Rent GPU Strategy (2026 pricing reality)

The rule

Rent to finish, own to iterate.

You write and debug kernels on your local 3090. You take the exact same kernel, run it once on a rented H100 to measure the real numbers, and terminate the pod within 2 hours. Anything else is burning money.

The five vendors that matter (2026)

Vendor

H100 80GB $/hr

H200 141GB $/hr

B200 180GB $/hr

MI300X 192GB $/hr

Notes

RunPod

$<phone_number_or_numberic_id_or_random_id_137> (Community/Secure)

$<phone_number_or_numberic_id_or_random_id_138>

$<phone_number_or_numberic_id_or_random_id_139>

$<phone_number_or_numberic_id_or_random_id_140>

Best UX. Docker templates. Serverless option.

Vast.ai

$<phone_number_or_numberic_id_or_random_id_141> (spot)

$<phone_number_or_numberic_id_or_random_id_142>

—

$<phone_number_or_numberic_id_or_random_id_143>

Cheapest but least reliable. Interruptible.

Lambda

$<phone_number_or_numberic_id_or_random_id_144>

$<phone_number_or_numberic_id_or_random_id_145>

$<phone_number_or_numberic_id_or_random_id_146>

—

Reserved 3-year rates lower. Reliable.

Prime Intellect

$<phone_number_or_numberic_id_or_random_id_147>

$<phone_number_or_numberic_id_or_random_id_148>

$<phone_number_or_numberic_id_or_random_id_149>

$<phone_number_or_numberic_id_or_random_id_150>

Aggregator; wide variety of GPUs.

Modal

~$<phone_number_or_numberic_id_or_random_id_151>*

~$<phone_number_or_numberic_id_or_random_id_152>*

—

—

Serverless-only (pay per second). Best for spiky loads.

*Modal charges per-second so effective rate depends on cold-start amortization. As of mid-2026, Modal’s H100 is ~$4/hr equivalent but its true value is zero-idle billing.

Pricing warning: these are Q2 2026 anchor points. Cloud GPU spot prices are volatile: check getdeploying.com or cloud-gpus.com for live comparisons before you rent.

Verification URLs (bookmark, do not memorize prices):

When to use which vendor

RunPod

Default choice for 95% of your rentals. Docker templates are excellent. Community Cloud (cheaper, no SLA) is fine for one-shot benchmark runs. Secure Cloud (SLA-backed, ~30% more) for anything you can’t interrupt. Persistent volumes let you save your Docker container between sessions.

Use for:

  • H100 kernel benchmarks (2-hour session, measure, terminate).

  • FA3 / WGMMA / TMA experiments.

  • Multi-GPU experiments (8x A100 or 8x H100 for training/TP experiments).

Setup workflow:

  1. Create a Pod template with your pinned CUDA / PyTorch / triton versions.

  2. Mount a persistent volume for your ~/kernels git repo.

  3. Start pod, git pull, run benchmark, save nsys/ncu reports to the volume, terminate.

  4. Total wall time per experiment: 30-90 min. Cost: $1-4.

Vast.ai

Use when Vast is 40%+ cheaper than RunPod for the same card (usually true for A100 80GB and older H100 stock). Interruptible spot pricing is fine for training a small model overnight if you have checkpointing.

Do NOT use Vast for:

  • Time-sensitive benchmarks (host may pause).

  • Anything with sensitive data (community-hosted; the host root user can inspect).

  • Multi-node work (aggregated hardware is unreliable).

Verification workflow: Vast lets you filter by DLPerf score, reliability rate, and machine location. Always sort by reliability > 98% and rent from 3-strike-verified hosts.

Lambda

Use when you want a proper VM experience. Longer sessions (multi-day training), reserved capacity, engineering-team-friendly. Their 1-Click Clusters let you spin up 8x H100 with NVLink for ~$25/hr on-demand.

Not competitive on hourly spot. Their strength is reserved pricing (3-year contract terms drop H100 to <$1.50/hr, but that’s not your play).

Prime Intellect

Use as a price aggregator. They pool capacity from many providers and often surface the cheapest specific GPU. Good for esoteric configurations (H100 SXM with NVLink between 2 GPUs specifically, MI300X for AMD kernel work, etc.).

