. Power is cheap.
Speed to power is the real bottleneck.
USA slides — recreated from the talk
INDIA slides — companion view, with sources
Reference video: “200GW Hiding in Grid, Sodium Batteries 10x Cheaper, Wave-Powered Datacenters w/ Ramez Naam” Moonshots with Peter H. Diamandis, EP #280 · youtube.com/watch?v=LUFJE5QkINg
The main thesis
AI's limit isn't chips or money.
It's how fast we can plug in.
$50
capex per watt
vs ~$3 of power over 5 years. Waiting costs far more than electricity.
3–7 yrs
to connect to the grid
So builders go behind the meter: gas, batteries, solar.
~100 GW
hiding in the grid
If data centers flex ~100 hours a year.
14×
cheaper batteries since 2010
Solar + storage now looks like baseload. Nuclear, fusion and space come next.
IN
India has the same race but a different bottleneck: plenty of new solar, but too few transmission lines and too little firm evening power.
Thesis summarised from the episode and the presenter's slides.
USA
Themes
Agenda of the talk.
AI is
Power
Hungry
USA
Compute is
Expensive
Power is Cheap
USA
USA
AI CapEx / Watt vs 5 Year Energy Cost
As presented in the talk. 5-year energy cost ≈ 1 W × 43,800 h at US industrial power prices.
INDIA
India view
Same math in India: GPUs cost 10–20× their 5-year power bill
DATA CENTER POWER PRICE
₹7–10/kWh
state utility tariff
₹4–7/kWh
green open access / group captive
The chips are imported at global prices, so capex per watt is about the same as in the US. Cheaper power makes compute an even bigger share of the cost.
Power prices: GRI — India data centres: powering the 9 GW ambition. Our calculation: 1 W × 43,800 h × tariff, at ₹88/$, PUE ignored.
SPEED TO POWER
IS THE PROBLEM
3-7 YEARS ON GRID
USA
INDIA
India view
India builds solar in 18 months.
The wires take 3–5 years.
50 GW
of renewable capacity stranded with no grid connection (June 2025)
−42%
FY25 transmission lines built vs target (8,830 of 15,253 ckm)
USA
Source: Lawrence Berkeley National Laboratory, “Queued Up” — time from interconnection request (IR) to interconnection agreement (IA). Values approximated from the slide.
USA
Large-Load Grid Connection Waits Are Increasing
Over 7 months, delivery times increased by 6-7 months
Approximate values digitized from Bloom Energy data center power reports.
USA
ERCOT’s current large load interconnection queue
Source: Electric Reliability Council of Texas
INDIA
India view
India's data center load is set to grow 4–5× by 2030
ELECTRICITY USED
13 → 57 TWh
2024 → 2030
SHARE OF INDIA'S DEMAND
0.8% → 2.6%
AI-dedicated capacity: ~24× growth
USA
AI Power Demand vs. Projected Grid Buildout
North America
USA
2030 AI Demand vs. U.S. Grid Buildout
Illustrative central estimates with ±20% uncertainty
INDIA
India view
India 2030: there's enough power being built.
The gap is the wires.
271 GW
record peak demand, met on 21 May 2026
>300 GW
non-fossil capacity (Jul 2026), 54% of the total, heading for 500 GW by 2030
POWERING AI, FASTER
Behind the Meter
USA
Large Siemens gas turbine during assembly
Photo: Siemens press picture, Wikimedia Commons, CC BY-SA 3.0
US gas-fired turbine wait times as much as seven years; costs up sharply
Large, Multi-
100MW Turbines
USA
Siemens SGT-A45 trailer-mounted mobile gas turbine at Bayat Power-1
Siemens SGT-A45 mobile (trailer-mounted) 41 MW gas turbine, Bayat Power-1. Photo: Mahbubehp, Wikimedia Commons, CC BY-SA 4.0
ONE OPTION : SMALL GAS TURBINES
[~ 1/3 OF PROJECTS]
18 – 36 Month
Backlog
USA
INDIA
India view
India's power plants on standby: an idle gas fleet
20.1 GW
gas fleet left after ~5 GW of long-idle plants were retired (Apr)
12.5 GW
could be used as fast peaker plants to back up renewables
Behind the meter in India often means green open access
Captive or group-captive solar and wind contracts at ₹4–7/kWh, vs ₹7–10 from state utilities. Higher LNG demand from data centers is starting to bring stranded gas plants back.
