Teaching on borrowed compute · Greg Chism

What we have previously relied on can no longer be guaranteed

Our courses depend on access that providers can shorten, restrict, or withdraw.

Educational computing · Evidence checked October 8, 2026

Three supports beneath a course are disappearingGitHub Classroom is closed. New Colab Pro for Education sign-ups are closed. AWS's new Free Tier free plan lasts at most six months. The course continues above these changing supports. The course must keep runningEvery student. Every assignment. Every term. GitHub ClassroomCLOSED Colab Pro for EducationNO NEW SIGN-UPS AWS Free Tier12 → 6 MONTHS A full course needs a dependable foundation.

Classroom 50 is a workable successor; external stewardship and Codespaces quotas remain.

October 10, 2021 – October 10, 2026

The free layer keeps moving

Select a year, then a change.

Colab Pro for Education was closed to new users by September 21, 2026, per Greg Chism’s update; the exact effective date is unconfirmed. Studio Lab was observed closed by October 8. This is a selected history, not a count of all free services.

Free GitHub repos remain; hosting, private-team limits and classroom workflows differ.

2022 Heroku free plans end · 2023 Replit stops new education teams · 2024 Replit education closes · 2025 Glitch hosting ends; AWS free plan becomes ≤6 months · 2026 GitHub Classroom closes; Colab education offer and Studio Lab closed to new users when checked.

Current free options · UArizona InfoSci · 9 non-cyber degrees

Free still works within limits

✓ Good classroom work

↗ Move beyond free when…

Best-fit assessment, not a full-degree replacement. Free quotas do not reserve capacity. Use public / approved teaching data; export work.

Proposed solution · compute + repository continuity

Make course infrastructure a shared service

Compute allocations and repository stewardship support one teaching serviceACCESS supplies the planned CPU and GPU classroom allocation; NAIRR Classroom is an additional eligible AI-course route. A campus-owned repository program supplies live Git hosting, independent backups and restore tests. A local team manages course spaces, permissions, grading links and retention. Institutional support and storage must outlast individual awards. ACCESS CPU + GPUNAIRR Classroom: additional AI routeTest · benchmark · apply · verify · teach CAMPUS-OWNED REPOSITORIESGitHub Free / GitLab EducationManaged forge + independent backup option MANAGED TEACHING SERVICELabs + accounts + supportPrivate repos + course templatesGrading links + export toolingBackups + restore rehearsalsNamed owner + annual operating budget Supported labsPrivate submissionsCourse archives Keep Git history + LFS + course records; test recovery on another host before the next term.

No student credit card in this proposed service.Federal compute map ↗

Compute awards expire; institutional Git ownership, off-host backups and funded support must outlast them. A mirror alone is not a backup.

Illustrative first teaching cycle

Size the allocation.
Build the support.

Adjust the cohort to see demand.
These are planning estimates, not awarded capacity.

CPU: 15 weeks × 8 core-hours / student / week.
GPU: 6 weeks × 2 GPU-hours / student / week.
Both include a 25% planning allowance.

72,000

CPU core-hours

12 × 40 × 15 × 8 × 1.25

2,400

GPU-hours

4 × 40 × 6 × 2 × 1.25

480 course seats · 160 in AI pilots

Course seats may include the same student twice.
GPU type, simultaneous demand, memory, storage, and support must be benchmarked separately.

12–10 WEEKS BEFORE

Recruit & scope

Inventory labs, repos, owners, storage and retention needs.

10–8 WEEKS BEFORE

Benchmark & apply

Request compute and education licenses; fund Git support.

BEFORE THE FIRST LAB

Rehearse & onboard

Test class load, private access and a restore on another host.

DURING THE COURSE

Teach & support

Back up nightly; track use and failures; file midterm reports.

AFTER THE COURSE

Export & plan again

Export and archive; final report within 30 days; renew.

Proposed schedule, not review-time guidance. Convert hours to resource-specific credits; approval is conditional. Fund storage and staff separately.

The goal: compute and course history that remain usable through the next teaching cycle.

Repository continuity plan

Proposed repository program, not an existing campus service or completed migration. Start with two courses before scaling to all nine degree programs.

1 · Keep a viable free route

Use GitHub Free for ordinary Git hosting now. Assess an institution-sponsored GitLab for Education application for private class teams and additional features; annual renewal and use restrictions still apply. Classroom 50 is a workable assignment-management successor, externally maintained by the Fifty Foundation. Keep course repositories/configuration in an institution-controlled GitHub organization; verify Team/Enterprise education eligibility, retain exports and rehearse an alternate submission path. It adds no Codespaces compute allocation.

