H100 Rental Prices Compared: $1.49-$6.98/hr Across 15+ Cloud Providers (2026)

H100 pricing cannot be expressed as one universal on-demand rate: providers package H100 GPUs in different node sizes, regions, capacity products, and service tiers. On its current Capacity Blocks page, AWS lists P5.48xlarge Capacity Blocks at $41.528 per hour ($5.191 per H100) in several U.S. regions and $37.76 per hour ($4.720 per H100) in several other listed regions. AWS Capacity Blocks are prepaid capacity products, not on-demand EC2 pricing. Comparisons below identify the product and commitment basis rather than treating prices as interchangeable.

Key findings: The cheapest current H100 rental rates (per GPU-hour) by vendor include:

• AWS EC2 Capacity Blocks (P5.48xlarge): $5.191 per H100-hour in US East (N. Virginia), US East (Ohio), and US West (Oregon); this is prepaid capacity, not on-demand EC2 pricing (^[1]).
• Google Cloud: excluded from the numerical comparison because a current rate must be selected for a specific region and purchase option in Google Cloud’s pricing workflow (^[2]).
• Microsoft Azure (NC H100 v5) and Oracle Cloud (BM.GPU.H100.8): excluded from the current comparison because this article does not establish a region-specific, current public hourly price and matching purchase model from the providers’ first-party pricing tools.
• Runpod Pods: $2.89 per hour for H100 PCIe, $3.29 for H100 SXM, and $3.19 for H100 NVL on the displayed Pods price list (^[3]). These are SKU-specific Pod rates, not a market-wide benchmark. Other vendors are omitted from the current-price summary when a matching public first-party hourly rate was not verified.

Prices vary materially by region, capacity product, operating system, service tier, and commitment. Historical tracker snapshots and promotional rates are not current-price evidence. AWS Capacity Blocks are prepaid reservations, while Runpod’s displayed Pod rates are SKU-specific; neither should be used to infer a market-wide on-demand range.

Despite the drop, H100 rent remains a premium compared to older GPUs (A100s are now sub-$1/GPU-hr open-market (^[4])). The sustained price reductions reflect oversupply and competition, as well as long-term commitments and spot markets putting downward pressure (^[4]) (^[5]). We examine the historical context, current pricing data, multi-provider comparisons (with extensive citations), market dynamics, and projected trends below.

The NVIDIA H100 Tensor Core GPU is a Hopper-architecture accelerator offered in 80 GB configurations and used for large-language-model training, HPC, and cloud AI workloads. It is no longer NVIDIA's newest data-center architecture: NVIDIA positions B200 systems as Blackwell-based products. Released in 2022–2023, the H100 offered unprecedented performance (e.g. multi-Exaflop on DGX systems) and swiftly became the gold standard for cutting-edge AI research (^[6]). However, its sticker price at retail (~$25–40K/GPU (^[7])) makes ownership difficult for many, driving demand for rental solutions.

In parallel, the GPU-accelerated cloud computing market has exploded in recent years. The global “GPU-as-a-Service” market was only $3.34 billion in 2023 but is projected to reach $33.9 billion by 2032 (^[4]). This growth is driven by AI/ML needs, making high-end GPUs like the H100 in continuously tight supply. Industry analysts note that “GPU prices remain sky-high in 2025” and that renting is often the only viable option for smaller players (^[7]). As adoption grows, vendors have been incentivized to lower prices: Thunder Compute reports H100 rental rates falling from $8/hr to $2.85–3.50/hr in 2025 (^[4]).

Rental models: Cloud providers typically charge by the hour (or minute) for GPU instances. Prices depend on instance type, region, and commitment. On-demand (no-commitment) is highest; spot/preemptible instances can be 60–90% cheaper (^[8]) (^[9]); and 1–3 year commitments (Reserved/Savings Plans) offer up to ~45–50% further discounts (^[10]) (^[9]). In practice, many users mix and match: e.g. burst on-demand or spot when needed, while using reserved instances for steady-state workloads. The market spans major hyperscalers (AWS, Google, Azure, etc.), specialist AI/cloud vendors (Lambda Labs, CoreWeave, Runpod, etc.), and even AI-focused marketplaces (Vast.ai, etc.).

This report compares a limited set of publicly displayed H100 prices checked in August 2026. It does not cover every vendor or establish a market-wide price range. Several providers require a calculator, account access, or a sales quote for current pricing. The comparison uses current first-party pricing only where a matching hourly product is publicly displayed; historical material is context only, not current-price evidence.