USA
US GRID IS BUILT FOR PEAK, AND OFTEN UNDER-UTILIZED
INDIA
India view
India's spare grid capacity is at noon, not 4 a.m.
64 GW
midday drop in net load (2026), up from 25 GW in 2023
65 GW
evening ramp as solar fades, more than double 2023
Flexible AI loads in India should run hardest at midday and back off in the evening.
DUKE · NICHOLAS INSTITUTE · 2025
Rethinking Load Growth
Assessing the Potential for Integration of Large Flexible Loads in US Power Systems
Tyler H. Norris, Tim Profeta, Dalia Patino-Echeverri, and Adam Cowie-Haskell
Report NI R 25-01 · CC BY-NC 4.0
Read the report ↗
ABSTRACT (SUMMARY)
The study measures how much new, flexible load the existing US grid could absorb without building new capacity. It covers 22 of the largest balancing authorities, which serve 95% of US peak demand. Its key concept is “curtailment-enabled headroom”: new load that fits under planned peaks as long as it can be briefly scaled back during the tightest hours.
1.7–2.5 h
average length of a curtailment event
~90%
of curtailment hours still keep at least half of the new load running
PJM 18 GW
largest single region (at 0.5%), then MISO 15, ERCOT 10, SPP 10
USA
DUKE · NICHOLAS INSTITUTE · 2025
Rethinking Load Growth
Assessing the Potential for Integration of Large Flexible Loads in US Power Systems
Tyler H. Norris, Tim Profeta, Dalia Patino-Echeverri, and Adam Cowie-Haskell
Report NI R 25-01 · CC BY-NC 4.0
Read the report ↗
“100 for 100”
100 hours / year flexibility..
..unlocks  ~100GW of data center capacity. Rapidly.
~= $5 Trillion in DC CapEx
In the report: 98 GW at 0.5% curtailment (about 44 hours a year at full output) and 126 GW at 1% (about 88 hours).
USA
Duke Nicholas Institute, 2025 · “100 for 100” framing as presented in the talk.
USA
Agentic
Infrastructure
USING BATTERIES AND SPARE
OFF-PEAK  GRID CAPACITY TO
CONNECT DATA CENTERS FAST
1 MW / 4 MWh Turner Energy Storage Project, Pullman, WA
1 MW / 4 MWh Turner Energy Storage Project, Pullman, WA. Photo: UniEnergy Technologies, Wikimedia Commons, CC BY-SA 4.0
Fills Batteries at Night When Grid is Free
USA
INDIA
India view
Cheap panels made India a solar superpower: 53× in 12 years
44.6 GW
solar added in FY 2025-26, a record
#3
in the world for installed renewable capacity
MNRE installed capacity as of 31 March each year · SolarQuarter · 164.59 GW (Jul 2026) · PIB
BATTERIES, BATTERIES EVERYWHERE
USA
Lithium-ion battery pack of a BMW i3
Li-ion battery pack, BMW i3. Photo: RudolfSimon, Wikimedia Commons, CC BY-SA 3.0
USA
Lithium-ion battery pack price, $/kWh (real). BloombergNEF annual battery price survey.
INDIA
India view
India's battery storage tariffs fell 86% in three years
USA
Sodium batteries
INDIA
India view
India is betting on sodium: no lithium imports needed
RELIANCE × FARADION
Jamnagar gigafactory
Bought UK sodium-ion pioneer Faradion and owns 100% since Oct 2024. Production planned from 2026: grid-scale BESS, commercial and telecom packs.
KPIT × TRENTAR
3 GWh cell line
KPIT's own Na-ion cells reach 100–170 Wh/kg. Technology transferred in Feb 2025, aimed first at electric two-wheelers.
WHY IT MATTERS
Sodium is abundant everywhere. That means cheaper storage without depending on lithium supply chains.
