AI COMPUTE PRICING

GPU Spot Pricing Intelligence

Hourly GPU rental prices across every major cloud, normalized to canonical SKUs. A leading indicator of AI capex before it reaches earnings.

2 YEARSParquet · JSON · CSVHourly snapshots, daily feature build
6,216
Price observations
38
Providers
18-100x
Price dispersion
not_ready
Honest equity verdict
01Sample

See the data

Representative records in the exact shape we deliver. Real provenance and full slices are shared under license.

Datacenter SKU, on-demand

Representative, not a real quote. Schema is reconstructed from the variables the research paper measures; there is no published schema file. Price is per GPU-hour; the panel keeps price > 0.

gpu_prices.jsonlrepresentative
{
  "observed_at": "2026-05-14T08:00:00Z",
  "provider": "provider-a",
  "canonical_sku": "H100_SXM_80GB",
  "raw_instance_name": "8x H100 80GB SXM5",
  "region": "us-east",
  "price_type": "on_demand",
  "price_per_gpu_hour_usd": 2.05,
  "gpu_count": 8,
  "capacity_available": 12,
  "source": "live_poll"
}

Consumer SKU, marketplace/spot

Representative. Consumer RTX class is the long-continuous-history segment of the panel.

gpu_prices.jsonlrepresentative
{
  "observed_at": "2026-05-14T08:05:00Z",
  "provider": "provider-b",
  "canonical_sku": "RTX_4090_24GB",
  "raw_instance_name": "RTX 4090 24G",
  "region": "eu-central",
  "price_type": "spot",
  "price_per_gpu_hour_usd": 0.31,
  "gpu_count": 1,
  "source": "archived_snapshot"
}
02Schema

Record shape

Every field, its type, whether it can be null, and a representative value.

FieldTypeConstraintDescription
observed_attimestamp · UTCrequiredWhen the price was polled or snapshotted.
e.g. 2026-05-14T08:00:00Z
providerstringrequiredCloud or marketplace provider the quote came from.
e.g. provider-a
canonical_skustringrequiredNormalized GPU identifier after collapsing marketplace spellings.
e.g. H100_SXM_80GB
raw_instance_namestringnullableOriginal provider listing string before normalization.
e.g. 8x H100 80GB SXM5
regionstringnullableProvider region or zone for the quote.
e.g. us-east
price_typestringrequiredon_demand or spot/marketplace term.
e.g. on_demand
price_per_gpu_hour_usdfloat · USD/GPU-hourrequiredNormalized rental price per single GPU-hour. The panel keeps price > 0.
e.g. 2.05
gpu_countint · GPUsnullableGPUs in the underlying instance, used to normalize to per-GPU.
e.g. 8
capacity_availableint · instancesnullableCurrently available, unrented capacity for the SKU.
e.g. 12
sourcestringnullablelive_poll or archived_snapshot (the panel backfills from archives).
e.g. live_poll
03What's included

Cross-Provider Price Surface

Normalized hourly spot and on-demand rates per SKU per region across 25+ providers. One schema, every cloud.

Capacity & Utilization

Provider capacity snapshots - total, available, and rented instances per SKU - a direct read on supply tightness.

Equity & Token Overlay

GPU-hour pricing joined to AI-infrastructure equities, options implied volatility, and DePIN compute tokens.

04Methodology

How it is built

  1. 01

    Collection

    Poll each provider public pricing surface, hourly where available, and backfill from archived snapshots.

  2. 02

    SKU normalization

    Map heterogeneous instance names to a canonical SKU set so prices are comparable across providers - the many marketplace spellings of an 80GB SXM H100 collapse to one identifier.

  3. 03

    Per-hour price basis

    Normalize every quote to USD per GPU-hour and filter the panel to price greater than 0.

  4. 04

    Fixed-set point-in-time reading

    Read trends within a fixed provider set. The panel composition changes over time, so a recent median up-tick can be an artifact of new, pricier providers entering rather than a real price rise; this is treated explicitly.

  5. 05

    Coverage-depth handling

    Consumer GPUs have the longest continuous history; datacenter SKUs are broad in the current cross-section but thin historically. The asymmetry is handled explicitly throughout.

05Evals

How we validate

What each evaluation measures and how it is run. Where no benchmark is published, we show the methodology and say so.

GPU-price to AI-equity signal

Measures

Whether the cross-provider price and dispersion panel predicts the AI-infrastructure equity complex.

Method

Build price and dispersion features; train walk-forward models against forward equity returns; validate with purged k-fold cross-validation and embargo, tested against a label-shuffled null and a naive-momentum baseline.

Result

Honest published null. Under naive evaluation the backtest looked strong; under rigorous evaluation the edge disappears - shuffle-test p-value near 1.0. Verdict: not_ready. We would rather report a null than sell false alpha.

Cross-provider price dispersion

Measures

How widely the same accelerator is priced across the market.

Method

Take the latest price per provider per SKU for SKUs listed by at least three providers; compute the priciest-to-cheapest ratio and coefficient of variation.

Result

Descriptive and real: order-of-magnitude dispersion - 18.3x for H100 SXM, 96.5x for A100 SXM, 100x for RTX-4090, with coefficients of variation between roughly 0.5 and 2.6.

06Graders

Ground truth

What correct means for this data, and how it is established.

Ground truth

Forward realized equity returns of the AI-infrastructure complex for the predictive test; the observed cross-provider quotes themselves for the descriptive layer.

How it is established

Walk-forward modeling of price and dispersion features against forward returns, scored under purged k-fold cross-validation with embargo and benchmarked against a label-shuffled null and a naive-momentum baseline.

Agreement

Not a human-rater task. Significance is measured against the null, and the signal fails it (shuffle-test p near 1.0).

07Application

AI Capex Leading Indicator

Spot-price moves and capacity exhaustion across providers front-run datacenter buildouts and chip demand, visible before guidance or earnings.

Supply-Chain Positioning

Map compute scarcity to the public AI supply chain - GPUs, networking, power, memory - without waiting on quarterly disclosures.

Compute Cost Benchmarking

Track the true market price of every GPU class over time. Benchmark internal compute spend or model the cost curve of a training run.

08Environment & integration

How you load it

Delivery

S3, REST API, Parquet

Formats

Parquet, JSON, CSV

Auth

Licensed for internal research and model development; redistribution requires a separate agreement. Sourced from public pricing surfaces; no PII or MNPI.

Cadence

Hourly snapshots with a daily feature build.

Request access.

Restricted-scope evaluation access for qualified teams. We share real samples, full schema, and provenance under a mutual NDA.