Skip to content
The engine

The engine
behind Epoch Learn.

Eco Foundry routes, power-caps, carbon-shifts, and evaluates every job in the Free Compute Program. On-device models take the light work, energy-efficient eco servers the heavy lifting, and each request lands on the lowest-power option that still meets the moment. Every model runs wrapped in structured, evidence-based pedagogy — and this page is the dossier: every figure cited.

thesis: close the access gap · more learning per watt · structured AI, not a raw chatbot

01 · the gap02 · energy & cost03 · learning
eco-foundry · engine room · scroll drives the load wave
01 · The gap

The bottleneck is compute,
not capability.

Talent is everywhere; the machines are not. The people most likely to be left out of the next epoch are the ones already furthest from a GPU.

“The estimated growth impact in advanced economies could be more than double that in low-income countries.”

IMF Working Paper WP/25/76[3]
~5%

Africa's AI talent with the compute they need

Just 1% have on-premise GPUs; the next 4% rent cloud by the hour. Within-Zindi-network estimate across 11,000 data scientists.[1]

< 2%

of world data-center capacity is in Africa

Deployment without local training capacity: the “Compute South.”[2]

~540×

per-capita data-center electricity gap, US vs Africa

< 1 kWh per person in Africa (2024) against ~540 kWh in the US.[7]

10–30×

more expensive GPUs, relative to GDP per capita

The same card costs an order of magnitude more where incomes are lowest.[6]

40 / 15 / 3

notable AI models in 2024: US / China / France

~90% came from industry; training compute doubles roughly every 5 months.[4]

2.6 B

people still offline

27% internet penetration in low-income vs 93% in high-income countries.[9]

A

Compute North vs. Compute South

A handful of countries host the world’s training capacity; the rest are left to deploy what others build. Eco Foundry is built for the South of that map.[5]

B

Iterate in 30 minutes, or wait 6 days

A G7 startup can run a training iteration roughly every 30 minutes. An African peer may wait up to 6 days for the same loop. Frontier training runs cost $78M (GPT-4) to $191M (Gemini Ultra) in compute alone, rising ~2.4× a year since 2016.[1][8][4]

02 · Energy & cost

More learning per watt,
by design, not by hope.

Global data-center electricity is on track to roughly double, from 415 TWh in 2024 to about 945 TWh by 2030, slightly more than Japan uses today, with AI the main driver. Eco Foundry’s answer is three settled techniques, stacked.

PROJECTED: IEA BANDS
eco-foundry · scheduler
$route smallest-capable-model WebGPU(browser) ▸ edge ▸ cloud
$powercap gpu=250W→150W → energy −13.7% · time +6.8%
$shift carbon-aware schedule → emissions −5% (next-day) … −20% (weekend)
stack model + power + schedule → more completed projects per kWh
reason every watt not spent on overhead is a watt of access returned
15.3–75.8%

training-energy cut, Zeus

Jointly tuning batch size and GPU power limit, no accuracy loss (USENIX NSDI ’23).[10]

−13.7%

energy from a 250W→150W power cap

For +6.8% time, averaged over BERT/MLM on 2–400+ GPUs (MIT Lincoln Lab).[11]

22–33%

energy saved by power-capping, replicated

Krzywaniak et al.: 22–32.5% on RTX 6000, 24–33% on V100. EnvPipe: up to 28.4% on a 16-GPU pipeline with <1% slowdown.[12]

5–20%

emissions cut by carbon-aware time-shifting

~5% shifting to next-day, ~20% to the weekend, without breaching SLAs. Operationalized by Google’s Carbon-Intelligent Computing.[14]

$3.84 / hr

multi-provider average H100, on-demand

Up 18% YoY (June 2026). AWS P5 ~$6.88, GCP A3 ~$11.06, Azure NC H100 v5 ~$12.29.[17]

$1,300–3,000

one 15-week class, at on-demand prices

30 students at 2 A100-hrs/week. Eco Foundry’s job is to make that figure approach zero for the people who can least pay it.[17]

GPU-hours per dollar, owned fleet vs rented cloud

UNDP estimates $1M of owned GPU capacity supports ~50% more African researchers, at twice the daily GPU hours, for 5+ years, versus ~2 years for the same $1M in cloud credits.[6]

03 · Learning

Structured AI tutoring helps learning.
A raw chatbot can hurt it.

Hands-on, scaffolded instruction beats lecturing, and a well-designed AI tutor beats both. Hand the same students a raw, unguided chatbot instead, and the technology can make them worse. The difference is not the model, it is the design around it.

