combine
Solar energy × local intelligence

Intelligence
powered by
sunlight.

Solar-assisted compute for small LLM labs. We’re developing a practical way for independent builders and small teams to adapt open-weight models and run private AI on their own hardware.

Discuss a lab pilotExplore the concept

Building in stealth · Research & prototyping

Explore the concept

Compute with sunlight, at different scales.

Start with a small LLM lab, then explore the yard, a data center and the device inside.

1 / 4

Swipe or scroll sideways to compare. Use the arrows or labeled views on desktop.

Small LLM lab · initial focus
01 / Outdoor PV02 / Storage & inverter03 / Accelerator workstation

A separate hardware class for developers building their own models.

Illustrative architecture
Build your own, at a smaller scale

Your model.
Your workstation.
More sunlight.

Explore small language models, adapt an open-weight model to your domain, or run private local inference on a solar-assisted workstation.

Adapt selected parameters of a pretrained model to your data. Model size, precision and accelerator memory determine what fits.

Adjust the planning assumptions
Daily PV energy
Workstation session energy
Daily energy coverage ceiling

Workload buttons load illustrative power and duration presets, not measured benchmarks. Adjust them for your hardware. Daily PV energy can be shifted only with suitable storage; coverage is an energy ceiling, not uptime. Inverter limits, hourly weather and additional cooling loads still need design.

What would need to be validated?

This is the initial system concept using existing accelerators, adequate RAM/VRAM, local storage and energy scheduling. It is not a released chip, a training-speed benchmark, or a claim that small-scale hardware can train a frontier model. Training from scratch is limited to appropriately small models and datasets. Fine-tuning and inference are the nearer-term use cases.

The daily energy example is a budget, not a promise of uninterrupted runtime. Electrical protection, battery usable capacity, cooling, safety and site conditions require design. Hugging Face’s PEFT documentation explains parameter-efficient model adaptation.

3D site model · daylight

Drag with a mouse to rotate. Camera view shows the fixed viewpoint; swipe sideways to compare illustrations.

Illustration, not camera footage
A single fixed camera

See what is visible.
Flag what isn’t.

Full-size trailers share one row of bays with an open apron for vehicle movement. Try a passing truck to see the limits of the view.

A05 · Clear

Sample check · 09:42

Sample statuses use simple sightline checks in this model. They are not AI results or a claim of field accuracy. Staff verify before directing a move.

Scene dimensions & physical assumptions

Illustrative geometry: 16.15 × 2.60 m trailers, 4.10 m overall height; 3.75 × 18 m bays; a 26 m open apron. Trailer tires and landing legs meet the pavement. A fixed camera sits 8 m above ground, 28 m in front of the bays, beyond the apron, with a 47° vertical field of view in a 16:9 image. These are model inputs, not a proposed installation specification.

One directional sun produces consistent shadows. The panel is tilted toward it, with power storage on the pole. Its depicted size is illustrative and independent of the original 0.5 W energy example. Real solar, storage, wind loading, foundations, camera optics, and truck turning paths require site-specific engineering.

Fifteen sample sightlines per bay meet pavement for an empty bay or the roof for a parked trailer. If any are blocked by another modeled trailer, the bay is marked unknown. This conservative geometry example excludes glare, weather, lens distortion, image resolution, and recognition errors. Even “Clear view” leaves some empty bays hidden by neighboring trailers. Camera-placement reference: Axis Object Analytics guidance.

Commercial data center · expansion concept

Click the PV panels, racks, cooling plant or electrical equipment.

Interactive cutaway
Selected component

Server hall

A commercial-scale horizon

Compute at scale.
Design the whole site.

A future solar-assisted compute campus would combine accelerator racks with a utility connection, storage, redundant power paths, cooling and resilient network access.

Adjust the planning assumptions
Daily PV energy
Daily facility demand
Daily energy coverage ceiling

Coverage is a daily energy ceiling, not solar uptime. Firm power must still support nights and low-sun periods. Battery losses and hourly dispatch are not simulated.

Explore candidate solar regions ↓
Commercial design boundaries

Conceptual only: no operating facility, rated rack capacity, certification, customer reservation or commissioning claim. PUE is total facility power divided by IT power. Daily demand = IT load × PUE × 24 hours. Cooling, power-conversion and other facility overhead are represented by that PUE assumption.

