Pulse H100 SXM Hyperscaler OD — Methodology v1.0
Median on-demand list price per GPU-hour for H100 SXM across hyperscaler providers.
Permanence note. This URL is canonical for this methodology version. It will not be redirected, removed, or repurposed once published. Future revisions land at v1-1, v2-0, etc., with this page preserved unchanged.
Plain-language summary
Pulse H100 SXM Hyperscaler OD is a daily price assessment for the NVIDIA H100 SXM 80GB GPU on the hyperscaler segment of the cloud GPU market, on a on-demand basis. The published value is the median of normalized on-demand prices reported by hyperscaler providers in the contributing list below, expressed in USD per GPU-hour.
Formal definition
For each assessment date t, let pi,t denote the normalized on-demand price (USD per GPU-hour) reported by provider i in the hyperscaler NVIDIA H100 SXM 80GB contributing set on date t. The published value is
Index_t = median_i ( p_i,t )
over the contributing providers that meet the data-quality criteria for date t. Equal weight is applied to all contributing providers; no volume-weighting or market-share-weighting is used.
Source family
Hyperscaler — large-scale public cloud platforms with global infrastructure, enterprise SLAs, compliance certifications, integrated storage, networking, and ML services, and reserved or committed-use options. The hyperscaler universe is exhaustive at four providers: AWS, Azure, GCP, and OCI.
Source families are never blended into a single number. Hyperscaler pricing for a given GPU consistently runs several multiples above neocloud pricing for the same hardware. The spread reflects real differences in the product being purchased — infrastructure reliability, compliance, networking, support, ecosystem — not market inefficiency. Pulse follows established price-reporting-agency practice in publishing separate assessments per family.
Data sources
Contributing providers and the per-provider pricing model:
| Provider | Instance / shape | Pricing model | Region |
|---|---|---|---|
| AWS | p5.48xlarge | Per-instance hour, divided by 8 GPUs. | us-east-1 |
| Azure | ND96isr_H100_v5 | Per-VM hour, divided by 8 GPUs. | eastus |
| GCP | a3-megagpu-8g | Component pricing: standalone GPU rate plus VM CPU/RAM divided by 8 GPUs. | Americas aggregate (us-central1, us-east1) |
| OCI | BM.GPU.H100.8 | Per-GPU-hour native; no instance normalization. | Uniform global |
All providers are collected daily. The assessment anchor time is 18:00 UTC; data collected within the assessment window (typically 17:00–19:00 UTC) is attributed to the 18:00 UTC assessed_at timestamp for that date. Collection runs 365 days per year.
Filter rules
- Pricing type. Only on-demand prices contribute to this series. Reserved, committed-use, calendar-mode, and defined-duration pricing variants are excluded, as is spot pricing (tracked separately under its own series).
- GPU canonicalization. Each provider uses different identifiers for the same physical GPU. Pulse maintains a canonical GPU model registry that maps every provider-specific name to a single canonical model. Only prices that resolve to the canonical NVIDIA H100 SXM 80GB model contribute to this series.
- Region. US-region pricing for hyperscalers (per the regions in the contributing-provider table). Neocloud providers that publish a single global price are included without region filtering.
- SKU canonicalization. Hyperscaler pricing APIs return many SKU variants per instance type. Each collector filters to exactly one canonical on-demand SKU per instance type (Linux operating system, shared tenancy, on-demand market option, no Capacity Block reservation, no bundled software).
- Anomaly review. A provider price that deviates by more than 50% from the current series median is flagged for review. Flagged prices are included unless determined to be a data-quality issue (API error, stale data, misclassified SKU). Any exclusion is logged in the public corrections page.
Aggregation method
Headline statistic
Median (P50). The median is robust to outliers and avoids giving disproportionate weight to a single provider's pricing decision. It is also easier to defend as a neutral reference point than a trimmed mean or a single percentile.
Supporting statistics
Alongside the headline median, Pulse calculates and stores P25, P75, min, max, and the contributing provider count for each assessment. Individual provider assessments are stored with full provenance (raw collection ID, collector version, methodology version, normalization note).
Weighting
Equal weight across all contributing providers. No volume-weighting, market-share-weighting, or tiering is applied.
Worked example
For a published assessment date, the four hyperscaler prices are sorted and the median taken. For example: $6.88 (AWS), $10.00 (OCI), $10.98 (GCP), $12.29 (Azure). The median is the average of the two middle values (($10.00 + $10.98) / 2 = $10.49). Published value: $10.49 per GPU-hour.
Publishability thresholds
Publishable: 3 of 4 major hyperscalers (AWS, Azure, GCP, OCI) report a valid price. Ideal: 4 of 4 hyperscalers report. Unpublishable: fewer than 3.
Publishability is evaluated each assessment date. The series may publish on most days but be marked unpublishable on any date where contributing-provider count falls below the threshold; on such dates, no headline value is published, the underlying provider-level data remains stored, and the series resumes publication on the next date that meets the threshold.
Edge cases and known limitations
- Bundling. Hyperscaler instance prices include the cost of bundled vCPUs, RAM, local storage, and high-speed networking. Dividing instance price by GPU count produces a per-GPU-hour figure that includes these bundled components. This is intentional — it reflects what a buyer actually pays per GPU-hour — and is not decomposed into GPU-only and non-GPU components in the headline. Where a provider exposes component pricing (notably GCP), the all-in figure includes all components.
- SKU filtering. Hyperscaler pricing APIs return many SKU variants per instance type (operating system, tenancy, market option, capacity reservation status, bundled software). Each Pulse collector filters to exactly one canonical on-demand Linux SKU per instance type so that one instance produces one assessed price. Without these filters, non-GPU pricing components (Capacity Block reservation fees, OS license surcharges, dedicated tenancy premiums) would enter the assessment.
- OCI bare-metal only. OCI publishes flagship-GPU instances as bare-metal shapes (BM.GPU.*) under uniform global pricing. There is no per-region variation to filter.
- Carry-forward window. If a provider's API is unreachable on a given date, that provider's most recent valid assessment from the preceding 3 calendar days is used in the index calculation. The underlying data remains a truthful record of what was collected on each date — no synthetic rows are inserted into the assessment table — and carried-forward contributions are flagged in the published output via the
is_carried_forwardfield.
Source code
The implementing code lives in the public repository: https://github.com/pulsebenchmarks. Each tagged release of the repository corresponds to a methodology version; the v1.0 release implements this page.
Changelog
v1.0 — Initial published methodology. Defines source-family segmentation, headline median statistic, equal-weight aggregation, provider list, publishability thresholds, and the carry-forward policy described above.