Catch the fire at one acre, not one hundred thousand.
A grid of tethered aerostats holds a continuous 360° watch over California's fire country. On-board AI sees the smoke column in the first ninety seconds. A firefighting drone parked at the same mast is over the flame in under ten minutes, while the fire is still one truckload of water wide.
Detection is nearly solved. Response is not.
California already watches itself well. What it cannot do is reach a new fire before the wind does. Between the alert and the first drop of water sits a gap of twenty to forty minutes, and that gap is where the entire loss column comes from.
The watch already exists
ALERTCalifornia runs 1,260 mountaintop cameras, 915 of them with AI smoke detection, and flagged roughly 3,600 fires in 2025, more than half of them before any 911 call. Detection is not the bottleneck any more.2
The response is still ground-bound
An engine or a helicopter still has to drive or fly from a station that may be 40 km away over mountain roads. In a Santa Ana or a Diablo wind event, a fire doubles in area roughly every few minutes. The asset arrives to a different fire than the one that was reported.
Where the twenty minutes go
Table view
Put the eye and the extinguisher on the same mast.
The insight is not a better camera. It is that the camera and the firefighting drone should occupy the same point in space, so that the moment detection happens, the response is already inside the cell it needs to reach.
Continuous 360°
Each aerostat holds station at 450 m above ground with a gyro-stabilised turret: 4K visible, near-infrared, and cooled dual-band MWIR/LWIR thermal. Not a 2-minute sweep but a continuous stare, day and night, powered up the tether.
Ninety seconds
Two channels in parallel. Edge AI classifies the smoke column against the terrain behind it, while a cooled dual-band MWIR/LWIR head looks for heat that should not be there. Heat shows at the instant of ignition and works at night; smoke carries further and sees over ridges. Because the median fire is inside the field of view of about three nodes, parallax fixes its position to under 100 m with no terrain model needed.
Two minutes to airborne
A recon VTOL leaves the garage first and confirms with thermal. Suppression drones follow. Batteries are swapped robotically and charged from thin-film solar on the envelope, so the node is never waiting on a grid connection.
Under ten minutes
Three drones, 45 kg of Class A foam each, delivered onto a fire that is still measured in square metres. Release authority stays with a human at the watch floor. If it does not hold, CAL FIRE now has a confirmed fix and a live overhead feed before the first engine rolls.
What one of these actually is
Not a weather balloon. A streamlined, all-weather tethered aerostat (the same class of platform that has flown persistent surveillance for decades), carrying a sensor turret, a solar array and a drone garage, held on a tether that also carries its power and its fibre.
Key
- Helium envelope: 320 m³, all-weather to 60 kt
- Thin-film photovoltaics: 18 kWp on the upper surface
- Tail fins: passive weathervaning into wind
- Sensor turret: 4K visible + NIR + cooled dual-band MWIR/LWIR
- Drone garage: shelter, robotic battery swap
- Tether: load member + power conductors + fibre
- Winch & mooring: 20-minute downhaul on command
- Battery bank: 120 kWh, rides through the night
- Drone pad: 3 suppression VTOL + 1 recon VTOL
- Helium reclaim: closed-loop recovery on downhaul
Design point
- Station altitude
- 450 m AGL
- Detection radius
- 30 km
- Response radius
- 18 km
- Cell area
- 842 km²
- Payload mass
- 40–120 kg
- Availability
- ≥ 92%
- CAPEX / node
- $1.03M
- OPEX / node·yr
- $160k
Two bands, two different jobs
Smoke and heat are not alternatives, and the choice of infrared band matters more than the choice to use infrared at all. Heat exists at the instant of ignition, where a smoke column takes minutes to build and is close to invisible at night. Heat is also stopped by any ridge, where smoke rises over it. Carrying both costs less than picking the wrong one.
| Band | Flame contrast | vs sun-baked rock |
|---|
The Planck curve rises far more steeply with temperature at 4 µm than at 10 µm, which is why satellite fire products (MODIS, VIIRS) work in the 4 µm band. The original turret specification here called for LWIR only. That was the wrong band and has been corrected to a cooled dual-band MWIR/LWIR head.