Instructor lead + college IT · before the next course

2 · Establish campus ownership

Inventory course repos, sizes, LFS, issues, wikis, artifacts and grading links. Give each course a responsible owner and successor. Keep submissions private; do not make student work public to escape a quota.

Course owner + college IT · weeks 12–10

3 · Pilot a managed fallback

Evaluate a supported campus Git service first. If none fits, pilot Forgejo or self-managed GitLab on an institution-managed host with authentication, private projects, quotas and isolated grading runners. Budget staff, durable storage and recovery separately from compute awards.

College IT + central IT · weeks 10–4

4 · Preserve more than Git

Take versioned nightly Git and LFS backups to independent storage; export issues, wikis, release files and course assessment links through supported paths. Define retention, access and deletion rules. A live mirror can copy deletions and cannot replace a separate backup.

Named service operator · nightly; alert on failures

5 · Prove recovery, then scale

Before teaching, restore a private sample course on a second host and confirm history, LFS, permissions and grading links. Proposed pilot targets: at most 24 hours of data loss and recovery within one business day. Validate these targets before promising them.

Operator + instructor · pre-term and after major changes

6 · Close the teaching cycle

Give students portable code exports, retain institutional assessment records under approved policy, and archive releasable research code/data with Libraries/ReDATA when appropriate. Review storage, restores, licenses and allocation renewal before the next cohort.

Instructor + IT + Libraries · each term

Pilot gate: proceed only after a tested export/restore, correct private permissions, a named operating owner, a storage budget and a plan for license or allocation expiry. ACCESS/NAIRR supply reviewed compute, not a permanent repository or records-retention guarantee.

GitHub Free ↗GitLab Education ↗Backup boundaries ↗UArizona archiving ↗

Free options · all 9 degrees

Codabench & classroom alternatives

Classroom 50 and Codespaces checked October 13; Modal and eight similar platforms checked October 16; earlier resource comparisons retain October 11 dates. Cloud credits and paid-service boundaries ↗. Free access supports particular assignments; it does not reserve a whole class or fund a semester service.

1 · DevelopCPU / notebook / GPU quota
2 · SubmitPredictions or packaged code
3 · EvaluateScoring queue + leaderboard

Codabench: free competition hosting and a shared CPU queue with a 20-minute job cap. Organizers supply their own queue for code submissions. GPU workers require separately supplied compute.

Free, open-source assignment workflow · evaluate before adoption

Classroom 50

Workable successor for GitHub Classroom assignment distribution, repositories, submissions, feedback and grading.

Boundary: Fifty Foundation project and hosted app are not ours. Requires a Team/Enterprise organization; check education eligibility. No additional Codespaces compute; Actions grading has separate quotas. ACCESS remains the CPU/GPU classroom route.

Classroom 50: project and license · Classroom 50: prerequisites and education eligibility · Classroom 50: managing Actions cost · Codespaces · free personal-account quotas

Advertised free for courses ≤50 students

CodeGrade

Programming labs with browser editing, immediate autograding and instructor feedback.

Boundary: Check free-plan execution/LMS limits and roster scope. No verified free GPU entitlement or campus deployment; grading records need approved retention.

CodeGrade · bounded free classroom autograding

Adjacent project; current free entitlement unverified

ChaGrade

Challenge-linked individual/team homework and instructor grading beyond leaderboard scores.

Boundary: Live page labels version 0.1 / Work In Progress. Support, quotas and current Codabench integration need verification; not a ready college-wide baseline.

ChaGrade · challenge-linked classroom workflow

Hosted service paid; open-source self-hosting

EvalAI

Custom prediction/code evaluation and organizer-controlled remote workers.

Boundary: Hosted plans $125–850/month; remote evaluation $100/month plus own compute. Self-hosting requires funded staff, infrastructure and storage. No verified hosted free tier.

EvalAI · paid hosted challenge plans · EvalAI · open-source self-hosting

Free accelerator notebooks; quota varies

Kaggle Notebooks

Checkpointed small-model GPU experiments and training for a competition.

Boundary: Check each account’s GPU quota, available hardware, session timer and saved disk. No reserved class seats or distributed/long-running training promise.