Panel and rack counts illustrate arrangement and are not a construction specification. Site assessment must include hourly generation, grid capacity, land, permits, heat rejection, water use where relevant, fiber routes and economics. Batteries shift energy and incur losses; they do not create the daily shortfall. Reference: Australian Government data-center energy guidance.

Every useful check
has an energy cost.

Combine, A Circle+ CompanyConceptual illustrative model.
Solar conditions
EnergyCapture → inference → status
Solar energy collected
2.00 Wh/day
5 equivalent full-sun hours
Whole-device energy use
1.92 Wh/day
80 mW average · assumed
Runtime with no sun
≈ 25 hours
From full usable storage
Energy surplus+0.08 Wh/day available to replenish storage.
Panel: 0.5 W peakCollection factor: 80%Usable storage: 2 Wh
Concept assumptions · hardware varies by workload · loads include sensing, compute and radio · peak power and hourly storage behavior not modeled.

The illustrative 0.5 W panel and assumed 80 mW image workload leave only 0.08 Wh/day of clear-day margin and a cloudy-day deficit. Measurements will determine capture interval, panel size, and storage. The atlas describes solar resource, not this device’s electrical output.

Explore the energy-model assumptions ↓.

Founded by Brian3× startup founderTechstars SF alumnus
Meet the founder ↓
01 / The first product

Your own models.
Start with one workstation.

For independent AI builders, research teams and small LLM labs. The first proposed system pairs an existing accelerator workstation with appropriately sized solar power, storage and software that schedules flexible workloads around available energy.

01

Choose a useful workload

Adapt an open-weight model to your own data or run private local inference. Match model size and precision to available accelerator memory.

02

Size the whole system

Use existing compute chips with a suitable solar array, inverter, usable storage and cooling. Grid support covers shortfalls when required.

03

Schedule and measure

Run flexible jobs during productive solar hours. Track task quality, completion time and total energy per successful run against a conventional workstation.

Initial scope: local inference and parameter-efficient fine-tuning. Training from scratch is limited to appropriately small models and datasets. A workstation needs its own power system; the tiny edge-device illustration is a separate application.

02 / A second use case: yard operations

Visibility has a cost.
Measure the whole job.

Beyond the lab, a proposed solar-powered camera can check marked yard bays where installing power is inconvenient. This separate example shows how to test an operational use case against its full cost.

Compare today’s checking effort with a complete deployment: camera, panel, storage, mount, gateway, connectivity, cleaning, and support.

A scoped paid pilot would test the value. If validated, the proposed model is hardware plus an ongoing service for bay status, device health, and model updates.

What to compare with powered cameras

Get a site-specific quote for camera coverage, trenching or cabling, power, connectivity, and support. Compare equivalent coverage and freshness; continuous video and periodic snapshots serve different needs.

What does manual checking cost?

Edit the example inputs. USD-equivalent labor cost only.

$416Monthly checking labor

20.8 staff hours each month.

Baseline expenditure, not promised savings. Some manual checks remain necessary. No Combine pricing or payback is assumed; inputs stay in this page.

03 / Progress & proof

Build. Measure.
Earn the next step.

Research and prototyping are ongoing. Public evidence today is an interactive concept and an assumption-based energy model. Measured hardware results and confirmed deployments are still to come.

Available now

A testable product direction

A small-lab buyer profile, a local model-development workflow, interactive system concepts and transparent energy budgets.

Next evidence

One instrumented prototype

Whole-workstation power measurements, repeatable model tasks, a sized solar/storage design and a supervised small-lab pilot.

The proposed 12-week validation plan
Weeks 01–04

Define & profile

Target 10 small-lab interviews. Choose a model and dataset, agree output quality and memory needs, and measure workstation power on a conventional supply.

Weeks 05–08

Integrate & compare

Integrate a suitably sized power system. Compare the same inference or adaptation task with and without energy scheduling. Test checkpoints, interrupted power and thermal limits.