| Range | Pixel on the ground | Fire for a 1% signal |
|---|
At 30 km a single pixel covers 37 m² of ground, and a fire the size of a matchbox still lifts it by one percent. Sensitivity sits three to four orders of magnitude away from being the constraint. What actually limits thermal detection is atmosphere, clutter and terrain.
| Station altitude | Smoke column | Ground-level heat |
|---|
Why this does not change the fleet size
Keeping smoke as the long-range layer holds the detection radius at 30 km, comfortably above the 18 km response radius. Drone speed stays the binding constraint and the fleet stays at 148 nodes.
Going thermal-only would cut the detection radius to 16 km, which falls below the response radius. Detection would become binding for the first time and the fleet would need 189 nodes, a 28% increase to buy a system that is blind behind every ridge.
Station keeping earns its keep a second time
Because a tethered node holds position to ±20 m, every pixel maps to a fixed patch of ground permanently. That allows a thermal baseline per location, "this rock reads 65 °C at 14:00 in July", and anomaly detection against it rather than an absolute threshold. A free-flying balloon can never build that baseline, because its pixels never look at the same place twice. This is a second and entirely independent argument for the tethered architecture.
Sun glint is the hard false alarm
Sun-baked rock is handled by band choice alone. Specular glint off water, glass, metal roofing and solar panels is not: it is spectrally a 5,800 K source, and temperature by itself cannot separate it from flame. Three defences, in order of strength: glint is strongly angle-dependent so it does not appear from two viewpoints at once, and the lattice already gives 2.8× node overlap; the dual-band Dozier method solves sub-pixel area and temperature together, which separates a 900 K flame from a reflection; and the per-pixel baseline learns where glint recurs and when.
Above 450 metres, altitude stops buying anything.
Three separate ceilings limit how far a node can see, and they bind in a specific order. Climbing higher defeats terrain masking, but only until the sensor itself runs out of pixels. Past that point the extra altitude costs regulatory pain and buys nothing.
| Station altitude | LOS horizon | Terrain factor | Effective r | Nodes to cover SRA | Airspace |
|---|
Two constraints, one crossing point
The finding that sizes the whole programme
At 450 m the detection radius reaches 30 km, which would cover California with just 51 nodes. But a drone flying at 150 km/h can only reach 18 km inside a ten-minute budget.
So the fleet is sized by how fast the drone flies, not by how far the camera sees. That is what sets 148 nodes, and it means detection comes with roughly 2.8× redundant overlap for free, which is exactly what stereo geolocation needs.
Where they go, and how many
A hexagonal lattice laid over California's fire-prone terrain. Move the response-time budget and the lattice re-sizes itself; every number below recomputes from the node count the lattice actually produces.
Belts are hand-drawn generalisations of California's fire terrain, scaled so their total equals the 31-million-acre CAL FIRE State Responsibility Area (125,453 km²). They are an illustrative planning geography, not the official Fire Hazard Severity Zone shapefile. Real siting would run against FHSZ, LANDFIRE fuel models, a viewshed analysis and land access.
Tethered or patrolling?
A free-flying balloon is cheaper per unit, flies far higher and sees further. It also loses the one thing the whole concept depends on: a fixed address for the drone to come home to.
| Dimension | Tethered aerostat (stationary) | Free-flying (patrolling) | Winner |
|---|
Verdict: tethered, with a small free-flying surge tier
Tethered wins on every axis that carries value. Power up the tether means an unlimited energy budget and a 120 kg payload instead of 25. Absolute station-keeping means the drone garage is always in the cell it serves. Fibre in the tether means zero-latency full-motion video with no satellite bill.
A free flyer can do none of this. It drifts, so it cannot host a drone base; it carries everything it burns, so its payload is a fraction; and it needs continuous airspace coordination rather than one permanent charted obstruction.
