Solar From First Principles: From Car-Roof Panels to Intermittent Compute as Buyer of Last Resort
Solar already beats farming per square metre, and cells are already cheap. What actually stops us covering the world is deployment cost and the fact that electricity can't be boxed up and shipped like tomatoes. Breaking the cost down per square metre, then looking at one path out.
The thesis in one line: solar output per square metre already crushes agriculture, and cells are already cheap. What blocks “covering the world” is (1) one-time deployment cost — the ~10x of headroom sits in installation and soft costs, not in the cell — and (2) the fact that electricity, unlike a tomato, cannot be boxed up, shipped, or stored. The most promising way out: convert surplus or stranded power on-site into an interruptible, fungible, low-bandwidth digital commodity (bitcoin today, asynchronous AI compute at the frontier), backhaul over satellite, and bypass the grid entirely.
1. Starting point: solar on a Model 3 roof
- About 1.5 m² of curved roof, usually parked in shade or a garage → realistically 0.4–0.7 kWh/day ≈ 3–7 km of range per day, and near zero in a Halifax winter.
- That’s about $30/year of electricity. Payback runs into the high teens or twenties of years. It is not a range solution.
- Where it does make sense: off-grid camping, and offsetting Sentry mode or cabin-overheat drain. Use a portable folding panel plus a power station — you can aim it, it doesn’t damage the car, and it’s more efficient than a fixed wrap.
2. First principles: cost and payback
- Cells and modules already bottomed out. Chinese FOB TOPCon runs about $0.087/W; the bare cell is roughly 3% of system cost. “Panels are expensive” is an out-of-date diagnosis.
- The payback equation:
years = cost per m² ÷ (capacity factor × electricity price × 8.76) - The revenue side has a physical ceiling. Annual revenue per square metre is capped by the solar constant at roughly $60/m²·yr. Efficiency can give you at most 2x, electricity price 2–3x, incident energy 0x. You will never reach one-year payback by pushing revenue alone.
- Efficiency is not the payback bottleneck — it’s a second-order variable. The real value of perovskite tandems (LONGi’s 34.85% lab cell, above the single-junction theoretical limit; ~43% theoretical for a dual junction) is the low-temperature printable manufacturing process, not the record itself.
- Conclusion: one-year payback requires roughly 10x cost reduction, and it cannot come from the cell (3% of cost). It has to come from deployment.
3. Cost per square metre (NREL residential, ~$2.70/W ≈ $540/m²)
| Category | Item | $/m² | Share |
|---|---|---|---|
| Module | Cell + encapsulation (delivered) | $70 | 13% |
| Inverter | Incl. microinverters/optimizers | $64 | 12% |
| Structural BOS | Racking/rails | $22 | 4% |
| Electrical BOS | Wiring/disconnects/meter | $50 | 9% |
| Installation labor | On-site hours | $66 | 12% |
| Permitting/interconnect (PII) | Municipal permits, inspection, interconnection | $20 | 4% |
| Sales & customer acquisition | Field sales, advertising, quoting | $84 | 16% |
| Overhead | G&A/insurance | $66 | 12% |
| Net profit | Installer margin | $68 | 12% |
| Sales tax | — | $16 | 3% |
| Total | ~$540 | 100% |
What this shows
- The part that actually generates electricity is under 13% of cost (about 3% as a raw commodity).
- Roughly 69% is “getting it up there, wiring it in, and selling the job” — deployment and soft costs, none of which have anything to do with solar physics. The 16% customer-acquisition line is a monster unique to residential.
- The same cells at utility scale cost $1.06/W ≈ $212/m², 40% of residential. That’s the proof: the bottleneck is scale and standardization, not the cell.
- The 10x path is to drive deployment cost toward material cost: make installation a by-product of something you were going to do anyway (factory-integrated, paint-on, standard building materials), plus plug-and-play, plus no permit.
4. Growing electricity vs growing food (gross revenue, $/m²·yr)
| Category | Example | $/m²·yr |
|---|---|---|
| Row-crop staples | Corn/soy/wheat | $0.08–0.21 |
| Open-field cash crops | Cotton, common produce | ~$1–10 |
| Greenhouse vegetables | Tomatoes (10–30 kg/m²) | $30–90 |
| Solar (at retail self-consumption rates) | — | ~$60 |
| High-value CEA | Saffron, microgreens, cannabis | $400–several thousand |
- Solar beats row-crop agriculture by 300–750x per square metre and sits in the same band as greenhouse vegetables — with no seeds, fertilizer, harvest, spoilage, or sales.