Weeks 09–12

Pilot & decide

Target a four-week supervised lab pilot. Compare completed jobs, quality, energy use, operating effort and total cost. Timing depends on parts, capacity and site access.

Inside the chip: energy-model assumptions

The original interactive illustration is available in the device view above ↑.

Energy-aware scheduling is proposed; this illustration uses fixed whole-device average loads: 5 mW for low-duty sensing, 80 mW for periodic images, and 2 W for video. No capture interval or measured performance is implied.

Collected energy = panel rating × equivalent full-sun hours × 80% collection factor. Daily use = average power × 24 hours. No-sun runtime = 2 Wh usable storage ÷ average power. Gateway energy, hourly charging, peak-current limits, aging, and thermal effects are excluded. Daytime capture and nighttime sleep need separate profiling.

04 / A wider horizon

Start with useful hardware.
Build toward a platform.

The ambition is reusable solar-powered compute hardware, energy scheduling software, and interfaces that make physical-world signals useful to larger AI systems. A useful small-lab system is the first proposed test of that foundation; yard sensing is a separate application.

First / Small LLM labsThen / Yard sensing & integrationsResearch / Commercial compute & orbital research
Long-term technology partners and customers

Combine intends to serve frontier AI developers such as OpenAI and Anthropic, alongside chip makers such as NVIDIA. These are prospective customers and collaborators; no current relationship is claimed. Evaluation kits, firmware licensing, or custom silicon would follow technical evidence and demand. Larger systems need their own power, cooling, memory, and connectivity design.

05 / Global solar atlas · Research horizon

Where could sunlight
support more compute?

Explore nine candidate regions for future solar-powered compute installations, plus orbital concepts. This is an early resource screen for a broader ambition; the initial focus is a solar-assisted workstation for small LLM labs.

NASA POWER · Satellite-informed · 2001–2020
Lowest monthly mean:< 44 to < 5≥ 5 kWh/m²/day

Select a point or choose a region. Regional samples, not proposed land parcels.

—Annual mean
—Lowest monthly mean

NASA data for this point

Strong solar resource is one screening factor. Every ground location still has night, cloud risk, and seasonal variation. Storage or backup, cooling, fiber, land access, grid connection, and project economics determine feasibility.

Data, method & limits of this map

Ground figures are global horizontal irradiation: solar energy arriving on a horizontal surface, in kWh/m²/day. NASA POWER’s ALLSKY_SFC_SW_DWN climatology provides 2001–2020 monthly and annual means from satellite-derived SYN1DEG data. The lowest value is the minimum of the 12 monthly means, not the worst individual day. Values do not measure PV electricity, sunshine duration, or guaranteed uptime.

Nine samples illustrate sunny regions across several continents; this is not an exhaustive global ranking or validated site recommendation. The roughly 1° solar grid cannot establish parcel-level shade, land suitability, or local extremes. Hourly data, engineering, and local diligence are required. Retrieved September 16, 2026. No live satellite feed or Combine satellite mission is represented.

Sources: NASA POWER climatology · NASA data resolution · Natural Earth geography. Rendered with D3 and world-atlas. Orbital concepts are research illustrations, not deployment commitments.

Brian, founder of Combine
06 / Founder

Brian

3× startup founder. Techstars SF alumnus.

Brian is building Combine in stealth as a Circle+ company, with research and prototyping at the intersection of solar energy, AI, and hardware. The immediate goal is a measured prototype and a useful first pilot with a small LLM lab.

More About Me
Kenyan flagAmerican flag
TechstarsMicrosoft for StartupsAWS Startups
Pitch DeckBusiness Model

Private investor materials. Contact Brian for access.

For small LLM labs & independent AI builders

Bring one model.
Let’s test its potential.

Explore a scoped pilot for model adaptation or private inference. Test whether solar-assisted operation suits your workloads, site and budget.

Discuss a lab pilot

Opens Brian’s LinkedIn. Pilot partners and deployments are not yet confirmed.

Hardware reference examples

TI energy-harvesting power management and ST’s camera/AI reference illustrate component categories, not selected parts or supplier relationships. Larger processors need a larger power system; NVIDIA’s Jetson specifications provide a separate reference. Existing Axis logistics camera systems are alternatives to evaluate.