GPU Rental Markets

Research dump (July 2026) answering: Where can I rent GPUs? How do you technically access a rental GPU? Can you write custom kernels without owning one? From Questions. Related: GPUs, Frontier Chips. Key citations verified July 2026.

Where can I rent GPUs?

The market has stratified into ~5 layers with a 3–12x price spread for identical silicon (H100: ~$1.29/hronmarketplaces\to$7–12/hr on hyperscalers; market median ~$2.45–3/hr). SemiAnalysis’s ClusterMAX 2.0 rating tracks 209 providers (84 rated) — it’s that crowded. Newest development: compute is becoming a listed commodity, with GPU futures launching on ICE and CME in 2026.

Hyperscalers — pay for compliance/ecosystem, not compute value

• AWS — P5/P6 instances, Capacity Blocks for short-term reservations; H100 ~$3.90–6.88/hr on-demand
• Azure — ClusterMAX Gold; also rents heavily from neoclouds ($19B Nebius deal, plus Nscale, Lambda, CoreWeave)
• GCP — A3/A4 instances, TPUs as alternative, Dynamic Workload Scheduler
• Oracle OCI — Gold; bare-metal superclusters (incl. AMD MI300X), OpenAI Stargate anchor

Neoclouds (GPU-native clouds)

• CoreWeave — the only ClusterMAX Platinum; K8s-native, earliest Blackwell/Rubin allocations, public; reserved ~$1.45/hr beats nearly everyone
• Nebius (ex-Yandex) — Gold; cheapest committed H200 ($2.30/hr),$19B Microsoft contract, EU+US
• Lambda — developer-first, “1-Click Clusters” with InfiniBand, IPO track
• Crusoe — stranded-energy angle, builds Stargate Abilene for OpenAI
• Fluidstack — Gold; lean team assembling giant training clusters fast (Mistral, Meta)
• Together AI — inference API brand + rentable Slurm/K8s GPU clusters
• Voltage Park — nonprofit-owned (~24k H100s), H100 from $1.99/hr
• TensorWave — AMD-only (MI300X/325X), $350M Series B
• Verda (ex-DataCrunch, Finland) — cheapest H100 spot floor (~$0.80–1.14/hr), EU green energy
• Mid-tier: Hyperstack (~$1.90/hr),\ast \ast Vultr\ast \ast ,\ast \ast DigitalOcean/Paperspace\ast \ast ,\ast \ast Scaleway\ast \ast (wontheEuropeanCommissionsovereign-cloudcontract),\ast \ast OVHcloud\ast \ast ,\ast \ast CUDO\ast \ast ($1.82/hr), Civo, Latitude.sh ($1.66/hrbaremetal),\ast \ast Hetzner\ast \ast ,\ast \ast MassedCompute\ast \ast ,\ast \ast GMICloud\ast \ast (building1GWsovereignAIinfrainJapanonVeraRubin),\ast \ast Corvex\ast \ast (compliance-heavyBlackwellcloud),\ast \ast WhiteFiber\ast \ast ,\ast \ast ThunderCompute\ast \ast ("VMwareforGPUs"—virtualizedH100$1.38/hr)
• Bitcoin-miner pivots: IREN (~$500MAI-cloudARR,directGPUcloud),\ast \ast AppliedDigital\ast \ast ($7.5B hyperscaler lease), TeraWulf ($19B Anthropic lease), Hut 8, Bitdeer, Hive

Marketplaces / aggregators / brokers

• Vast.ai — the original “Airbnb of GPUs”; per-second billing, interruptible auctions, H100 from ~$1.33/hr, consumer cards from cents/hr; cheapest floor, reliability varies by host
• RunPod — dev-focused hybrid: Community vs Secure tiers, serverless with sub-200ms FlashBoot cold starts, Instant Clusters
• SF Compute — true order-book spot market for H100/H200 blocks; buy by the hour and resell unused time; owns no GPUs; $40M Series A; positions as the physical base for compute derivatives
• Shadeform (YC S23) — one API/console over 20–30+ providers at list prices, no markup
• Prime Intellect — compute exchange over 12+ clouds + decentralized training + RL environments hub; $130MSeriesA@$1B (July 2026, Radical Ventures; NVentures, Intel Capital, Dell participating)
• Mithril (ex-Foundry) — spot/preemption arbitrage “omnicloud”; customers Cursor, LG, Arc Institute
• Hydra Host — bare-metal brokerage (Brokkr); $100M Series A (June 2026, Kindred Ventures, NVIDIA participating)
• TensorDock — vetted-host marketplace, full KVM VMs
• NVIDIA DGX Cloud Lepton — NVIDIA now runs its own rental marketplace routing developers to partner neocloud capacity
• Price comparison: getdeploying.com, gpucloudprices.com, computeprices.com

Financialization (the novel 2026 wave, all early-stage)

• Ornn — GPU compute futures with ICE (May 2026; “trade AI compute like oil”)
• Silicon Data — SDH100RT H100 rental-price index; first listed compute futures with CME Group (May 2026)
• Pluto — YC-backed, 4-person startup; aiming to be the “first CFTC-regulated exchange for AI infrastructure” (DCM/DCO applications filed, not yet approved)
• Stoa Exchange — cash-settled futures/options on GPU hardware
• Capacity Derivatives — OTC forwards/swaps on Canton blockchain
• Compute Exchange (TCEX) — auction-based exchange, co-founded with auction theorist Paul Milgrom

Decentralized / DePIN

Akash (on-chain reverse auctions), io.net (Solana, pivoting enterprise), Aethir (real enterprise revenue), Hyperbolic (H100 at $1.29/hr—cheapesttrackedinoneJuly2026index),\ast \ast Salad\ast \ast (gamer{s}^{\prime }idle4090s$0.20/hr), Nosana, Spheron, Render (3D rendering more than ML), Gensyn (verifiable decentralized training, testnet stage). SemiAnalysis rates essentially all DePIN “Underperforming” for serious multi-node training — they compete on price for inference/batch jobs.

