At a glance
- 69%have no direct GPU accessAccess is the bottleneck. The median wait for a compute job is 14.7 hours.
- 72%delayed a project in 12 monthsDelay has a cost: 793 delays and 195 abandoned projects in all.
- 41%run on the grid alonePower is part of the problem: a median 3 hours lost a week, and 30% report heat limits.
- 48%need local data controlControl matters as much as price: 43% have hit a provider limit.
- 131can host a pilotDemand is real and specific: 72% want solar with battery or grid backup.
In the validated model, a 2 kWp array with 5 kWh of storage covers about 97% of a small lab's annual load at 5.0 kWh/m²/day.
01
Executive summary
Small AI labs outside the big compute hubs are blocked less by talent than by access to GPUs, reliable power and control over their data. Combine closes that gap with solar-assisted compute: an accelerator workstation, a solar array and storage, sized and scheduled together.
This paper draws on three bodies of evidence gathered over eight months, from 1 February to 1 October 2026: a survey of 600 AI practitioners in 42 countries, 60 structured interviews, and an open energy model validated against pvlib and the NREL System Advisor Model.
The engineering case holds up. In the validated model, a 2 kWp array with 5 kWh of storage covers about 97% of a small lab's annual load at 5.0 kWh/m²/day, and sun-aware scheduling cuts grid import further. Combine's next step is to measure real workstation power on the workloads respondents run most: LoRA and QLoRA fine-tuning, and local inference.
02
The compute gap
Two in three respondents have no GPU of their own, and the cost shows up as delayed and abandoned research rather than as a line on an invoice.
The sample is the people Combine is built for. Independent AI builders are the largest group (22%), followed by research scientists (17%), graduate researchers (17%) and university faculty (14%). Their work is adaptation, not frontier training: LoRA and QLoRA fine-tuning (39% together), local inference (19%) and retrieval with embeddings (13%). Most work with open-weight models from the Llama, Mistral, Gemma and Qwen families.
Source: Combine survey of 600 AI practitioners in 42 countries, 1 February to 1 October 2026
Only 9% (52) have a dedicated local GPU; 23% share a limited one. The median respondent uses about 20 compute hours a month and waits 14.7 hours for a job to start.
| Region | Respondents | No direct GPU | Delayed a project | Median wait (h) | Billing or geographic barrier |
|---|---|---|---|---|---|
| Africa | 240 | 78% | 73% | 15.8 | 16% |
| Oceania | 30 | 80% | 67% | 15.6 | 23% |
| South America | 90 | 70% | 72% | 14.1 | 24% |
| North America | 45 | 64% | 84% | 14.3 | 9% |
| Europe | 45 | 58% | 67% | 13.3 | 24% |
| Asia | 150 | 54% | 68% | 13.5 | 22% |
Note: Oceania has 30 respondents, so its shares move more with each answer.
Source: Combine survey of 600 AI practitioners in 42 countries, 1 February to 1 October 2026
Africa has both the largest share of respondents and the highest rate of no GPU access among large groups. North America's high delay rate is a reminder that queues and quotas, not only geography, hold work back.
03
Power and cost
Most respondents have grid power most of the day, but not all of it, and they pay a high price for each kilowatt-hour they report.
Source: Combine survey of 600 AI practitioners in 42 countries, 1 February to 1 October 2026
The median respondent has grid power 22 hours a day (middle half: 20.3 to 23.1) and loses 3 hours a week to outages (middle half: 1.5 to 5). Oceania (4.6 hours a week) and Africa (3.3) report the most. 30% report ambient heat that constrains their hardware, and 64% describe their internet as variable (45%) or unreliable (19%). Both point to a machine that keeps working offline and is designed for heat.
04
Privacy, policy and access
Price is the most named single barrier, yet the barriers Combine removes by putting compute on site are together named first by more than half of respondents.
Source: Combine survey of 600 AI practitioners in 42 countries, 1 February to 1 October 2026
A cloud discount fixes none of these. A machine on the lab's own site, running open-weight models, fixes all three.
05
Demand for solar-assisted compute
Respondents want solar compute with a safety net: 72% choose a solar setup backed by a battery or the grid, and 71% say grid backup is a requirement.
Preferred energy setup: solar, battery and grid backup (43%), solar-first with grid backup (29%), grid-first with a solar offset (20%), not sure (9%). Interest in a solar-assisted workstation has a median of 52 out of 100; 57% score it 50 or more and 13% score it 70 or more. Interest is broad but considered, which is what a buyer of lab equipment should sound like.
Source: Combine survey of 600 AI practitioners in 42 countries, 1 February to 1 October 2026
No single way to buy or deploy dominates. Pay per job (25%), buying an appliance (20%), an institutional contract (19%), leasing (19%) and a shared subscription (18%) all draw similar support, and deployment preferences split almost evenly between a personal workstation, a shared lab node, a regional hub and a mix. Combine should therefore be one machine with several ways to pay for it. The features people want most are the ones the architecture already implies: private inference (19%), offline-first operation (17%), local fine-tuning of open models (17%) and energy-aware scheduling (17%).
Source: Combine survey of 600 AI practitioners in 42 countries, 1 February to 1 October 2026
06
What the interviews add
The 60 interviews turned the survey's numbers into six product requirements, led by offline operation and local data control.
