Table of Contents
GPU rental is the purchase behind most AI work today — and the market quietly repriced this year. The weekly Cloud GPU Price Index sits at $3.53 per H100 GPU-hour (the median of 36 posted on-demand rates from 34 providers), the neocloud median runs $3.21 against a posted hyperscaler median of $6.98, and one-year H100 contracts climbed 38% between October 2025 and March 2026 (from $1.70 to $2.35 per GPU-hour per SemiAnalysis). Behind the numbers sits a machine worth understanding once: why neoclouds price the same chip at a third of the cloud sticker, what reserved, spot and serverless actually trade, and the arithmetic — idle hours, failed jobs, unit minimums — that decides what your workload really pays. This guide is the buyer’s decision table for renting compute in 2026, plus the honest economics for anyone eyeing the rental business itself.
Every rate below is a posted list price read from provider rate cards this month, labeled as such — negotiated contract rates are private, and this guide says so instead of pretending otherwise.
Key Takeaway
- 💸 The posted market is two-tiered: neoclouds median ~$3.21 per H100 GPU-hour on-demand against a posted hyperscaler median ~$6.98 — a 2.2× sticker gap driven by cost structure, not magic (list prices; large buyers negotiate below both).
- 🕐 Rent by workload, not by brand: on-demand for bursts and proof-of-concepts, reserved/committed for steady 24/7 work (discounts up to 60% to three-year terms), spot for interruptible batch work (Nebius from $0.79, CoreWeave spot $2.46), serverless per-second for small jobs.
- 📉 Reserved is funding the buildout: CoreWeave carries $99.4 billion of signed-but-undelivered revenue and reported $2.078 billion in Q1 2026 revenue at a 56% adjusted EBITDA margin — the contracted side, not the sticker, pays for the GPUs.
- 🧮 The unit traps cost real money: a “per GPU-hour” can mean a share of an 8-GPU minimum (AWS, Azure, Google, CoreWeave list that way) or a true single GPU (RunPod, Verda); an idle-but-billed reserved hour doubles your effective rate before failures even enter.
- 🇵🇭 The Filipino builder’s math: a $25,000-$40,000 H100 earning $14,000-$28,000 a year at 80% utilization is the reference point any freelancer, agency or startup should check before renting first — and the B200 generation’s $6-$18 rates compress the payback for the newest silicon.
What GPU rental actually costs across the market (October 2026 rate cards)
| Offer class | Per GPU-hour (posted) | Read of the range |
|---|---|---|
| H100 on-demand | $1.80 – $12.29 | Neoclouds $1.80-$6.16 (Hyperstack to CoreWeave); hyperscalers $6.88-$12.29 (AWS to Azure) |
| H100 reserved / committed | $2.97 – $7.87 | From a 3-year AWS Instance Savings Plan to a 1-year Azure reservation; DigitalOcean $3.26 at 12 months |
| H100 spot | $0.79 – $2.46 | Nebius dynamic from $0.79; Together $1.99; Azure $2.27; CoreWeave $2.46 (60% below its on-demand) |
| B200 on-demand | $6.00 – $14.24 | Hyperstack to AWS P6; Lambda $6.69; Nebius $8.50; CoreWeave $8.60; Oracle $14.00 |
| GB200 NVL72 rack access | $10.50 – $27.04 | CoreWeave on-demand to Azure pay-as-you-go; a 3-year Azure reservation reaches ~$11.90 |
| Serverless (T4-class per-second) | $0.50 – $0.81 | Koyeb $0.50; Modal/Cerebrium $0.59; Baseten $0.63; Replicate $0.81 |
| Blended 1-year contract (SemiAnalysis series) | $1.70 → $2.35 | Oct 2025 → Mar 2026 — +38% on year-long contracts (a different market from hourly cards) |
| Weekly index (median of posted H100 on-demand) | $3.53 | 36 rates across 34 providers; the per-provider median reads $3.42 |
Read the table as three different markets stacked in one list: hourly flexibility (spot and on-demand), term commitment (reserved), and unit economics per chip class (H100 vs B200 vs rack-scale GB200). The buyer’s error the market punishes hardest: comparing an on-demand sticker to another on-demand sticker across providers that price different units — some quote a share of an eight-GPU instance, others a single GPU.
Why neoclouds charge a third of the cloud for the same chip
Why is GPU rental cheaper on neoclouds? The honest answer is structure, and it explains every line above. A hyperscaler like AWS or Azure carries a catalog of hundreds of services, deep compliance portfolios and enterprise sales machinery; a neocloud — CoreWeave, Lambda, Nebius, Crusoe — sells GPU capacity almost alone, runs fewer hardware generations harder, and often holds cheaper power and colocation deals. The pricing gap is therefore real but bounded: posted hyperscaler rates are list prices that large committed customers negotiate down, so the real-world gap narrows with commitment — while the public sticker gap stays wide because the committed rates stay private.