H100 Release and Early Pricing

When NVIDIA announced the H100 in mid-2022 and it entered general availability in 2023 (^[11]), it was the fastest GPU for AI training. At launch, public on-demand rental pricing was extremely high. For instance, in mid-2023 AWS’s P5 instances (8×H100) were often listed above $60/hr (~$7.50/GPU-hr) (^[12]) and Google’s A3 instances around $88/hr (~$11/GPU-hr) (^[13]). These figures match NVIDIA’s historical pattern of high launch prices, as seen with prior A100/A6000 tiers. Academia & industry narratives at the time noted that while H100 power justified its cost-performance, cost remained a major barrier for widespread use.

However, by 2025 the supply of H100s increased substantially (due to added production, alternate chips, and enterprise-focused deals). Cloud providers have also advanced specialized hardware (e.g. AWS Trainium, Google TPUs) which competes with H100. To maintain competitiveness, many began cutting prices. A notable turning point occurred in June 2025: AWS announced a ~44% price reduction on P5 instances (H100) across regions (^[10]) (^[14]). This brought AWS H100 GPU rental to roughly half its former rate. Similar cuts followed at other clouds. A DataCenterDynamics report confirms: “On-demand price … 44% for the H100 instance” (^[14]).

Beyond hyperscalers, many smaller GPU-cloud vendors instituted competition-driven price wars. Throughout 2025, daily tracking sites (e.g. GPUCompare) documented multiple providers slashing H100 rates (see Section on “Daily Pricing Updates” below). By late 2025, aggregated analyses expect H100 rental to settle around $2–4/GPU-hr in most markets (^[4]) (^[15]). For perspective, ThunderCompute’s September 2025 snapshot still showed Azure H100 at $6.98 and GCP at $11.06 (^[16]), but as of fall 2025 current rates are much lower (^[9]).

A key market consideration is price volatility and regional variation. Indicators include GPUCompare’s mid-2025 reports of providers continuously lowering H100 prices by up to 20–25% in single moves (^[17]) (^[18]). Reported insights note “continued price reductions across board” (^[19]) and H100 reaching “new lows” (^[20]). The H100 price volatility surpasses that of previous-gen GPUs, reflecting fierce competition and excess capacity. Regionally, even within the same cloud provider, pricing differs; for example, the U.S. West Coast regions are consistently ~10–30% above East US rates (^[21]). Users can often reduce costs by choosing slower regions or leveraging preemptible instances.

Looking forward, experts anticipate this downward trend to continue, albeit with fluctuations due to demand surges. The ThunderCompute analysis projects global GPU rental market growth from $3.34B (2023) to $33.91B by 2032 (^[4]), which should force pricing down over time. The imminent arrival of next-gen GPUs (e.g. NVIDIA H200) may further depress H100 prices as customers upgrade. On the other hand, sustained AI hype could keep demand high, partially offsetting price drops. For now, the evidence points to H100 rentals easing toward a competitive equilibrium of roughly $2–3/GPU-hr for on-demand usage (^[4]) (^[22]).

This article presents selected H100 pricing examples from large clouds and specialist providers. Values are shown only where a current first-party page provides an applicable hourly price. The entries use different purchase options, including prepaid capacity and Pods, so they are not an on-demand market comparison.

Note: The AWS entry is a prepaid Capacity Blocks reservation rate, not on-demand EC2 pricing. AWS currently lists reservation-component effective rates of $41.528 per P5.48xlarge hour ($5.191 per H100) in listed U.S. regions and $37.76 per hour ($4.720 per H100) in listed Asia-Pacific, European, and South American regions. These rates exclude an operating-system charge when applicable; Linux is listed at $0, while other supported operating systems have separate charges. Google Cloud, Azure, and Oracle require a region- and purchase-model-specific check for a like-for-like comparison.

Notes: The Runpod figures are displayed Pod rates for the named SKUs. Other specialist-cloud, marketplace, promotional, and historical offers are omitted because they are not a verified, like-for-like current comparison.

These figures make clear that many non-hyperscaler platforms are now cheaper per GPU-hour than the big three clouds. For instance, Lambda’s $2.99 (^[23]) and Runpod’s $1.99 (^[24]) compare favorably to half of AWS’s last MSRP. (AWS’s own bulk-reservation prices can approach $1.90–$2.10/GPU-hr (^[9]), but that requires 1–3 year commitments.) This limited table does not establish a current market-wide on-demand range or ranking: its entries use different regions, dates, service tiers, and purchase models.