Kaggle · free accelerator notebooks · Kaggle · floating weekly GPU quota

Base free tier remains available

Free Google Colab

Short shared notebooks, CPU analysis and opportunistic accelerator exercises.

Boundary: GPU access and resource/session limits fluctuate. Free Colab is separate from the closed new-signup education Pro offer; never promise class GPU availability.

Google Colab Pro for Education · offer closed

Free ephemeral CPU sessions

mybinder.org

Small Python/R notebooks launched from a reproducible repository, including metadata/text workshops.

Boundary: 1–2 GB RAM; ten-minute idle shutdown; aims for ≤6 hours or one CPU-hour; 100 concurrent sessions per repo. Download work; no GPU or persistent workspace promise.

Binder · free ephemeral teaching notebooks

Five GPU minutes/day for signed-in Free users

Hugging Face ZeroGPU / Spaces

Short Gradio inference demos and model-behavior comparisons; static exhibits on free Static Spaces.

Boundary: Two own ZeroGPU Spaces only for eligible personal accounts ≥30 days old. Shared queues and Gradio constraints; organization hosting paid. Ordinary compute-backed Space creation also needs a paid plan.

Hugging Face ZeroGPU · short shared GPU demos · Hugging Face Spaces · current creation and persistence boundaries

15 executions/month; bounded CUDA exercises

cuda.live

Tiny CUDA kernel/compiler demonstrations independent of the Codabench workflow.

Boundary: Advertised T4 16 GB, five runs per 15 minutes; per-execution allowance is not GPU-hours. Existing October 8 verification, no reserved class capacity.

cuda.live · optional CUDA micro-exercises

Conditional recurring credit · $30/workspace/month; three Starter seats

Modal

DS: a small pinned training/inference run. MLIS: public-corpus OCR, embeddings or retrieval experiments.

Boundary: Payment method required; GPU, CPU and memory share credits. Set explicit usage and net-spend limits. Shared Endpoints tokens cost extra. Academic grant is an application, not funding in hand.

Modal Starter pricing and Shared Endpoints boundary · Modal billing: payment method required · Modal budgets and out-of-pocket spend limits · Modal academic credit grants

Paid usage / possible trial · Free Hobby subscription; no recurring compute credit verified

Cerebrium

DS model serving; MLIS public-corpus embeddings or transcription.

Boundary: Free plan fee and included storage do not make execution free. Trials are discretionary/time-limited and may transition to paid billing.

Cerebrium Hobby subscription versus metered compute · Cerebrium trial terms

Unconfirmed free lead · Landing page advertises 150 compute hours/month, no card

Saturn Cloud

Potential DS and MLIS notebook alternative, pending an authorized account check.

Boundary: 150 compute hours is not verified as 150 GPU-hours. Current pricing emphasizes paid hosted resources; reconcile device allowance, storage, inactivity and new-account access.

Saturn Cloud free-compute landing claim · Saturn Cloud current hosted pricing

Paid comparison · Prepaid compute; no recurring free allocation verified

Runpod

Funded DS/MLIS jobs with explicit checkpoint and external export.

Boundary: Pod launch needs credit; stopped storage still bills and can be deleted when funds run out. Not a zero-cash fallback or preservation repository.

Runpod credit balance and storage lifecycle

Limited model trial · Select-model demonstrations, then billing

Replicate

DS model-behavior comparison; MLIS controlled metadata/transcription demonstrations.

Boundary: No stable free monthly grant verified. Private deployments can bill setup and idle time. A model API does not replace arbitrary CUDA training.

Replicate limited free models and billing scope

Paid API on Free accounts · No included monthly API credits for Free users

Hugging Face Inference Providers

DS inference evaluation and MLIS embeddings/metadata tasks when funded.

Boundary: Separate from ZeroGPU/Spaces. Current API pricing overrides older free-credit snippets; no historical change date established.

Hugging Face Inference Providers: current free-account credits

Proposed pilot: develop in a bounded notebook environment; evaluate approved assignments through a Codabench queue attached to reviewed ACCESS compute where permitted. Budget operators, isolated submission workers, durable storage and exports separately. Validate scheduling, network/container rules and real class load before promising this route. Keep code, scoring bundles, submissions and results portable; a leaderboard is not the full course grade.

Research & evidence

    → / Space reveal text, then advance · ← step back · F full screen · S sources · N notes · Escape closes panels

    Presenter notes