GPU and neocloud pricing
What GPU capacity costs to rent, as each provider publishes it, and what GPU systems cost to own.
Coverage
GPU classes from the A100 to rack-scale NVL72 systems, priced by hyperscalers and neoclouds.
GPU classes
| GPU class | Maker | Architecture |
|---|---|---|
| H100 | NVIDIA | Hopper |
| H200 | NVIDIA | Hopper |
| B200 | NVIDIA | Blackwell |
| B300 | NVIDIA | Blackwell Ultra |
| GB200 NVL72 | NVIDIA | Grace Blackwell, rack scale |
| GB300 NVL72 | NVIDIA | Grace Blackwell Ultra, rack scale |
| MI300X | AMD | Instinct, CDNA 3 |
| MI325X | AMD | Instinct, CDNA 3 |
| MI355X | AMD | Instinct, CDNA 4 |
| L40S | NVIDIA | Ada Lovelace |
| A100 | NVIDIA | Ampere |
What each price is recorded against
- Provider segmentHyperscalers, the large general-purpose clouds, and neoclouds, the clouds built around GPU capacity. The two are kept separate.
- Pricing typeOn-demand, spot or preemptible, and reserved.
- RegionEach provider's published regions, mapped to one region list.
- TermHourly on-demand rates and reserved terms in months, as each provider states them.
- ConfigurationGPU variant, interconnect and GPUs per node, recorded with every price.
For what GPU systems cost to own, see GPU systems pricing.
How to compare GPU cloud prices
Two prices compare fairly only on one basis. Each price describes one configuration and term at the moment it is captured.
The price as published stays on record beside its normalized form, so each adjustment can be checked.
- Per GPU-hour
- Providers price an instance or a whole term. Bring each price to one GPU for one hour, in U.S. dollars.
- GPU variant
- One GPU class can be offered in more than one form factor and memory size. Compare the same variant.
- Interconnect and node size
- Compare offers with the same GPU interconnect and the same number of GPUs per node.
- Pricing type
- On-demand, spot or preemptible, and reserved prices are different offers. Compare each with its own type.
- Term
- Reserved prices hold for a committed term in months. Compare them term for term.
- Region
- Providers name their own regions. Map each to one region list, then compare within a region.
- Capture time
- A price describes its offer at the moment it is captured. Compare prices from the same point in time.
Neoclouds and hyperscalers
Every provider in the data sits in one of two segments.
To set a neocloud price beside a hyperscaler price, put both on one basis first.
- NeocloudsCloud providers built around GPU capacity for superintelligence (SI).
- HyperscalersLarge general-purpose cloud providers.
- One segment per sourceEach provider source carries one segment, recorded with every observation.
- Shown separatelyHyperscaler and neocloud prices are kept apart, and the Rillor Compute Index shows the two segments separately in every series that uses them.
Observation schema
One record per offer per capture. The price as published is kept beside its normalized form.
| Field | Type | Meaning |
|---|---|---|
| observation_id | string | One identifier per offer per capture |
| captured_at | UTC timestamp | When the price was captured |
| source | enum: pricing_api, published_page | Where the observation came from |
| raw_snapshot_id | string | The stored raw capture the observation was read from |
| provider_segment | enum: hyperscaler, neocloud | Provider type |
| Field | Type | Meaning |
|---|---|---|
| provider_id | string | Provider, from the Rillor source register |
| region | string | Provider region, mapped to the Rillor region list |
| gpu_class | enum: H100, H200, B200, B300, ... | One of the tracked GPU classes |
| gpu_variant | string | Form factor and memory variant as the provider lists them |
| interconnect | string | GPU interconnect as the provider lists it |
| gpus_per_node | integer | GPUs in the priced instance or node |
| pricing_type | enum: on_demand, spot, reserved | Spot includes preemptible offers |
| term_months | integer | Commitment term. Zero for on-demand and spot. |
| usd_per_gpu_hour | decimal | Price for one GPU for one hour, in U.S. dollars |
| price_as_published | decimal, currency, unit | The price exactly as the provider states it |
| availability_as_published | string | The provider's own availability field, unchanged |
| reported_available | boolean | Whether the provider reports the offer as available to rent |
How it is captured
From the provider's published price to a record you can audit back to the original response.
- Sources
- Official provider pricing APIs and published pricing pages, used within each source's terms.
- Timing
- Each observation records its capture time as a UTC timestamp.
- Raw snapshots
- Each raw response is stored unchanged and linked to the observations read from it, so any value can be traced to its source.
- Normalization
- Prices are converted to U.S. dollars per GPU-hour. Per-instance and whole-term prices are brought to one GPU for one hour, and each offer is tagged with GPU variant, interconnect, node size, region and term.
- Corrections
- A corrected observation is stored as a new version, and the original stays on record.
What availability means
Availability means what a provider reports as available to rent.