AWS (Amazon EC2)

AWS offers H100 under its P5 instance family. The flagship is p5.48xlarge (192 vCPU, 8×H100 80GB). Prior to mid-2025, AWS’s on-demand price for p5.48xlarge in USD regions was about $60.54/hour total (^[25]) (i.e. $7.57/GPU-hr) – one of the highest H100 rental rates. However, on June 5, 2025 AWS announced an “up to 45% price reduction” for GPU instances, specifically 44% off H100 (P5) on-demand (^[14]). Post-cut, the p5.48xlarge on-demand went to roughly $33–$34/hr (~$4.1/GPU-hr). According to third-party trackers, current on-demand P5 pricing is about $3.90 per GPU-hour in key regions (^[9]).

For a current public AWS reference, the Capacity Blocks page lists effective reservation-component rates for P5 H100 capacity: $5.191 per H100 in specified U.S. regions and $4.720 per H100 in specified Asia-Pacific, European, and South American regions. These are prepaid Capacity Blocks reservation rates, not on-demand EC2 rates. AWS bills the operating system separately while instances run; Linux is listed at $0, but other supported operating systems add a charge (^[1]).

AWS Spot and Reservation Pricing

AWS Spot and Savings Plans can have different pricing and interruption or commitment terms from Capacity Blocks. Obtain the current rate for the exact P5 instance, region, operating system, and purchase option from AWS before budgeting; this article does not establish a universal AWS spot or committed-use H100 rate.

Microsoft Azure

Azure offers H100-capable VM families, but this article does not establish a current region-specific public hourly rate on the same purchase basis as the selected AWS, Google Cloud Spot, and Runpod examples. Use Azure’s pricing page or calculator with the intended region and VM configuration before making a comparison (^[26]).

Box-by-box, Azure’s ND and NC families have historically been priced above AWS, reflecting differing market positioning. Users report ~$7/GPU-hr as typical (^[16]). As with AWS, Azure offers 1–3 year Reservations for up to 40–50% discount, but Microsoft’s commitment terms are usually around 30% off (depending on region). Azure also has spot (low-priority) VMs for GPUs, though public data on fallback H100 spot rates is scarce as of Nov 2025.

Google Cloud Platform (GCP)

Google Cloud offers H100 GPUs in A3 VM configurations. Its pricing workflow requires a selection of configuration, region, and purchase option; this article does not preserve a verified current numerical rate for a matching A3 configuration. Do not compare a Google Cloud on-demand or Spot selection directly with AWS Capacity Blocks, marketplace offers, or other service tiers (^[2]; Google Cloud Spot VM pricing).

Oracle Cloud

Oracle offers H100 GPUs via bare-metal instances (the BM.GPU or BMH series). A standard BM.GPU.H100.8 configuration packs 8×H100 GPUs into one machine. Oracle lists BM.GPU.H100.8 as an eight-H100 Hopper shape. This article does not establish a current public price for a specified Oracle region and purchase model, so it does not normalize or rank an Oracle hourly rate.

Other large clouds (currently no official H100 on-demand from Alibaba Cloud or AWS competitor Alibaba has new GPU instances announced, but no public H100 figures as of late 2025). Similarly, IBM Cloud started offering H100 in late 2024 but pricing details are not widely published, so we omit them here.

Lambda, CoreWeave, and Paperspace

Lambda Labs, a popular AI cloud, rents H100 GPUs by the unit or in 8-GPU clusters. Lambda’s 8×H100 “Lambda Cloud” instances list at $2.99/GPU-hr (^[23]) for SXM (NVL3) usage. (Dividing $23.92/hr billed for the 8-GPU instance). Individual GPU rentals are available at similar per-GPU rates (i.e. $2.99 for a node counted per GPU). This is far below AWS on-demand.

CoreWeave also offers 8×H100 HGX nodes; their pricing is around $49.24/hr per node (^[27]), which normalizes to $6.16/GPU-hr. CoreWeave thus is pricier than Lambda, reflecting its niche HPC focus and InfiniBand support.

Paperspace’s “dedicated” H100 VM (1×80GB) is $5.95/GPU-hr (^[28]). This is a dedicated single GPU instance, competitive with Azure. Paperspace also provides discounts for multi-GPU bookings, but on-demand single price stays near $6.

Runpod, Vast.ai, TensorDock, HPC-AI, and Cudo

Runpod publishes separate Pods prices by GPU SKU and service option. Its displayed Pods rates are $2.89 per hour for H100 PCIe, $3.29 per hour for H100 SXM, and $3.19 per hour for H100 NVL. Prices and availability vary by SKU and service, so use the provider page for the current selection (^[3]).