ProjectSizeCostCost per wattStatus
Masdar / EWEC, Abu Dhabi (UAE)5.2 GWdc solar + 19 GWh battery → 1 GW firm 24/7~$6.0B$1.15 /Wdc
$6.0 /W of firm output
2027 (building)
Edwards & Sanborn, California (USA)875 MWdc solar + 3.3 GWh battery (1.3 GW interconnect)~$1.76B (phase 1+2 financing)$2.0 /Wdc solar
$1.35 /W of interconnect
Operating
Gemini, Nevada (USA)966 MWdc / 690 MWac solar + 380 MW / 1.4 GWh battery>$1B≥ $1.0 /Wdc
≥ $1.45 /Wac
Operating (2024)
Turpan, Xinjiang (China)900 MW PV + 100 MW thermal storage (phase 1 of 1 GW)~$0.84B (RMB 6B)$0.84 /W
(little storage)
Under construction
$/W figures are my own division of press-reported cost by nameplate; costs are mostly financing totals or announced budgets, not audited. Batteries are a different share of each project, so rows are not like-for-like. Firm 24/7 power costs far more per watt than daytime solar.
BASELOAD SOLAR + BATTERIES, NOW, IN ULTRA-SUNNY PLACES
UAE 1 GW 24/7 SOLAR + BATTERY PROJECT
COST: $6 / W.   [Design: 5GW solar + 19GWh battery for 1GW “baseload”]
USA
INDIA
India view
India is already buying 24/7 clean power for less than the grid tariff
₹5.25
/kWh SECI 1 GW round-the-clock auction (Aug 2026), fixed for 25 years
₹2.70
/kWh Morena (MP) firm, dispatchable solar + storage
Caveat: critics say “round-the-clock” contracts are not yet as firm as true baseload.
ON THE HORIZON
USA
Nuclear Fission and Fusion – The
Power of the Atom
USA
USA
U.S. Nuclear Fission Horizons
U.S. goal: +35 GW new nuclear by 2035 • Triple capacity by 2050
INDIA
India view
India's nuclear horizon: 8.8 → 100 GW by 2047
LARGE REACTORS
700 MWe PHWR
An Indian design and the backbone of the fleet. 10 units approved in 2017.
SMRs (BARC)
BSMR-200 · SMR-55
Plus a 5 MWth high-temperature gas reactor. At least 5 SMRs planned by 2033.
SHANTI ACT (DEC 2025)
Private sector in
Replaces the 1962 Atomic Energy Act. Private firms can now build, own and operate plants.
NUCLEAR ENERGY MISSION
₹20,000 crore
Budget 2025-26 funding to develop and deploy SMRs.
IN
Nuclear is ~3% of India's power today. Data centers could be its first private buyers.
USA
Projected First Units: Likely Cost per MWh & Schedule Risk
U.S. fission leaders — public targets, disclosed budgets, and recent slippage / inflation signals, mid-2026
USA
What We Do Repeatedly Gets Cheaper
The economics of fission: learning by building
INDIA
India view
India's flywheel: 10 identical reactors ordered at once
SANCTIONED COST (2017)
₹1.05 lakh cr for 7 GW
≈ ₹150/W ≈ $1.7/W*
US FIRST-OF-A-KIND (VOGTLE 3&4)
>$30B for ~2.2 GW
≈ $14/W
*2017 sanction cost converted at ₹88/$. Not inflation-adjusted, and not like-for-like.
NUCLEAR FUSION?
$6 Billion of Private Capital
> 40 Private fusion companies
USA
USA
After the Fusion Industry Association global fusion industry report (timeline & company count). Simplified.
USA
Major Fusion Design Families
A simple map of the main approaches to fusion energy
USA
Fusion Startups: Claimed Timelines to Power
Leading private fusion companies and their stated commercialization milestones, mid-2026
ⓘ Timelines are company-stated / aspirational.
INDIA
India view
India's fusion effort is state-led, and private startups are just starting
Tokamak schematic
IPR GANDHINAGAR
ADITYA-U · SST-1
India's own tokamaks. SST-1 is the country's first fully superconducting tokamak.
ITER cryostat (3D render with human figure for scale)
ITER MEMBER SINCE 2005
3,850-t cryostat
India built ITER's largest component, a 30 m × 30 m vacuum chamber.
Model of the PRAGYA tokamak vacuum vessel
PRANOS FUSION (2024)
PRAGYA tokamak
India's first private tokamak. Compact (R = 0.40 m), $417k angel round in 2025.
IN
US: 40+ companies, $6B. India: one startup. Its edge is ITER-grade manufacturing.