Interviewees came from 30 countries, 22 of them in Africa and 19 in Asia. Most were independent builders (16), graduate researchers (12) and startup founders (9). Their work centred on LoRA fine-tuning (21) and QLoRA fine-tuning (12). Each interview focused on one area: compute access and cost (19), silicon and technical scope (11), privacy and control (11), power and solar fit (10), and product and adoption (9).
Note: Each interview is coded to the one implication it pointed to most strongly.
Source: Combine structured interviews, 60 participants in 30 countries, 1 February to 1 October 2026
Three of these confirm what the survey shows: offline-first operation, grid backup and shared deployment. Two shape the engineering roadmap directly: size to the workload, and measure power before committing to silicon. The interviews also set what each pilot must bring: GPU model, memory and training configuration; an hourly power trace and outage log; cloud invoices or credit usage where shared; the procurement decision-maker; and the site's solar resource.
07
Engineering evidence
A small node is enough: in the validated model, 2 kWp of solar with 5 kWh of storage covers about 97% of a small lab's annual load at 5.0 kWh/m²/day.
Combine's energy model simulates one node hour by hour for a year: a solar array, storage, an inverter and one accelerator workstation running a 4-hour fine-tuning session and 8 hours of inference a day (2,442 kWh a year). It covers 54 scenarios across three sunlight levels, three array sizes, three storage sizes and two scheduling strategies. On 2 November 2026 (Gate G1) it was checked against two industry-standard solar models on real hourly weather.
| Reference site | pvlib (kWh/kWp a year) | NREL SAM | Agreement | Heat loss | Daytime cell temperature |
|---|---|---|---|---|---|
| Douala | 1,362 | 1,329 | +2.5% | 7.7% | 40 °C |
| Accra | 1,635 | 1,633 | +0.2% | 6.8% | 39 °C |
| Addis Ababa | 1,613 | 1,583 | +1.9% | 5.3% | 34 °C |
Note: 1 kWp fixed open rack, 10° tilt, 14% system losses, 96% inverter efficiency.
Source: Combine Gate G1 cross-check, 2 November 2026; PVGIS SARAH2 typical meteorological years
- Sizing. 2 kWp with 5 kWh covers about 97% of the load with real seasons. The remaining gap falls in the rainy months; a larger array mostly adds surplus.
- Storage. Going from 5 to 15 kWh cuts grid import from 75 to 29 kWh a year at the Accra reference.
- Scheduling. In the tightest case, moving flexible sessions into sunnier hours cuts grid import from 286 to 107 kWh a year, in exchange for 45 sessions a year moved to a later day.
- Weather data. Two satellite datasets differ by 7 to 20% on annual sunlight at the same site, so every sizing is tested against both.
The same method scales. A 1 MW data center hall at a PUE of 1.3 needs 31.2 MWh a day; a 5 MWp field covers about 60% of it at 5 sun hours, with firm power and storage carrying nights. The model, data and code are public in the Combine repository.
08
Design principles and roadmap
The evidence asks Combine to build one machine, offline-first and backed by the grid, and to prove it with measured power before any custom silicon.
- Offline-first, local data. Private inference and fine-tuning work with no internet, with checkpoint recovery through power interruptions.
- Solar-first, never solar-only. Battery and grid backup are standard, and every report shows solar fraction beside any deferred jobs.
- Sized to the workload. Each node is sized from the lab's real workload and sunlight, tested against both NASA POWER and PVGIS.
- Built for heat. Enclosure and cooling are designed for sites where cells average 34 to 40 °C in daylight.
- One machine, several ways to pay. Purchase, lease, institutional contract, pay per job or a shared subscription, with transparent usage metering for shared lab nodes.
- Measure before silicon. Whole-workstation power, memory and heat are measured on real workloads before the energy management chip moves beyond simulation.
November to December 2026: power and memory traces for LoRA, QLoRA and inference on the open-weight models respondents use most.
Hourly solar from pvlib, monthly seasons and both weather datasets, replaying the scheduler on the measured traces.
Start with the 131 respondents who can host a pilot, with a documented baseline and an agreed workload at each site.
A written feasibility note on the energy management block, decided on measured data.
09
Methods and limitations
Data collection. Questionnaires and structured interviews ran from 1 February to 1 October 2026. Participants were recruited through online invitations sent by email, against regional quotas for the survey (Africa 40%, Asia 25%, South America 15%, North America and Europe 7.5% each, Oceania 5%), and consent was obtained from survey respondents and all 60 interviewees. Interview notes and transcripts are Combine proprietary data and are not published; this paper reports only the coded themes and product implications drawn from them.
How figures were computed. Shares are of all 600 survey respondents unless a group is named. "Median" is the middle value; "middle half" is the range between the 25th and 75th percentiles. Interview counts are records coded to each focus area and product implication. Energy model figures come from the published Gate G1 cross-check.
Limitations. The sample is a quota sample, not a random sample of AI practitioners, so shares describe these respondents, not a population, and all-respondent totals reflect the chosen regional mix. Grid hours, outage hours, wait times and tariffs are respondents' own estimates. Interest and pilot readiness are stated intentions, not commitments. Workstation power in the energy model is an assumption until the measurement stage reports.