Also relevant: they run distributed training as a service if you ever want to touch that (unlikely for inference eng).

What you actually spend on (13-month budget)

Realistic monthly rental need by phase

Phase

Months

What you need

$/mo

Phase 0-1

M1-M2

Nothing. Local 3090 is fine.

$0

Phase 2

M3-M4

Occasional H100 to compare kernel numbers.

$10-25

Phase 3

M5-M6

H100 for FA2/FA3. Real WGMMA numbers.

$30-60

Phase 4

M7-M8

2x-8x A100 for TP experiments; H100 for engine benchmarks.

$50-100

Phase 5

M9-M11

H100 for FP8; occasional MI300X to test AMD path.

$30-60

Phase 6

M12-M13

Multi-GPU for TP+PP+EP experiments. Rare 8x H100 bursts.

$60-150

Phase 7

M13+

Sustained mini-cluster for capstone deployment.

$80-200

Total 13-month rental cap: $500-1200. If you’re spending more, you’re using cloud as an IDE. Stop.

The “one-shot benchmark” protocol

Before you rent:

  1. Write the entire benchmark script locally.

  2. Test on your 3090; make sure it runs (correctness only, not perf).

  3. Commit to git.

  4. On the rented pod: git clone, run.sh, save results to persistent volume.

  5. Terminate within 2 hours.

  6. Analyze results locally.

Never SSH into a rented pod and start writing code. If you’re editing code on a rented pod, you’re paying $2/hr to type.

Free / near-free options

  1. Google Colab Free — T4 GPU, 12h session limit, kicks you out constantly. Useless for real work but fine for testing Triton syntax.

  2. Google Colab Pro ($10/mo) — A100 40GB (rare) or L4/T4. Better session limits. Worth it only if you can’t buy hardware.

  3. Kaggle Kernels — 30 hrs/week of P100 or T4. Free. Useful for public notebooks / portfolio pieces.

  4. Modal $30/mo free credits (new signup) — covers ~15 hours of H100. Use for one benchmark run.

  5. Vast.ai promotional credits — sometimes $10-25 on signup.

  6. AWS/GCP/Azure free credits — if you sign up as a startup or via GitHub Student, sometimes hundreds of dollars. But their default rates are 2-3x RunPod, so use credits only for what you can’t do elsewhere.

Do NOT use university cluster access (if Zoho has one) for personal portfolio work. IP contamination risk. Keep personal repos on personal hardware.

Common cost mistakes to avoid

  1. Leaving a pod running overnight — $2/hr × 10 hrs = $20 for nothing. Set a phone alarm to terminate.

  2. Renting an 8x H100 pod when 1x would do — 8x is $16-24/hr. Only rent 8x when your experiment requires 8-way TP.

  3. Downloading model weights inside the pod every session — use persistent volumes. Llama-3-70B weights = 140GB = ~10 min download at pod-startup = $0.30 burned every time.

  4. Using on-demand when spot would work — for benchmarks, spot is fine. You lose the pod, you re-run in an hour.

  5. Buying reserved capacity — don’t. You’re not at that scale. On-demand only.

Bandwidth reality (India-specific)

  • Downloading a 70B model weights = ~140GB. On a 100Mbps home connection = 3+ hours. On a 300Mbps fiber = ~1 hour.

  • Solution 1: rent a pod in Singapore (RunPod SEA) or India (RunPod India via secure partners) to minimize latency.

  • Solution 2: cache HuggingFace weights on a persistent volume. Download once, mount many times.

  • Solution 3: for the very largest experiments, download the weights inside the pod’s Docker and never bring them to your local machine.

Cross-references

  • Own GPU decision: 01_own_gpu_choice.md

  • Budget by month: 04_budget_scenarios.md

  • 2026 landscape context: 05_2026_gpu_landscape.md

  • Serving benchmark hygiene: 13_discipline/03_benchmark_hygiene.md

One line: rent RunPod H100 for 2-hour sessions, do everything else on your local 3090. Total 13-month cloud spend: aim for <$600.