Pre-seed · Robotics, Physical AI & AI Infrastructure
The autonomy that sets the cost of lunar helium-3
Ithil decides where to dig, and keeps machines digging where no one can repair them.
Benjamin Lee, Founder · ithilenergy.io
Why now
AI needs clean, firm power faster than the grid can deliver it
Global data-center electricity, TWh
+17%
Data-center electricity growth in 2025
200 MW
Google's fusion offtake with Commonwealth Fusion (2025). Microsoft signed with Helion in 2023.
19
NASA lunar landings planned for 2027 and 2028
Sources: IEA, Energy and AI (base case); company announcements; NASA CLPS manifest
The fuel
Helium-3 is the cleanest fusion fuel. It costs 20× too much.
Price per kilogram of helium-3
Today
Earth supply
What fusion can pay
To compete as fuel
88 GWh
Electricity from one kilogram
Few neutrons
A fraction of what deuterium-tritium fuel releases
Sources: DOE National Isotope Development Center; market pricing
The supply
Earth makes 3–4 kg a year. 2030's data centers alone would need 11,000.
One dot: Earth's entire annual supply, ~3.5 kg
The whole grid: helium-3 for 2030 data-center demand, ~11 t a year
0.7–1.7 Mt
Estimated helium-3 on the Moon, deposited by solar wind over billions of years
~11 t = IEA 2030 data-center demand (~945 TWh) ÷ 88 GWh per kg
The problem
Every kilogram means moving 100,000 tonnes of lunar soil
At 4–15 parts per billion, cost per kilogram is set by two things nobody has solved.
01
Finding the richest ground
Grade varies several-fold from site to site and has never been measured on the Moon itself. Every dig in the wrong place burns power and machine life.
02
Keeping machines working
No humans, abrasive dust, 14-day nights and a signal delay. No excavator has ever operated there. Apollo's seals failed after two EVAs; a commercial plant needs 5,000 hours.
Both are AI and robotics problems.
The solution
Ithil decides where to dig, and keeps the machines digging
Vendor-neutral autonomy software. It runs on anyone's excavator, rover or processing plant.
Find the richest ground
Orbital titanium and solar-wind maps fused with rover samples into a grade map with uncertainty. Active learning picks the next sample site.
Keep machines working
Remaining life estimated from force, current, vibration and thermal telemetry. Faults recovered on board. Work scheduled around the 14-day night.
Optimize both at once
The richest ground is often the rockiest and furthest from power. Ithil trades grade against wear and energy to maximize kilograms over the fleet's life.
Product
One planner for grade, wear and power
Orbital mapsTitanium, soil maturity
Rover samplesIn-situ grade
Machine telemetryForce, current, thermal
Grade modelMap with uncertainty
Wear modelRemaining useful life
Joint plannerMaximizes kilograms over the fleet's life, within the power and thermal budget
Dig
Haul
Process
Idle
Runs on NASA's cFS and F Prime flight frameworks. Ithil sits on top; it doesn't replace them.
Learns From Earth fleets first. Every processed batch becomes a ground-truth grade measurement.
The technical core
The hard part: learning without breaking machines
Formally, constrained stochastic control over a mixed fleet, where each machine's wear is hidden, irreversible and depends on how it's run.
Wear can't be measured, only inferred
Remaining life is estimated from force, current, vibration and thermal telemetry, a prognostics problem with training data on Earth.
The wear model changes by site
Tool–soil interaction in low gravity is poorly characterized, so the controller adapts online instead of running fixed gains.
Failure can't be explored
Failure is permanent and fleets are tiny. That rules out standard reinforcement learning: we use conservative policy improvement.
Open-loop digs; it doesn't optimize
Open-loop control is enough for site preparation. It can't trade output against remaining life under a power cap.
Market
Budgets for this autonomy exist today. The fuel market is the prize.