The neocloud side has its own engine: the business is debt-funded hardware. The provider signs multi-year take-or-pay contracts (CoreWeave’s Master Service Agreement with Microsoft dates to February 2023; the recent class of contract wins includes IREN’s $9.7 billion Microsoft agreement, Nebius’s $27 billion Meta deal and CoreWeave’s $14.2 billion Meta arrangement), borrows against that contracted cash flow to buy the GPUs, and then must sell nearly every hour — an unsold GPU-hour is gone, because a neocloud cannot slide idle capacity into database or storage services the way a hyperscaler can. That single fact prices every tier: high on-demand rates buy flexibility, deep reserved discounts reward certainty, and spot prices exist to convert would-be-idle capacity into revenue.
The buyer’s decision table: which rental shape fits which workload
| Workload | Fit | Why | Watch |
|---|---|---|---|
| Proof of concept / one-off training | On-demand (or single-GPU rental) | No commitment to amortize; shop per-GPU pricing (RunPod, Verda, Lambda’s cheaper on-demand class) | Eight-GPU minimums at hyperscalers price you a node, not a chip |
| Steady 24/7 production (12+ months) | Reserved / committed | The deep discounts (up to 60% at three years) are built for contracted cash flow | Idle-but-billed hours; count the utilization you will actually reach |
| Interruptible batch / checkpointed training | Spot | 60% below on-demand at CoreWeave; dynamic floors at Nebius ($0.79) | Interruptions are the contract — the work must resumable-checkpoint |
| Small inference, spiky traffic | Serverless | Per-second billing ($0.50-$0.81 T4-class); zero idle cost | No VM access on some platforms; cold starts; model-fit limits |
| Multi-week training runs | Blended (reserved base + spot overflow) | Base capacity reserved; surge absorbed on spot | Two vendors’ reliability profiles to manage |
The arithmetic behind the GPU rental table, made concrete: at a $2.50 reserved rate, using only half the hours you pay for turns the effective compute cost into $5.00 per used hour; if failures also consume 8% of those used hours, the real cost is about $5.43 per useful GPU-hour. The rate card told you $2.50; the utilization told the truth. Add egress, storage and data-transfer charges (excluded from GPU-hour quotes) and the comparison finishes on the invoice, not the index.
The rental business itself: who earns, who borrows, who pays
For freelancers, agencies and startups weighing the business side — the ring’s “gpu rental business” question — the reference economics are now public. A used-generation H100 costs roughly $25,000-$40,000 at market; at 80% utilization and the stabilized $2-$4 hourly spot band, one card grosses $14,000-$28,000 a year — which is why the H100’s sub-a-year payback period has stretched, and why the blackwell-class B200 (roughly $30,000-$35,000 per unit renting $6-$18) compresses payback for whoever holds the newest silicon. The structural winners in the wholesale tier are the players who locked power before the rush (IREN and Applied Digital both arrived from bitcoin mining with long-term power agreements priced below the market), because a single GPU rack burns 30-40 kilowatts and power procurement is the differentiator that new entrants cannot buy retroactively.
WorldNgayon Insight: What changed: GPU rental matured into a two-tier capital market — retail hourly cards on top, multi-billion contracted backlogs underneath. Why now: anchor contracts (Microsoft-Meta-class) made neocloud debt fundable at scale. Why it lasts beyond the news cycle: pricing structure (flexibility premium, commitment discount, spot floor) is now the industry’s visible operating law, updated weekly by the public indexes. Who is affected: AI startups pricing roadmaps, freelancers and agencies choosing rent-vs-build, investors reading backlog quality. What to do next: price your workload per useful GPU-hour (utilization and failures included), not per posted hour. Risks, opportunities, costs, skills: the risk is contract lock-in at the top of a rate cycle (the +38% contract move cuts both ways); the opportunity is spot-plus-reserved blending for bursty teams; the cost traps are unit minimums and idle commitments; the skill gap is capacity planning. Connection to work and money: the rent-vs-build decision is the compute twin of the licensing decision in our AI model licenses guide — both reward reading the contract before the demo. The buyer’s economics here also flow through the national compute ladder and the inference-cost structure we track. The gpu-rental-guide lane is where the buyer’s how buyers decide question gets decided.
