Legit, verified sources of free compute an African ML researcher can get in hours or days — no committee, no year-long cycle. Verified July 26, 2026. Free tiers change monthly; re-check quarterly.

The strategy in one line: stack Kaggle + Colab for ~50–60 free GPU hours a week, add Modal’s $30/month recurring credits for modern GPUs, use the free inference APIs instead of GPUs wherever possible, and graduate to TPU Research Cloud for anything heavy. None of this requires an institution; most of it doesn’t require a credit card. Each entry below can be expanded into a standalone one-page guide for grantees — that’s the “raise the bottom” publishing play.

Tier 1 — Instant. Working within the hour.

1. Kaggle Notebooks — best raw value, zero paperwork

What you get: Choice of P100 (16 GB) or dual T4 GPUs, ~30 GPU hours/week, plus ~20 TPU hours/week; 20 GB persistent storage; sessions up to ~12 hours.

Requires: Free account + phone (SMS) verification. No card, no institution.  Time to access: Instant.

Gotchas: The killer feature for unreliable power: “Save & Run All” executes in the background, so a 9–12 hour training run finishes even if your connection dies. Use fp16 (T4 has no bfloat16). ML training is explicitly allowed and documented.

kaggle.com/docs/efficient-gpu-usage

2. Google Colab (free tier)

What you get: NVIDIA T4 (16 GB), roughly 15–30 GPU hours/week (quota is dynamic and unpublished), sessions up to ~12 hours.

Requires: Google account only. No card, no phone.  Time to access: Instant.

Gotchas: Disk is wiped on disconnect — checkpoint to Google Drive or Hugging Face Hub every epoch. GPU not guaranteed at peak hours. Training allowed; banned uses are mining, remote desktop, file hosting, multi-account quota evasion.

colab.research.google.com

3. Modal — the sleeper pick

What you get: $30 in real credits every month, recurring, on modern GPUs (T4 up to H100/B300 bursts): ~50 T4-hours or ~7–8 A100-hours monthly.

Requires: GitHub-account signup, no card.  Time to access: Instant.

Gotchas: Code-first serverless (no notebook UI) — a learning curve, but jobs run detached and survive disconnects, which is exactly right for intermittent connectivity. Training and fine-tuning fully allowed.

modal.com/pricing

4. Lightning AI (free tier)

What you get: ~15 credits/month (≈20–25 T4 GPU hours) in a persistent browser VS Code/Jupyter studio.

Requires: Account; a verification queue and phone check have been reported. No card.  Time to access: Instant to a day.

Gotchas: Credits burn much faster on bigger GPUs; stick to T4-class. Persistent environment is the draw.

lightning.ai

5. GitHub Codespaces (CPU) + Cloudflare Workers AI (edge inference)

What you get: Codespaces: 120 free core-hours/month — a reliable cloud dev machine for data prep, tokenizers, evaluation. Cloudflare: ~10,000 free neurons/day running open models (Llama-class, Whisper, embeddings) at the edge.

Requires: GitHub / Cloudflare accounts. No cards.  Time to access: Instant.

Gotchas: Neither gives you a training GPU; both cover everything around the GPU. Saturn Cloud (T4-class, recurring free hours) is a fine backup notebook platform.

docs.github.com · developers.cloudflare.com/workers-ai

Tier 2 — Days. One short form.

6. Google TPU Research Cloud (TRC) — the biggest prize on this list

What you get: Free access to fleets of Cloud TPUs (v2/v3/v4-class, on-demand + preemptible), initial ~30 days, routinely renewable. This is the documented heavy-lift workhorse behind much of Masakhane’s actual training.

Requires: A short application describing the project; no institution needed — independent researchers get accepted; you agree to share results openly. Turnaround commonly 1–7 days.  Time to access: Days.

Gotchas: The one real barrier: TPUs are free but the small VM + storage bill runs on your own GCP billing account, which needs a card. Legit fix: the $300/90-day GCP trial (card for identity only, never auto-charged) covers those pennies. This TRC + trial pairing is the highest-value legit stack available to anyone, anywhere.

sites.research.google/trc

7. Hugging Face community GPU grants

What you get: Free GPU upgrade for a public demo Space supporting research or community work — many 2025–26 grants visible on the Hub.

Requires: A working Space + a short justification (open a discussion titled “Apply for community grant” or use the Space settings button). No card, no institution.  Time to access: Days, discretionary.

