H100 GPU Price per Hour — August 2026 (India & Global): Spot vs On-Demand, Trend and Forecast
H100 GPU Price per Hour — August 2026 (India & Global): Spot vs On-Demand, Trend and Forecast
As of August 6, 2026, the cheapest H100 rental is ₹209/hr on-demand in India (E2E Networks) and $0.99/hr spot on Vast.ai/Spheron; typical on-demand is $2.50/hr, hyperscalers $7.50–$8.50/hr. Prices live from the gpuprice.in feed updated 2026-08-06.
*Published August 6, 2026 · Updated from the live gpuprice.in pricing feed (source: Parallel API) · This is the dated August 2026 snapshot of the H100 price tracker.*
What does an H100 cost per hour in India in August 2026?
In India in August 2026, you can rent an NVIDIA H100 for ₹209/hour on-demand (E2E Networks' current listed deal) down to $0.99/hour spot (≈₹83/hr) if you route through Vast.ai or Spheron and tolerate interruption. Reported India-hosted H100 rates sit in the ₹219–₹400/hour range (CloudMinister lists ₹219–250/hr, JarvisLabs lists ₹255.15/hr, E2E lists ₹350–400/hr on its GPU rental page), so the ₹209/hr deal is at the bottom of the range. The single most important number for an Indian buyer is that a dedicated, always-on H100 on a global neo-cloud like Spheron or Vast.ai costs about $2.50/hr (≈₹208/hr), which is effectively the same as the best in-country deal — latency and billing currency, not sticker price, are what separate India-hosted from global rental.
How to read this H100 price tracker
This is the dated August 2026 snapshot of the gpuprice.in H100 price tracker. Every price below is pulled from the live feed at `/root/penecho-demo/public/pricing-data.json`, `last_updated 2026-08-06T09:41Z`, source Parallel API, and reflects what a provider lists today. Because GPU rental pricing moves weekly (H100 spot briefly traded above on-demand on Spheron in July 2026), treat this as the price *right now* — not a guaranteed future rate. The companion page spot vs on-demand H100 2026 explains when to pay the on-demand premium.
Access date for all prices in this article: 2026-08-06. If you are reading this later, re-check the live pricing table.
H100 price per hour — live table (August 2026)
The table below is the full H100 block from the live gpuprice.in feed (on-demand and spot, per GPU per hour, USD unless noted). Providers are consumer "neo-cloud" marketplaces through hyperscalers.
| Provider | H100 On-Demand ($/hr) | H100 Spot ($/hr) | On-Demand → Spot Savings |
|---|---|---|---|
| Vast.ai | $2.50 | $0.99 | 60% less |
| Spheron | $2.50 | $0.99 | 60% less |
| E2E Networks | $2.70 (₹209 deal) | $1.50 | 44% less vs on-demand |
| Nebius | $3.20 | $1.80 | 44% less |
| CoreWeave | $3.30 | $2.10 | 36% less |
| Lambda | $3.49 | $2.50 | 28% less |
| RunPod | $3.90 | $2.50 | 36% less |
| GCP (Google) | $7.50 | $4.50 | 40% less |
| Azure (Microsoft) | $8.00 | $4.80 | 40% less |
| AWS | $8.50 | $5.00 | 41% less |
*Data source: gpuprice.in live feed (last_updated 2026-08-06T09:41Z, source parallel.ai). E2E's ₹209/hr is the listed deal from the feed's deals block; its GPU rental page shows an H100 range of ₹350–400/hr.*
What the table tells you
Three things stand out in the August 2026 snapshot. First, the spot spread is huge but narrowing: spot-to-on-demand savings run 28–60%, with Vast.ai and Spheron offering H100 spot at under a dollar an hour. Second, hyperscalers still charge 3–4x the neo-clouds for the same H100 silicon — AWS lists $8.50/hr on-demand against $2.50/hr on Vast.ai/Spheron — so the "AWS price" you see quoted in most headlines is not the market-clearing price. Third, E2E's ₹209/hr deal makes India genuinely price-competitive with the global neo-clouds, which was not true earlier in the year.
Has the H100 price gone up or down recently?
Over the last few months the H100 on-demand price has been roughly flat, with the significant movement at the margins: spot capacity tightened in July 2026, and on Spheron H100 spot (≈$2.91/hr) briefly traded at or above on-demand (≈$2.54/hr) — a sign that interruptible capacity was scarce enough that the "discount tier" lost its discount. Compared with the May–June 2026 baseline in the previous H100 trend article, headline on-demand rates (AWS $8.50, Azure $8.00, GCP $7.50, Lambda $3.49, Vast $2.50) are unchanged, which is itself the story: after the sharp 2025–2026 run-up, on-demand H100 has plateaued. The volatile variable is now spot, and the cheap-tier leader (Vast/Spheron $0.99) is where you'll see the next move if B200 supply keeps pushing H100 demand sideways.
What is the H100 spot price vs on-demand price right now?