Images: tokamak schematic, Smith & Cowley, CC BY 3.0 (Wikimedia Commons) · ITER cryostat render, Image: ITER Organization · PRAGYA model, Photo: Soumya Pillai / ThePrint. The first schematic is illustrative, not the Indian machines.
USA
Lawson parameter vs ion temperature for fusion experiments (Wurzel and Hsu) TokamakStellaratorRFPZ PinchMagLIFFRCMTFSpheromakPinchMirrorSpherical TokamakLaser Indirect DriveLaser Direct Drive
WORKING
NUCLEAR FUSION
(NET ENERGY GAIN)
Net Energy Gain is <5 years away.
Commercial reactors: <10
Economics: Good on paper.
COULD ADDRESS FINAL 30%+ OF ELECTRICITY
STILL MANY UNKNOWNS ON CONSTRUCTION & ECONOMICS
Plot: Wurzel & Hsu, “Progress toward fusion energy breakeven and gain as measured against the Lawson criterion”, Phys. Plasmas 29, 062103 (2022); 2021 preprint version via Wikimedia Commons, CC BY-SA 4.0.
OUT OF THIS WORLD COMPUTE
USA
Rendering of a hexagonal-module orbital data center in sun-synchronous orbit above Earth
Artist rendering: hexagonal solar data-center array in sun-synchronous orbit. Wikideas1, Wikimedia Commons, CC0
SPACE BASED DATA CENTERS!
PROs
-   No grid delays
-   No local opposition
-   No terrestrial permitting
-   Terawatt-scale possible
CHALLENGES
-   Launch scale 🤯
-   Launch permitting
-   Launch costs must drop
-   Maintenance
USA
USA
How much launch mass for 1 GW of orbital AI?
Compared with SpaceX’s best annual launch pace — and humanity’s best pre-SpaceX year
INDIA
India view
India's first orbital data centers: pilots in 2026
200 kg
Pixxel × Sarvam AI — “Pathfinder”
Orbital platform to train and run LLMs in space. Launch targeted by end of 2026.
Pilot
NeevCloud × Agnikul Cosmos
AI compute modules on Agnikul's rocket upper stage, powered by solar in orbit.
$5M
TakeMe2Space
Seed round. Has already run AI in orbit on ISRO's POEM platform.
Computer Weekly · Pixxel × Sarvam · TakeMe2Space · ISRO LVM3 (~8 t to LEO; our estimate)
Panthalassa Ocean-2 wave-powered node at sea
Panthalassa Ocean-2 prototype at sea. Photo: Panthalassa, via Fox News
OCEAN DATA CENTERS: PANTHALASSA
PROs
-   No grid delays
-   No local opposition
-   No terrestrial permitting
-   Terawatt-scale possible
-   Ultra-low-cost power
-   Free conductive cooling
-   Chips run colder!
CHALLENGES
-   Entirely novel approach
USA
INDIA
India view
India's ocean power: ~40 GW of wave potential, still at the research stage
Survey of India Political Map of India, 13th edition 2026
Kavaratti, Lakshadweep
10 MW OTEC plant
Chennai
wave buoy
A&N Islands
no project yet
Map: Survey of India, Political Map of India, 13th ed., 2026 (© Government of India). Project markers added.
~40 GW
wave energy potential along India's ~7,500 km coastline
10 MW
OTEC power plant, project report under the Deep Ocean Mission
Kavaratti, Lakshadweep: India's first OTEC plant
NIOT is using the ocean's temperature gradient to run desalination off-grid (100 m³/day). A wave-powered navigation buoy is already working off Kamarajar Port, Chennai.
IN
No ocean data center projects yet. Islands are the natural first test sites.
USA
OPPORTUNITY: AI ENERGY USE vs HUMAN BRAIN
Lateral view of the human brain, lobes colour-coded
Frontal
Parietal
Occipital
Temporal
Nvidia GB200 NVL72 liquid-cooled rack with 72 GPUs, door open showing compute trays
AI rack:
72 GPUs,
liquid-cooled
Brain lobes: Gray's Anatomy fig. 728, H. V. Carter, public domain (Commons)
Nvidia GB200 rack (72 GPUs). Pokiiri, Wikimedia Commons, CC BY-SA 4.0
20 Watts
20 kWatts –  Inference
Gigawatts –  Training
INDIA
Synthesis
Same race, different bottlenecks
Summary of the preceding slides. Sources are cited on each slide.
References
Sources