TAM · annual, 2030s
$10B+
Helium-3 fuel for 2030's data-center demand: ~11 t a year at a fusion-viable ~$1M/kg
SAM · programs buying it now
$20B
NASA Moon Base program (total)
$2B+
Nuclear-decommissioning robotics on Earth
SOM · year-3 revenue target
[$__]
[__] machines under license plus [__] operations contracts
TAM is annual; SAM figures are program totals. Sources: IEA; NASA
Competition
Everyone else optimizes one machine or one fleet
Vendor-neutral
Picks sites by grade
Plans for wear
Built for the Moon
Ithil
Yes
Yes
Yes
Yes
Interlune, with Vermeer
No, own fleet
Own sites
Not public
Yes
Magna Petra
No, own hardware
Not public
Not public
Yes
Pronto, ASI Mining (retrofit)
Yes
No, haulage
Not public
No
OEM autonomy (Cat, Komatsu)
No, own fleet
No
Not public
No
Moat: neutrality plus cross-fleet wear data that no single vendor can collect. NASA's cFS and F Prime aren't competitors; Ithil runs on them.
Ithil's capabilities are designed, not yet built. Competitor entries from public information.
Business model
Software revenue first. A royalty on every kilogram later.
From 2026
Software license
Per machine, per year, on any vendor's hardware.
[$__ per machine-year]
From 2027
Operations contracts
Per mission on the Moon, per site on Earth: planning, monitoring and model updates.
2030+
Production royalty
Per kilogram extracted, once operators are producing helium-3.
Software margins from day one: [target gross margin __%]. Early revenue doesn't depend on helium-3 production.
Go-to-market
Earth first, then the Moon
Now – 2027
Earth pilots
Paid pilots in nuclear decommissioning and mining: revenue plus the wear and energy data the models need.
Design partner: [Name]
2027 – 2029
Lunar missions
Fly on NASA Moon Base and CLPS missions through hardware partners. NASA SBIR/STTR funds early work.
2030+
Lunar extraction
License plus royalty with helium-3 operators as production scales.
Channel: sold through hardware makers as their autonomy layer, and directly to mission operators.
Traction
The system is already architected
Designed as a Johns Hopkins systems engineering project, with design feedback from lunar hardware engineers.
HERMES: a full model-based systems engineering architecture for a helium-3 extraction plant, in SysML
Requirements traced to verification, including 5,000-hour operation and fault detection in under 5 s
Trade studies, interface definitions, risk register and test plan
Design feedback from
Kris Zacny, Honeybee Robotics Dr. George Sowers, Colorado School of Mines A former Blue Origin architect
Next build
[e.g. wear estimator trained on public excavator telemetry, run on a DEM regolith simulation]
Team
Benjamin Lee, Founder
Systems engineer and data scientist. Ithil's two problems are a decision under uncertainty and a reliability problem. He has shipped both.
Reliability from telemetry
Built telemetry and reliability analytics for Microsoft Copilot at production scale.
Decisions under uncertainty
Pricing and revenue optimization with statistical models.
The system, designed
Johns Hopkins M.S. in Systems Engineering; designed HERMES end to end.
Hiring next
Founding robotics and controls engineer with field-autonomy experience. [Name]
Financials
Three-year plan
Revenue
Customers
Headcount
Net burn
Year 1
[$__]
[__] pilots
[__]
[$__]
Year 2
[$__]
[__]
[__]
[$__]
Year 3
[$__]
[__]
[__]
[$__]
Runway: [__] months on this raise. Non-dilutive: [NASA SBIR/STTR applications, if any]
The ask
Raising [$__] pre-seed
To build vendor-neutral autonomy for lunar extraction, starting on Earth.
Use of funds
Engineering, incl. robotics and controls hire[__%]
Simulation and bench rig[__%]
Pilots and partnerships[__%]
18-month milestones
Wear estimator validated on Earth fleet data
First paid Earth pilot
NASA SBIR award or hardware-partner LOI
Contact · [email] · ithilenergy.io
Appendix · Energy check
Extraction uses under 1% of the energy the fuel returns
Energy per kilogram of helium-3
Out
Electricity from fusion
In
Process heat, 85% recovered
Source: University of Wisconsin Fusion Technology Institute, Mark II/III lunar miner design (heat to 700 °C, 85% recuperation)
Rotate your phone for a larger view or swipe to move between slides