- On a net basis the gap is wider still: staples require heavy reinvestment every year (total margins on 2024 corn, soy, wheat and cotton were negative), while solar has near-zero annual operating cost.
- So the bottleneck on “why not cover everything” is not revenue density. It is (a) up-front capex and the payback hurdle, and (b) a problem unique to electricity ↓
5. The problem unique to electricity
A tomato can be boxed, shipped, and sold whenever you like. A kWh generated at noon on an empty patch of land is worth almost nothing without transmission or storage. Crops come with storage and logistics built in — the crop is the physical commodity. Electricity doesn’t. That, not output density, is the real physical constraint on where you can put panels: not how much you produce, but whether it can be moved and held.
6. A model worth exploring, and its boundary conditions
The strongest version: off-grid or curtailed generation → consume it on-site with an interruptible, fungible, low-bandwidth load → backhaul by satellite → prefabricated roll-out material placed by machine (killing the labor line), rather than laid flat on the ground.
- Already happening. Bitcoin miners are the buyer of last resort for curtailed power. Engie has an 895 MW solar plant in Brazil; MARA acquired a 114 MW wind farm to mine behind the meter. One simulation puts solar-powered mining at 3.5-year payback versus 8.1 years selling to the grid.
- Bitcoin is a near-perfect fit: interruptible at zero cost, globally fungible output, tiny data footprint (satellite is already enough), no latency sensitivity.
- The core tension: cheap but intermittent power (~20% capacity factor) against expensive hardware that wants to run 24/7 means ASICs idle 80% of the time and payback stretches out. The mitigating trend: ASIC prices fell from $80/TH (2022) to $16/TH (2025), and the cheaper the hardware, the more intermittency it can absorb.
- AI is a poor fit as usually conceived: training wants 24/7, the chips are extremely expensive, and you have to move the data. The only version with a shot is batch, asynchronous, preemptible inference.
- What it’s actually worth: converting power to compute does not double the $60/m². It turns “curtailed = $0” into a positive number. You’re capturing a price spread — monetizing, not amplifying.
Correcting point 1 (laying panels flat)
- Racking is only 4% of cost, so “eliminate the racking” attacks the smallest target on the board. The real weight is labor plus electrical BOS plus soft costs.
- Flat-laid panels contradict the low-maintenance goal anyway: −15–20% from heat derating, dust that never self-cleans (and remote means dusty), and poor lifespan on flexible films.
- The better answer: prefabricated rolls placed rapidly by machine, keeping a minimum tilt and ground clearance — preserving both output and self-cleaning.
Three first-principles conditions for the model to hold
- Power must be near-free or curtailed — you’re earning the spread and being compensated for low uptime. That locks the model to remote or oversupplied places where the power would otherwise be worthless.
- Hardware capex must be cheap enough to absorb the idle penalty of a ~20% capacity factor.
- The load must be interruptible at zero cost.
The honest ceiling: difficulty and competition grind global miner profit toward breakeven, with electricity price as the only moat; ASIC economic life is about 1.29 years, with the e-waste that implies; and the opportunity-cost critique is strongest exactly where the grid is tight (Norway restricted mining in 2025) and weakest where power is stranded or in surplus.
7. The position that’s actually open
The opening isn’t building one more solar-powered bitcoin mine. It’s turning AI’s interruptible workloads into the next buyer of last resort for intermittent power — making a class of economically valuable AI computation behave like mining: run when the sun is out, stop when it isn’t, at zero loss.
Bitcoin is today’s answer because it satisfies three demanding conditions at once: interruptible at zero cost, fungible and globally saleable, and almost free of bandwidth requirements. Batch, asynchronous, preemptible AI inference is starting to satisfy the first two. Whether it satisfies the third depends on how much data the particular job has to move.
If that path opens, the model widens from bitcoin’s narrow niche to the whole pool of AI compute demand — and that’s the point at which “growing electricity” might actually cover the world.
Personal research notes. Cost and price figures come from NREL, public industry reporting, and company disclosures; they move quickly, so check the date before citing. Not investment advice.