We keep the provider's own field exactly as published, such as an in-stock flag or a stock level, and map it to one yes-or-no value for comparison. It is recorded at the moment of capture, beside the price, and describes rentable offers at that moment.
See RCI AvailabilityOwn
GPU systems pricing
What GPU systems cost to own. Specifications and price observations for NVIDIA and AMD systems, each typed by the kind of evidence behind it.
- What it holds
- System specifications, normalized configurations and price observations for HGX, NVL72 and Instinct systems.
- Sources
- Manufacturer specifications and system catalogs. Distributor, reseller and broker listings and quotes, where collection and use rights permit. Reported secondary sales with recorded provenance.
- Configurations
- Modeled by GPU family, GPU count, memory, CPU, interconnect, cooling, condition, location and seller type.
- Refresh
- Set per source. Every observation keeps its capture time and source.
What a GPU system price includes
A system price covers the whole build. Two eight-GPU servers with the same GPU board can differ in CPUs, host memory, network cards, storage, condition and support.
- GPUsThe GPU family and the number of GPUs.
- CPUs and memoryThe host CPUs and system memory.
- NetworkingThe network cards, and the interconnect between GPUs.
- StorageLocal storage for datasets and checkpoints.
- CoolingAir or liquid, each a separate configuration.
System types
- HGXNVIDIA HGX servers with eight GPUs joined by NVLink, such as the HGX H100, H200, B200 and B300. Air-cooled or liquid-cooled.
- NVL72NVIDIA GB200 NVL72 and GB300 NVL72 racks: 72 Blackwell or Blackwell Ultra GPUs and 36 Grace CPUs, with NVLink across all 72 GPUs. Liquid-cooled.
- InstinctAMD Instinct systems, such as the MI355X system of eight GPUs joined by AMD Infinity Fabric.
Each configuration also records the system's condition, such as new, refurbished or used, and the seller type, such as distributor, reseller or broker. Two systems that differ only in condition or seller are different configurations.
Evidence classes
| Evidence class | What the record asserts |
|---|---|
| Specification | System and component attributes as the manufacturer publishes them. |
| Listing | A public offer, seen at a stated time. |
| Quote | A price supplied on request for a stated configuration and quantity. |
| Indicative price | A reported price level without a verifiable counterparty. |
| Reported sale | A sale a source reports, without independent confirmation. |
| Verified transaction | A completed sale with confirming evidence. |
Each observation keeps its evidence class, so a listing is read as a listing and a quote as a quote.
How evidence classes workWhat it is used for
Decisions about renting, buying and financing GPU capacity.
Capacity planning
Compare what each GPU class costs to rent by region, term and pricing type before committing to capacity.
Procurement
Check a cloud quote against what providers publish, and a system quote against listings, quotes and reported sales for the same configuration.
Financing diligence
Rental and system prices for the GPU class behind a loan or lease, as context for collateral and residual value.
GPU financingThe Rillor Compute Index
These observations are the input to the RCI Rental, RCI Availability and RCI Hardware series.
Rillor Compute Index
Common questions
What is a GPU-hour?
One GPU used for one hour. Providers price an instance or a whole term, so each price is brought to U.S. dollars per GPU-hour before prices are compared. The price exactly as the provider states it stays on record beside it.
What are on-demand, spot and reserved prices?
On-demand is an hourly price with no commitment. Spot, which includes preemptible offers, is capacity the provider can reclaim at short notice. Reserved is a price for a committed term, recorded with its length in months.
Why do H100 prices differ between providers?
Each price describes one offer: a GPU variant, an interconnect and node size, a region, a pricing type and a term, at the moment it is captured. Two H100 prices can differ on any of these, and hyperscalers and neoclouds are kept in separate segments. Put both prices on one basis before comparing them.
Research
The newest articles from our Insights library, each with a full research report and its sources.
- From GPU-hour price to cost per training tokenHow to calculate LLM training cost from a GPU-hour price: compute per token, dense peak FLOPS, utilization in published runs and a worked example with dated public prices.GPU and neocloud pricing7 min readPDF report
- How to compare GPU system prices across configurationsTwo GPU servers rarely match part for part. Comparing GPU server prices across configurations takes matched builds, line-item adjustment, price per unit of performance or hedonic regression.GPU systems pricing7 min readPDF report
- How memory prices reach the cost of a GPU serverDated listings, filings and contract price forecasts show how memory costs reach a GPU server's price, and what lies behind the reported NVIDIA server price increase.GPU systems pricing7 min readPDF report
Notices
Rillor Compute Index. Rillor Compute Index series are for research and planning use. They are not offered for settlement of any contract.
Data rights. Source data is provided within the rights of each source.
Trademarks. NVIDIA, AMD, Ceph, Kubernetes and related product names are trademarks of their owners and are used only to identify products. No affiliation or endorsement is implied.