Recommendation: build the fleet tethered, and hold a tier of roughly 12 free flyers as a red-flag surge layer over Modoc, Inyo and the far northeast, terrain where node density will never pay for itself and detection-only coverage is still worth having.
Cost of continuous watch
Table view
A third option, a tethered multicopter on a 100 m line, is cheaper again but cannot lift a useful payload, cannot stay up in wind and has no room for a drone garage. It suits incident-scene comms, not a permanent watch.
The arithmetic of not letting it get big
California's fire loss is not spread across seven thousand fires. It is concentrated in about three. The entire economic case is one question: how often can you push a fire down one size class?
| Final fire size | Events / yr | Suppression / event | Damages / event | Total / event | Annual total | % of loss |
|---|
Three fires a year carry two thirds of the bill
| Downshift | Acres avoided | Loss avoided / event |
|---|
One megafire held to "merely" a large fire pays for the entire statewide build, thirteen times over.
| Scenario | Coverage | Downshift rate | Effective | Loss avoided / yr | Acres saved / yr | Return on cost |
|---|
Break-even is the effective rate at which avoided loss exactly equals the annual whole-life cost of the programme. It sits at 0.49%, which is about 36 of California's roughly 7,240 annual ignitions held one size class smaller than they would otherwise have reached. Put the other way: a single megafire demoted to a large fire funds 47 years of the programme on its own.
A deliberately pessimistic case still pays
Assume the system reaches only 15% of the state's ignitions and downshifts only 25% of those, so it works one time in twenty-seven. It still avoids $375M a year against an annual whole-life cost of $49M, a return of 8:1. The programme has to fall below an effective rate of 0.49% before it stops paying for itself at all.
The comparison that matters to a legislator
The full statewide build is 1.3% of CAL FIRE's annual budget, and its total construction cost is 0.6% of the insured losses from the January 2025 Los Angeles fires alone.3
What would falsify this
If drones cannot reliably suppress a sub-acre fire in field conditions, the downshift rate collapses and the case rests on detection alone, which ALERTCalifornia already does more cheaply. That single question is what the PoC is designed to answer first.
What it costs to build and to run
A bottom-up estimate in 2026 dollars: bill of materials per node, a learning curve across the fleet, programme-level costs that do not scale, and a ten-year whole-life view including a mid-life refresh.
| Item | USD |
|---|
| Item | USD / yr |
|---|
| Line | Basis | USD |
|---|
Cumulative spend against cumulative avoided loss
Table view
| Phase | Years | Nodes | CAPEX |
|---|
Strongly carbon-negative, by a factor of thousands
The system runs on solar collected on the envelope. Its only meaningful emissions are embodied in the hardware. Against that sits every acre that does not burn.
Three nodes, two fuel regimes, twelve months
The PoC is not a demo. It is designed to produce the two numbers the entire business case rests on: the real detection latency and the real downshift rate, under conditions that generalise to the rest of the state.
Butte Unit: Sierra foothills
Two nodes 30 km apart, giving a stereo baseline. Mixed conifer and oak woodland; Camp Fire (2018) and Dixie Fire (2021) country. Tests geolocation accuracy and the long-duration watch in terrain with severe masking.
San Diego Unit: coastal chaparral
A completely different fuel and wind regime: chaparral under Santa Ana offshore events, where fires move fastest and the response window is shortest. If the concept works here it works anywhere in the state.
Run inside CAL FIRE, not beside it
Detection runs in shadow mode against the ALERTCalifornia baseline for the first two phases, so every alert is scored against an existing operational system rather than against a claim. Live suppression trials happen only on CAL FIRE controlled burns.
| Line | USD |
|---|
| Metric | Target | Today |
|---|
What could stop this
Listed worst-first, including the one that could kill the programme outright. A proposal that does not name its own failure modes is not worth reading.
The next step is a $6.95M question.
Twelve months, three nodes and two fuel regimes will tell you whether a drone can reliably kill a sub-acre fire in the field. Everything else on this page follows from that one measurement. We are looking for a CAL FIRE unit willing to host it.
pz@xfaang.com · Xfaang