For Filipino readers: renting smart first, building never too early
The Filipino builder’s sequence is renting-first by arithmetic. A team validating an agentic service or fine-tuning on a small corpus pays on-demand or serverless rates in pesos-scale tens of thousands, not the seven-figure capex of a rack — and the single-GPU rental market (RunPod-class, $2-$3 hourly for an H100) makes burst experiments cheap and cancellable. The build-versus-rent line moves when three things hold at once: a steady 24/7 workload with predictable utilization, sustained contracted demand behind it, and power that is already secured — the same three-legged rule the wholesale tier just proved at billion-dollar scale. Power is the Philippine twist: a rack’s 30-40 kW appetite prices grid contracts and cooling before silicon, which is why the region’s data-center buildouts (our SEA scoreboard and the power-pipeline analysis) decide where local GPU hosting can even compete. For OFW tech workers, the ring’s buyer skills — reading rate cards, modeling utilization, blending spot with reserved — are contract-negotiation skills that travel to any market. The how buyers decide logic reads identically from any market.
How much does it cost to rent an H100 GPU?
Posted on-demand rates run about $1.80-$6.16 per GPU-hour at neoclouds (median ~$3.21 in tracked tables this month) and $6.88-$12.29 at hyperscalers; reserved terms drop into the $2.97-$3.26 range at multi-year commitments, and spot pricing reaches $0.79-$2.46. The weekly index median of posted on-demand rates sat at $3.53 in the tracked mid-October week — and none of these figures include negotiated discounts, storage or data transfer.
Why is GPU rental cheaper on neoclouds?
Structure: neoclouds sell GPU capacity almost exclusively, run leaner compliance and sales stacks, hold cheaper power and colocation contracts, and run single generations of hardware harder. The counterweights: posted hyperscaler rates are list prices negotiated down for large customers, and neoclouds trade ecosystem depth and reliability guarantees for the lower rate.
How do GPU rental discounts work?
The discount ladder buys the provider certainty to fund hardware: on-demand (no commitment, highest rate), reserved or committed use (up to 60% off at CoreWeave for multi-year terms, up to 35% at Nebius), spot (leftover capacity, 60%+ below on-demand, interruptible), and per-second serverless for small jobs. Reserved pricing is what lenders underwrite, which is why the discount is deepest at multi-year terms.
What is a neocloud?
A cloud provider built around GPU rental almost exclusively — CoreWeave, Lambda, Nebius and Crusoe are the classified giants — rather than a hyperscaler’s broad catalog. The business funds itself with multi-year contracted cash flow (take-or-pays from Microsoft and Meta-class buyers), buys GPUs on debt against that backlog, and must sell nearly every billed hour because idle capacity has no fallback service to absorb it.
Should I rent GPUs or buy them?
Rent until three conditions hold together: a steady 24/7 workload with high predictable utilization, sustained demand behind the spend, and secured power at reasonable rates. An H100 rents in the $2-$6 hourly band posted while the same card costs $25,000-$40,000 to buy and earns $14,000-$28,000 a year at 80% utilization — the rental premium buys flexibility, upgrades and zero idle risk; the build only wins on steady, well-utilized, power-secured workloads.
What GPU rentals are available for small jobs or bursts?
Single-GPU on-demand rentals (RunPod, Verda and similar price single cards rather than eight-GPU nodes), per-second serverless inference ($0.50-$0.81 for T4-class on Koyeb, Modal, Baseten, Replicate), and spot capacity from about $0.79 (Nebius) to $2.46 (CoreWeave) for interruptible batch work. Match the shape to the job: bursts and proof-of-concepts never need commitments.
What hidden costs should I watch in GPU rental?
Five: eight-GPU instance minimums that price a node behind a “per GPU-hour” quote; reserved hours billed while idle (halved utilization doubles the effective rate — the $2.50-vs-$5.00 arithmetic); failed-run consumption on spot; storage and egress/data-transfer charges outside the GPU-hour; and availability gaps — a published rate card is not a capacity guarantee.
Is there a GPU rental price index I can check?
Yes — two public reads: the weekly Cloud GPU Price Index (a posted-rate median across ~34 providers, $3.53 in the mid-October 2026 week) and daily-verified tracker tables that rank 35+ neoclouds against posted AWS, Azure and Google Cloud rates (neocloud H100 median $3.21 vs hyperscaler $6.98). Read the methodology: listed rates exclude negotiated terms, and per-provider medians differ from all-rate medians.
Financial Disclaimer
This article is editorial analysis of publicly reported prices, contracts and market data, not investment, financial or procurement advice, and no outcome is promised. Posted rates, contract discounts and market projections change; verify current terms with the named providers and institutions. Compute, technology and capacity markets carry substantial risk.
Sources: GPU Economy — the weekly Cloud GPU Price Index (methodology and receipts published); GPURentalPrices.com — daily-verified neocloud and hyperscaler rate tables; What is a neocloud and how does it price GPUs — the CoreWeave/Ornn/SemiAnalysis rate reading; The Neocloud Guide — published rates per GPU-hour read from rate cards. Additional context: Springbok Finance’s neocloud unit-economics review (H100 and B200 purchase-price and payback figures; utilization math). All rates are posted list prices with their read dates; negotiated contract rates are private.