Right now, H100 spot is $0.99/hr on Vast.ai and Spheron, versus $2.50/hr on-demand on those same providers — a 60% saving for interruptible capacity. Across all providers in the feed, spot runs 28–60% below on-demand, with the Neo-cloud spot leaders (Vast $0.99, Spheron $0.99, E2E $1.50, Nebius $1.80) well under half the hyperscaler spot (AWS $5.00, Azure $4.80, GCP $4.50). The catch is preemption: spot instances can be reclaimed when demand spikes, so they suit batch inference, fine-tuning and distributed training that can checkpoint and resume — not real-time services. If your workload must stay up, pay the on-demand rate; the 60% spot saving is real but conditional on interruption tolerance.
Is it cheaper to rent an H100 or buy one in India?
For almost every Indian buyer, renting an H100 is cheaper than buying one in 2026, unless you run it above roughly 70% utilization around the clock. An H100 SXM (80GB) costs approximately $30,000–$40,000 of hardware, and India adds import duty plus 18–28% GST on top, so an owned card starts well north of ₹30L once cabled into a server, power, and a machine room. Renting that same card at E2E's ₹209/hr for a full 730-hour month is ≈₹1.53L/month; at Vast/Spheron $2.50/hr on-demand it's ≈₹1.52L/month. At that monthly cost, continuous 100% utilization takes roughly 16–24 months to break even on ownership — and most projects run at 20–40% utilization, stretching effective payback to 3+ years. Rent unless you have a genuinely saturated, always-on workload. For the full India TCO math, see the DGX Station vs GPU cloud comparison.
How does H100 price compare to H200 and A100 in August 2026?
In August 2026, H100 at $2.50/hr on-demand sits between the older A100 at $1.67/hr and the H200 at $2.80–$3.10/hr on the same neo-cloud marketplaces. The pricing ladder across the feed is: A100 80GB on-demand from $1.67/hr (Vast/Spheron, $0.61/hr spot), H100 80GB from $2.50/hr on-demand ($0.99/hr spot), H200 141GB from $2.80/hr (Spheron) to $3.10/hr (Vast), and B200 192GB from $4.50/hr on-demand (Spheron). The buying decision is memory: H100's 80GB HBM3 fits a single 13B–70B model comfortably; H200's 141GB is the memory bridge for 140B-class models that H100 cannot hold; A100 is the price-efficient fine-tuning workhorse for 7B–13B. For a full workload-by-workload comparison, see H100 vs H200 vs B200 2026.
Where is the cheapest H100 rental in India?
The least expensive H100 rental accessible to an Indian user right now is $0.99/hr spot on Vast.ai or Spheron (≈₹83/hr) for interruptible capacity, and ₹209/hr on-demand via E2E Networks' current listed deal for guaranteed capacity. India-hosted H100 on-demand rates otherwise run ₹219–400/hr: CloudMinister lists ₹219–250/hr, JarvisLabs lists ₹255.15/hr, and E2E's GPU rental page shows ₹350–400/hr. Two practical notes for Indian teams: first, Neo-cloud spot at $0.99/hr is the true cost floor but comes with preemption risk; second, an India-hosted provider (E2E, JarvisLabs) gives INR billing, lower egress latency, and local compliance, which matters if you are subject to India data-residency rules — see E2E Networks vs Vast.ai for India GDPR.
How much does one month of continuous H100 compute cost in India?
One month of continuous H100 compute in India costs about ₹1.53L at E2E's ₹209/hr deal (730 hours), and roughly the same (≈₹1.52L) on Vast.ai/Spheron at $2.50/hr on-demand. Here is the math: 730 hours ÷ month, so ₹209/hr × 730 = ₹1,52,570/month; $2.50/hr × 730 = $1,825 ≈ ₹1,52,000/month (assuming ≈₹83/USD). Renting the same card on spot at $0.99/hr drops it to ≈₹60,000/month, but only if your workload survives preemption. Run the numbers against your projected utilization before committing — the monthly figure below a 70% utilization breakeven favors rental in every case we priced.
Will H100 prices fall when B200 supply increases?
We do not know, and we will not pretend to forecast it. The honest answer is that the relationship between B200 supply and H100 pricing is not currently forecastable from the public data. B200 rental has been volatile in 2026 (averaging ~$5–6/hr through mid-2026, with neo-cloud spot dipping to ~$3.35–3.70/hr while hyperscalers charge up to ~$14/hr), and B200 on-demand is available from $4.50/hr in this month's feed. On one hand, a flood of B200 capacity could relieve H100 demand and pull H100 prices down; on the other, H100 remains the default for many 70B-class workloads the B200 simply isn't needed for, so demand may prove sticky. State a directional guess, not a forecast, and re-check the live table monthly.
Does IndiaAI Mission offer subsidised H100 rates?
