One listed 8-GPU B300 system. The model weights themselves are not the capital cost.
Source observation ↗A physical path to intelligence sovereignty
Sunlight in.
Intelligence out.
This page asks a concrete version of the Household Sovereignty Index question: what would it take for one household-controlled site to generate, store and convert enough energy to operate frontier open intelligence without renting the model or the electricity that keeps it alive?
The installation below is an illustrative strong-form scenario, not a canonical HSI household. Every market price remains dated and classified; every sizing assumption remains visible.
- 01Harvest sunlight256 kW DC
465 modules in the modeled field.
- 02Store through dark1.5 MWh LFP
Roughly 68 hours at the design load.
- 03Run local compute8 × B300
One household-controlled Kimi K3 inference target.
- 04Serve intelligenceLocal by design
No rented model runtime or utility connection in this scenario.
sunlight stored electricity sovereign intelligence
Current modeled build
What would this stack cost today?
Market watchSep 12, 2026
Modeled acquisition + energy-plant range$1,069,736–$1,304,736
Not a quote. Land, freight, taxes, financing, permitting surprises and contingency are excluded.
15 pallets × 31 panels × 550 W. Modules only.
Source observation ↗1.5 MWh scenario scaled from a 1 MWh commercial LFP benchmark.
Benchmark ↗PCS, switchgear, racking, balance of system, cooling and electrical/civil allowance.
Design target
One local frontier model.
Designed without the grid.
This reference architecture assumes no utility connection and sizes the energy plant around a 20 kW continuous facility load. The modeled storage covers roughly 2.8 days of dark runtime, but that is not a site-specific reliability guarantee: strong-form temporal sovereignty still requires hourly weather, load and reserve simulation.
One single-node Kimi K3 inference target.
480 kWh of modeled energy demand per day.
465 × 550 W modules in the current scenario.
About 68 hours at the design load after the modeled usable-energy reserve.
Active model: Kimi K3. The daily watch surveys 4 open-weight families — moonshotai/Kimi-K, deepseek-ai/DeepSeek-V, deepseek-ai/DeepSeek-R, zai-org/GLM — and gates them identically; no vendor is privileged. It may advance the reference automatically only to a newer version inside the active family, and only when that release has a public official repository with a declared license and an exact vLLM recipe that still fits this same eight-GPU B300 profile. A successor that needs different hardware is held behind the reference-profile gate rather than silently changing the cost model. An eligible model from a different family is recorded as an alternative and never auto-promoted, because version numbers order only within a family and choosing between families is a Sovereign Compute Frontier measurement that does not exist yet. Last model check: Sep 12, 2026. Inspect the active serving recipe ↗
The cost frontier
The number should move.
Hardware, module and storage prices are not constants. This experimental tape keeps dated observations so a future visitor can see whether household-owned frontier intelligence is becoming cheaper or more expensive to assemble.
The interval is the modeled low–high range—not measured uncertainty and not a quoted price.
| Date | Compute | Solar modules | Storage scenario | Modeled total |
|---|---|---|---|---|
| Sep 12, 2026 | $507,236 | $67,500 | $345,000–$480,000 | $1,069,736–$1,304,736 |
| Sep 11, 2026 | $507,236 | $67,500 | $345,000–$480,000 | $1,069,736–$1,304,736 |
| Sep 10, 2026 | $507,236 | $67,500 | $345,000–$480,000 | $1,069,736–$1,304,736 |
| Sep 9, 2026 | $496,236 | $67,500 | $345,000–$480,000 | $1,058,736–$1,293,736 |
| Sep 8, 2026 | $496,236 | $67,500 | $345,000–$480,000 | $1,058,736–$1,293,736 |
| Sep 2, 2026 | $496,236 | $67,500 | $345,000–$480,000 | $1,058,736–$1,293,736 |
Latest complete observation: Sep 12, 2026. Failed source checks do not overwrite the last complete observation.
Research boundary
This is a market watch, not a hidden methodology change.
HSI already distinguishes Observed, Measured, Derived, Proxy, Scenario and Unobserved evidence. This page keeps that discipline. Retail listings are proxies; scaled storage and system sizing are derived scenarios; neither silently becomes a canonical national HSI constituent.
The canonical energy research still follows the separate SEI-A and SEI-T framework: annual energy sufficiency and temporal delivery are different questions.