Serverless GPU (rent by the second)

Modal (Python-native, sub-second cold starts), Replicate (acquired by Cloudflare), Baseten, fal.ai (generative media), Beam, Cerebrium, Koyeb, Lightning.ai, RunPod Serverless. Adjacent: token-priced inference (rent outcomes, not GPUs) — see Serving Models.

Sovereign / regional

HUMAIN (Saudi, ~600k GPUs planned by 2030), Core42/G42 (UAE), Yotta (India, 20k+ B300s), Sakura (Japan), SMC/Firmus (Singapore, immersion-cooled), Scaleway/OVHcloud (EU), TELUS (Canada). Fastest-growing new category in 2025–26.

How do you technically access a rental GPU?

Four access modes:

1. SSH into VM/bare metal (Lambda, hyperscalers, CoreWeave dedicated) — provider preinstalls the NVIDIA driver; Lambda images ship drivers + CUDA + Jupyter. Standard workflow is VS Code/Cursor Remote-SSH. Hyperscaler gotcha: default GPU quota is 0, must request increases.
   • Verified July 2026, Prime Intellect A100 80GB on-demand: raw Ubuntu 22.04 VM (full sudo apt access, not a locked container), driver present (nvidia-smi reports the CUDA version the driver supports) but no CUDA toolkit preinstalled — nvcc genuinely missing, unlike Lambda’s driver+CUDA+Jupyter baseline. Anything needing to JIT-compile a kernel at runtime (e.g. vLLM’s FlashInfer fused sampler) will crash until you install a matching toolkit yourself. Also: your own SSH public key must be registered in the Prime Intellect dashboard before the instance is created (key injection is first-boot-only — a bad/late key means terminate + relaunch, not a fixable-in-place problem); a non-default-named local key isn’t auto-offered by ssh, needs -i explicitly.
2. Docker container rentals (Vast.ai, RunPod pods) — your rental is literally a docker create on someone’s host, with an SSH daemon or Jupyter injected. Key division of labor: the host owns the kernel driver + libcuda.so; your container brings the CUDA toolkit (nvcc, libcudart). Container CUDA version must be ≤ host driver’s supported version.
3. Managed Slurm/K8s clusters (CoreWeave SUNK, DGX Cloud, reserved clusters) — SSH to a login node (no GPU), shared NFS home, srun/sbatch for GPU jobs, containers via Pyxis/Enroot.
4. Serverless (Modal, RunPod Serverless, Replicate) — never see a machine; decorate a Python function with gpu="H100". Modal’s stack: gVisor sandbox, lazy-loaded image FS, CUDA checkpoint/restore for sub-second cold starts.

Can you write custom kernels without owning a GPU? Yes.

Compiling and running CUDA/Triton kernels works on every rental type, including serverless — people run full CUTLASS/B200 kernel labs on Modal. Timing-based benchmarking (CUDA events, torch.profiler, Nsight Systems timelines) also generally works.

The one real constraint is hardware-counter profiling (Nsight Compute ncu): since driver 418, GPU perf counters are admin-restricted (ERR_NVGPUCTRPERM). You need bare metal, a dedicated VM, or --cap-add=SYS_ADMIN:

• Bare metal / dedicated VMs (Lambda, CoreWeave dedicated, Vast VM-mode): full ncu/nsys works
• Container marketplaces (RunPod, Vast Docker mode): hit-or-miss; Vast’s answer is “use a VM instance”
• Serverless: structurally blocked (no SYS_ADMIN in multi-tenant sandboxes)
• Avoid MIG/vGPU/fractional GPUs for kernel work — CUPTI profiling APIs unsupported, no P2P across slices. Rent a whole physical GPU.

Cheap/free options for learning kernel dev

• Google Colab free T4 — nvcc preinstalled, nvcc4jupyter cell magic, srush’s GPU-Puzzles; Kaggle similar
• Modal — $30/month free credits, per-second H100/B200 billing
• Lightning AI — free monthly GPU hours, SSH + any IDE
• LeetGPU — browser CUDA problems + free simulator for zero-cost correctness checks
• Tensara — competitive kernel challenges (CUDA/Triton/Mojo) on real T4/A100/H100, leaderboards vs PyTorch baselines
• GPU MODE Discord (~20k members) — free kernel submissions via popcorn-cli to donated T4→B200/MI300 hardware, ~15s turnaround, weekly lectures; all 60k+ submitted kernels open-sourced
• Raw SSH floor: RunPod/Vast spot RTX 3090/4090 from ~$0.10–0.40/hr