Not confirmed in a primary source, so we will not state a subsidised H100 number as fact. CloudMinister lists a subsidised H100 rate of ₹65/hr under IndiaAI Mission, but that figure is not corroborated by an official IndiaAI Mission tender document in the evidence set, and the available published IndiaAI numbers concern the B200 tender (L1 ₹290.70/hr, 12-month reservation ₹251.10/hr, a ~10% cut vs the prior ₹323 round; B300 ₹351–₹2,808/hr) — not H100. Per the source-quality rules, a claim needs a primary source; the IndiaAI H100 subsidy is unverified and must be confirmed against an official tender before it appears in print. If you are chasing subsidised compute, treat the verified B200 numbers as the referenceable ones and flag the H100 subsidy as unconfirmed.
What AI workloads is the H100 still the right choice for in 2026?
The H100 remains the right choice in 2026 for training and inference of a single 7B–70B model where price-per-token beats both A100 and B200, and where the 80GB memory tier is sufficient. For a 7B–13B fine-tune, A100 at $1.67/hr is the cheapest workhorse; for a single 70B-class model, H100's 80GB HBM3 and higher memory bandwidth make it the value pick at $2.50/hr; for a 140B-class model that exceeds 80GB, step up to H200 (141GB) at $2.80–$3.10/hr or B200 (192GB) at $4.50+/hr. The H100 is also the safe default for most production inference because of mature CUDA tooling, wide provider availability, and the largest spot pool in the market. Choose H100 when 80GB fits and price-per-token matters; let memory size, not sticker price, be the deciding factor when it doesn't.
H100 vs H200 vs China-export A100: what the August 2026 numbers mean for Indian AI teams
For an Indian AI team the August snapshot collapses to one practical decision rule. If your model fits in 80GB, H100 at $2.50/hr (or $0.99/hr spot) is the value sweet spot; if it needs more memory, pay for the tier that holds it — H200 at $2.80–$3.10/hr or B200 at $4.50+/hr. India-hosted options (E2E ₹209–400/hr, JarvisLabs ₹255/hr) are now within ~0–15% of the global neo-cloud price band, which removes the old "India is always pricier" caveat. The two traps to avoid: buying hardware for a workload you'll run at 20–40% utilization (rent instead), and chasing the headline on-demand savings from hyperscaler comparisons that quote AWS $8.50/hr as if it were the market — the real market floor is $0.99/hr spot on Vast.ai/Spheron.
Frequently asked questions
Is H100 still worth buying in August 2026?
No, for most teams. With H100 on-demand at $2.50/hr and spot at $0.99/hr, renting is cheaper than owning unless you sustain roughly 70%+ utilization around the clock — most projects run 20–40%, where effective ownership payback stretches past three years after India import duty and GST.
How much does an H100 cost to rent for a day?
At $0.99/hr spot, 24 hours is about $24 (≈₹2,000); at $2.50/hr on-demand it is about $60 (≈₹5,000); and at AWS's $8.50/hr it is $204. Daily rental only makes sense for short experiments — monthly commits cut the rate on most providers.
What is cheaper, H100 or A100?
A100 is cheaper ($1.67/hr on-demand vs $2.50/hr H100), but H100 is faster per token and fits a full 70B model in its 80GB memory. For 7B–13B fine-tuning, A100 is the better value; for single 70B-class work, H100 wins on cost per token.
Can I run a 70B model on an H100?
Yes. H100's 80GB HBM3 can host and run a single 70B-class model (quantized) comfortably at $2.50/hr on-demand or $0.99/hr spot — this is the core reason H100 remains the default for many 70B workloads even as B200 arrives.
What is the H100 price at AWS vs Vast.ai?
AWS lists H100 on-demand at $8.50/hr and spot at $5.00/hr; Vast.ai lists $2.50/hr on-demand and $0.99/hr spot. That is a 71% price gap on the same silicon — the difference is managed service, reliability guarantees, and egress differences, not the GPU itself.
Does IndiaAI subsidise H100 rental?
Unverified. An official IndiaAI Mission tender confirms B200 at ₹290.70/hr (L1) but no primary source in this evidence set confirms an H100 subsidy; the ₹65/hr figure attributed to IndiaAI is not corroborated and should not be stated as fact until confirmed from an official document.
Methodology, sources and data notes
Live table: all prices are the H100 block of `/root/penecho-demo/public/pricing-data.json`, `last_updated 2026-08-06T09:41Z`, source Parallel API, accessed 2026-08-06. gpuprice.in re-reads this file at write time rather than trusting a cached copy. INR conversions use ≈₹83/USD and 730 hours/month.
Primary sources (accessed 2026-08-06):
- Spheron — GPU Cloud Pricing Comparison 2026 — H100 $2.50/hr on-demand; notes July 2026 H100 spot (≈$2.91) briefly ≥ on-demand.
- E2E Networks — GPU Rental India — H100 ₹350–400/hr range; current ₹209/hr listed deal.
- JarvisLabs India — H100 ₹255.15/hr.
Unverified / do-not-state: IndiaAI Mission H100 subsidy (₹65/hr claim lacks a primary source). Forward H100 price forecasts are flagged as not forecastable from public data. Where a tier is missing in the feed (marked `null`), the provider is treated as not publishing that tier rather than inventing a rate.
Disclosure: gpuprice.in may earn commissions through provider affiliate links at no extra cost to you. Prices are real, current as of the access date, and change often.