EXTERNAL RCTS
+0.47 SD

exam-score gain from active learning

Failure rates fall ~34% → ~22%; students are 1.95× more likely to fail under lecturing (225 STEM studies).[18]

d = 0.71

effect of project-based learning

Medium-to-large, across 12,585 students; a 2026 meta-analysis reports g = 1.11.[19]

≈ 2×

learning gain, well-designed AI tutor

A GPT-4 tutor built with explicit pedagogy beat in-class active learning, with higher engagement (4.1 vs 3.6) and motivation (3.4 vs 3.1); 83% rated the AI’s explanations at or above their instructor’s (Harvard physics RCT, n=194).[20]

What “structured” means here

Active learningProject-based workScaffolded promptingSafety guardrailsHuman review where it matters

AI & big data is the #1 fastest-growing skill, 2025–2030: 170M jobs created against 92M displaced (net +78M), with 39% of core skills changing by 2030. Meanwhile ~17 million US schoolchildren lack adequate home internet, and 1 in 3 Black, Latino, and Native American households lack high-speed access. The skill the labor market wants most is gated behind the infrastructure the least-served don’t have.[22][23]

Why this matters

The 100× number is a ceiling, not a dial.

UPPER BOUND

Done together, four levers (a better Model, an efficient Machine, datacenter-grade Mechanization, and a clean-grid Map) have been shown to cut ML training energy by up to 100× and carbon by up to 1000×. We cite this as the size of the prize, not as a knob we can turn on a single class. It is a best-case ceiling across all four levers at once. Eco Foundry pulls the two levers a learning platform actually controls (the smallest capable Model and power-capped Machines) and schedules against the cleanest Map it can reach. We report what those deliver, not the ceiling.[16]

04 · How we measure it

One number
we hold ourselves to.

Per-watt claims are easy to cherry-pick. So we pre-declare a single metric, the unit it is measured in, and the threshold at which we admit it isn’t working.

> 25%

target energy saved per completed project

Versus an un-optimized cloud baseline running the same coursework, measured per completed project, not per training run, so efficiency can’t be bought by quietly dropping student work.

Metric
Energy (kWh) per completed student project
Baseline
Same curriculum on un-optimized on-demand cloud
Target
> 25% reduction, sustained across a cohort
Reassess if
< 15%, the approach is not earning its complexity
Levers counted
Smallest-capable-model routing · power cap · carbon-aware schedule
Reported as
Measured, with bands, never the 100× ceiling
Sources

Sources &
honest caveats.

  1. [1] UNDP Digital, “Only five percent of Africa’s AI talent has the compute power it needs” (May 2024). Analysis of ~11,000 data scientists in the Zindi network.
  2. [2] UNDP, “Time for Africa to lead the global AI revolution” (June 2025).
  3. [3] IMF Working Paper WP/25/76, “The Global Impact of AI: Mind the Gap” (April 2025).
  4. [4] Stanford HAI, 2025 AI Index Report, p.46. Frontier training-cost figures via Epoch AI.
  5. [5] Lehdonvirta, Wú & Hawkins, “Compute North vs. Compute South,” AAAI/ACM AIES 2024.
  6. [6] UNDP / Africa Green Compute Coalition (2025).
  7. [7] IEA, “Energy and AI” (April 2025).
  8. [8] Cottier et al., Epoch AI, arXiv:2405.21015.
  9. [9] ITU, “Facts and Figures 2024.”
  10. [10] You et al., “Zeus: Understanding and Optimizing GPU Energy Consumption of DNN Training,” USENIX NSDI 2023.
  11. [11] McDonald et al., MIT Lincoln Laboratory, arXiv:2205.09646 (2022).
  12. [12] Krzywaniak et al., Springer LNCS (2022); EnvPipe, pipeline-parallel energy reduction.
  13. [13] EnvPipe: energy-aware pipeline scheduling (up to 28.4% with <1% slowdown).
  14. [14] Wiesner et al., “Let’s Wait Awhile,” arXiv:2110.13234.
  15. [15] Radovanović et al., “Carbon-Intelligent Computing” (Google), arXiv:2106.11750.
  16. [16] Patterson et al., “The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink,” IEEE Computer 55(7) (2022), arXiv:2204.05149.
  17. [17] GetDeploying, H100 cloud GPU pricing (accessed June 2026).
  18. [18] Freeman et al., PNAS (2014). Meta-analysis of 225 STEM studies.
  19. [19] Chen & Yang, Educational Research Review (2019), 12,585 students.
  20. [20] Kestin et al., Scientific Reports / Nature (June 2025). Harvard physics RCT, n=194.
  21. [21] Bastani et al., PNAS 122(26) (June 2025). ~1,000 Turkish high-school students.
  22. [22] World Economic Forum, “Future of Jobs Report 2025.”
  23. [23] FCC, “Bridging the Digital Divide”; All4Ed (2018 ACS).
See it run

The numbers above,
as a live console.

The Eco Foundry Console walks the same routing, power-capping, and carbon-aware scheduling decisions in real time. Telemetry is simulated and labelled illustrative.

SIMULATED
Bring Eco Foundry to a classroom

For schools, educators & research cohorts

For classrooms, clubs, and student cohorts. The public playground needs no account.