Vast.ai is a GPU marketplace where independent hosts list idle GPUs. Vast’s lowest H100 listing has been as low as $1.87/GPU-hr (^[24]) (used NVIDIA 80GB PCIe GPUs). In July 2025, VAST.ai ran promotions cutting SXM H100 to $1.49 (^[18]). Vast.ai prices bounce with supply; it often presents the floor for what’s possible on the open market.

HPC-AI (a smaller cloud) advertises H100 at $1.99 (^[29]). TensorDock (Europe/US cloud) similarly offers H100 at $2.25 (^[29]) (for clients like FloydHub). Cudo Compute lists H100 at $2.45 in many clusters, and was reported to cut to $1.80/GPU-hr in June 2025 (^[17]). Other entrants like Hyperstack and Nebius have H100 pricing around $2.00–2.49 after cuts (^[17]).

Specialist-cloud, marketplace, spot, and committed-use offerings may have materially different prices and service characteristics. This limited comparison does not establish a market-wide discount, ranking, or sub-$2 current H100 rate; compare only like-for-like region, SKU, service tier, and purchase terms.

Spot and committed-use products have different availability, interruption, and commitment terms from on-demand and prepaid-capacity products. Check the provider’s current pricing page for the exact configuration, region, and terms before comparing options (^[30]).

The comparison uses only the published product pages for the entries in the table:

• Normalized Comparisons: When comparing per-GPU costs, larger instances distort price-per-GPU. Thus, most analyses (e.g. Thunder Compute) normalize to $/GPU-hr. Table 1 and 2 use this metric.

• Temporal Trending: GPUCompare’s daily updates document month-to-month declines. For example, in June–July 2025 multiple providers slashed H100 pricing by ~20–25% in short order (^[17]) (^[18]). By September, ThunderCompute notes general H100 availability around $2.85–3.50 (^[4]). The cloudfoundation blog reports AWS/Azure now at $3.9/$3.0 (^[9]), down from $7.57/$11.06 (^[16]). This chronological perspective (Jun 2025: cuts (^[17]) → Sep 2025: ~$3–4 (^[4]) → Nov 2025: similar) shows sustained downward pressure.

• Regional Variance: Anecdotal tracking indicates AWS West Coast regions ~15% higher than East (^[21]). These differences arise from data center costs and local supply. Users can often save by choosing an east/midwest zone. Azure also has up to ~30% spread between cheapest and priciest region (^[31]).

• Cost vs. Purchase: For context, the cost to buy an H100 is cited at $25–40K (^[7]). Renting at $3/GPU-hr full-time (24/7) costs ~$2,160/month, which would “pay off” the GPU in ~1–2 years. But owning incurs capital, power, facility costs (estimated $1000–2000/month for electricity alone (^[32])). The rental model avoids these and adds maintenance support, which explains why enterprises often prefer cloud rental for short-to-medium-term workloads.

• Cost-Performance: Analysts emphasize cost-per-computation, not just sticker price (^[33]). For instance, AWS’s price/performance improved over prior generations despite higher absolute cost (a theme in The NextPlatform’s GPU pricing articles (^[33])). Our focus here, however, is raw rental price; detailed cost-efficiency (FLOPS per dollar) is beyond this report’s scope.

• Citations and Method: Table values are drawn from official pricing pages or trusted compilations: e.g. AWS/Azure public calculators, Thunder Compute blog (Sep 2025 data) (^[34]), Go4Hosting analysis (^[22]) (^[9]), GPUCompare daily logs (^[35]) (^[17]), and vendor docs. Where providers only list multi-GPU rates, we divided by GPU count (noting that licensed per-instance often prices by node). We avoided any hidden fees or long-term commitments in these figures.

Startup GPU procurement: A CloudCombinator analysis of GenAI startups highlights the complexities of acquiring H100 capacity on AWS (^[36]) (^[37]). It notes that “popular GPU types (H100, P6-B200) may be out of stock” in busy regions (^[37]), forcing planners to use alternatives like GPU Capacity Reservations or AWS UltraClusters for guaranteed access (^[38]). This underscores that price alone isn’t enough; availability is also a factor. The blog recommends strategies combining spot buys, regional flexibility, and multi-cloud to manage costs—echoing our observation of price variation.