Gotchas: For hosting demos, not training. Free accounts also get 5 min/day of ZeroGPU use and unlimited free CPU Spaces.

huggingface.co/docs/hub/spaces-zerogpu

8. Azure for Students + GitHub Student Developer Pack

What you get: Azure: $100/year in credits, renewable, explicitly NO credit card. GitHub Pack: unlocks Azure’s $100, DigitalOcean $200/12mo, GitHub Pro (180 Codespaces hours), JetBrains, and dozens more.

Requires: Verifiable enrollment at any institution Microsoft/GitHub can check — many African universities work. Verification: minutes to ~2 weeks.  Time to access: Days.

Gotchas: African student IDs are often rejected on the first pass — resubmit with an official, dated enrollment letter. If a grantee is enrolled anywhere, this jumps to top-3 value.

azure.microsoft.com/free/students · education.github.com/pack

9. Vendor academic credits: Modal ($10K), Lambda ($5K)

What you get: Modal for Academics: up to $10,000 in credits via an application form. Lambda Research: up to $5,000 in GPU-cloud credits.

Requires: Short applications; criteria and turnaround unpublished.  Time to access: Days to weeks.

Gotchas: These are also Dusoma partnership targets — one approved grant covers an entire grantee cohort’s compute tranches.

modal.com/academics · lambda.ai/research

The free inference stack — often you don’t need a GPU at all

For most African-NLP work — translation baselines, annotation, data augmentation, synthetic data, evaluation — free LLM APIs replace local GPUs entirely. All of these are card-free as of July 2026:

• Google Gemini API (AI Studio): the most generous free tier; Flash-class models, thousands of requests/day in practice. Note prompts may be used for training — don’t send sensitive data.
• Groq: ~30 requests/min, ~1,000/day on Llama-3.x, Gemma, Mixtral at extreme speed — best free throughput.
• GitHub Models: GPT-4-class access with just a GitHub account (~150–1,000 requests/day).
• Cerebras (~1M tokens/day), NVIDIA NIM (~1,000 req/day), Mistral Experiment tier (~1B tokens/month, opt-in to data training, phone verification).
• OpenRouter: ~28 free models at 50 requests/day — a one-time $10 top-up lifts that to 1,000/day for life, one of the few places $10 buys a durable capability.
• Cohere trial keys: modest limits but includes Aya — the model family built for African and multilingual work.

Stacked together, that is thousands of free frontier-class requests per day, per person, with no card.

Africa-specific access — the honest picture

• KENET (Kenya): virtual labs, the CHUI HPC cluster, and a GPU-as-a-Service cluster for researchers at Kenyan member institutions — the most concrete national offering found. Contact info@kenet.or.ke.
• CHPC (South Africa): free allocations exist but require a South Africa-based PI; the realistic route for non-SA researchers is collaboration with an SA lab. Weeks, not days.
• ilifu (Cape Town): astronomy/bioinformatics only. WACREN/UbuntuNet: connectivity and federated identity, not compute. Nigeria/Ghana/Ethiopia: no verifiable open-access national program found — a gap worth naming publicly.
• Zindi: officially points competitors to Colab; some competitions ship sponsor credits, but no standing free-GPU program was verifiable in 2026.

Card workarounds that are legit

• GCP $300 trial: a debit card works, for identity only — never auto-charged, closes itself at $300. The intended, legit path.
• Safaricom M-PESA GlobalPay virtual Visa: an official Safaricom product funded straight from M-Pesa (3.5% FX fee). Works for many cloud signups; AWS/GCP sometimes decline prepaid/virtual cards without notice — try, don’t rely.
• Licensed fintechs Eversend, Bitnob, IntaSend (Kenya/Nigeria) issue legit virtual dollar cards with the same caveat.

Crackdown watch — what’s shrinking

• Oracle halved its Always-Free Arm servers June 15, 2026 (still the best free always-on CPU server if you clear the card wall).
• AWS SageMaker Studio Lab reportedly closes to new users July 30, 2026 — don’t build guides on it.
• Hugging Face cut its free Inference API allowance to a $0.10/month taster. Colab Pro for Education closed to new signups; the free tier survives.

The takeaway for Dusoma: the free floor is real but eroding — which strengthens the case for grants that include cash and owned credit pools rather than relying on providers’ generosity persisting.

Prepared for the Dusoma Foundation, July 26, 2026. All entries verified against official pages or dated 2026 sources except where noted as reported; full source URLs in the research annex. Each numbered entry is designed to expand into a standalone one-page grantee guide.