Cloud vs. On-Premises TCO: A Cudo Compute blog breaks down the total cost of owning an 8-H100 server. They estimate each H100 costs ~$30,971 part-price (^[39]), so an 8-card system ~$247,766 plus CPU ($25K) and extras (^[40]). They estimate a roughly $325,000–$425,000+ on-premises configuration (^[41]). In Cudo’s illustrated configuration, $19.60 per hour is the GPU component only; adding the listed CPU, memory, and storage brings the total to $22.68 per hour. At continuous use for a 30-day month, that total is $16,329.60 (22.68 × 24 × 30). Relative to the cited $325,000–$425,000+ estimate, the simple hardware-cost crossover is about 20–26 months before electricity, cooling, maintenance, financing, utilization, and configuration differences. It is an illustrative comparison, not a purchase recommendation. For many, annual or monthly GPU rental provides budget flexibility and eliminates maintenance overhead.

Regulatory procurement: Government and academic HPC centers often publish utilization cases. While specific H100 rental deals are proprietary, general budget reports (e.g. Argonne/Lawrence projects) note that hardware budgets hammered by rising chip prices make cloud an attractive option. For example, the Frontier supercomputer’s on-site GPU parts were amortized to ridiculously low hourly prices (Argonne essentially paid $0/GPU-hr after write-off (^[42])). This contrasts with cloud’s retail rates and hints why high-demand research might qualify for subsidized purchases instead.

The steep fall in H100 rental costs has several implications. For enterprises and researchers, AI training becomes more accessible: fine-tuning large models or running vision pipelines on H100 is now far cheaper than a year ago. This democratization fuels innovation but also squeezes margins for small GPU vendors. Market analysts observe “GPU rental market exploded” demand but concurrently pressure to “drive down costs” (^[4]). In effect, the anticipated GPU shortage of 2023–24 eased somewhat, as vendors responded by lowering prices and increasing supply chains (purchase from multiple chip fabs, interconnect tech improvements, etc).

For cloud providers, these price cuts reflect economies of scale and competitive necessity. AWS’s 45% cut was partly driven by customer pushback over scarcity pricing. Google and Microsoft risk losing GPU-hungry customers to smaller clouds, which offer more attractive hourly rates (albeit usually with less enterprise support). Conversely, providers expect to retain workloads by offering discounts for committed usage (1–3 years) – a reason AWS still advertises ~$1.90/GPU-hr savings plans (^[9]). Potentially, if H100 prices drop further, providers may shift emphasis to newer chips (like H200) as the premium offering.

Another aspect is inventory management. Some smaller clouds (e.g. Lambda Labs, Runpod) offering H100 at $2/GPU-hr risk hardware resale losses if H100 market prices fall below purchase cost. Indeed, Intel’s reports show per-GPU cost halving over generations (^[33]). If H100 commodity price falls, companies may offload old inventory cheap. Conversely, manufacturer reactions (e.g. Nvidia’s own pricing guidance) could support higher prices via limiting supply of discounts. But as of Nov 2025, H100 chips are no longer “limited edition” – plenty of refurbished or gamer-stock boards have appeared.

H200 is a memory-enhanced Hopper GPU, not a Blackwell product. NVIDIA said H200 systems would be available from global system manufacturers and cloud service providers beginning in the second quarter of 2024. NVIDIA introduced the Blackwell platform separately in March 2024; B200 is a Blackwell GPU. These product releases do not establish a universal future H100 rental price, so this report does not forecast a specific 2026 H100 rate.

Falling rent may hasten AI adoption in SMBs and academia. For startups and labs previously deterred by cost, H100 compute is reaching commodity price points. This could contribute to an AI boom or democratize research. However, some caution is warranted: extremely low prices often come with caveats (limited slots, older GPUs sold as H100, or multi-tenancy on slower networking). Additionally, as H100 rent flattens, differences between providers may hinge on value-adds (support, specialized interconnects, compliance). It’s also possible that in later 2026, if GPU supply outstrips demand, providers will deliberately throttle lower-end rates to protect profit margins – e.g. by promoting multi-year commitments or preemptible offerings over on-demand.

In summary, this article provides selected H100 pricing examples, not a definitive market-wide snapshot. Current comparisons must identify the provider, product, region, service tier, and purchase model; listed rates such as Google Cloud on-demand A3 High, AWS Capacity Blocks, and Runpod Secure Cloud are not interchangeable.

Key drivers have been broad market trends – surging supply, price wars, and alternative chips – all documented here with citations (^[14]) (^[17]) (^[4]) (^[22]). Users planning ML workloads should obtain a current quote or pricing-tool result for the exact GPU SKU, region, capacity type, and commitment term. The selected examples in this article do not support a universal H100 rate or a predicted spot or reserved rate.

For budgeting, obtain a current quote or use the provider’s pricing tool for the exact GPU SKU, region, capacity type, and commitment term. The selected examples are not a universal on-demand benchmark or coverage of the full provider market.