Data methodology

What the dataset contains

This dataset uses Verda’s official GPU pricing page. The check date shown in the Pricing Lab is the most recent successful automated extraction. We extract three single-GPU instance configurations, their on-demand bases and the self-service commitment discount schedule. The curve represents one supplier’s published tariff structure, not a global benchmark or transaction index.

Published inputs and calculated outputs

The source publishes base rates and percentage discounts. TensorCurve applies the formula base hourly rate × (1 − discount). The resulting term rates are labeled “Calculated.” They are not individually quoted prices, interpolations or forecasts. Four decimal places preserve the arithmetic; no rounding is applied before modeled totals are calculated.

Hardware memory, CPU and RAM fields are retained in the JSON snapshot and CSV download. A GPU instance, an instant cluster, a serverless container and a spot instance are distinct products; this dataset uses only the first category.

Time, geography and availability

The check date records when the public page was reviewed. It is not a supplier guarantee of price validity or availability. The extracted table does not attach a region or guaranteed start date to these rates. The “Public tariff” row therefore must not be read as a confirmed start-today offer. Future-start cells remain missing. Missing does not mean zero, sold out or unsupported.

Chart conventions

The homepage gives equal horizontal space to the four commitment categories. Its vertical axis is labeled and does not start at zero. Connecting lines are visual guides, not prices for unlisted durations. The lab plots actual month distances. Neither chart infers a forward-start premium. The two-year discount is outside the current four-term display, not an unavailable supplier term.

Planning totals

The lab models one GPU for 720 hours per month: exact calculated hourly rate × 720 × term months. This is a normalized planning amount, not an invoice or a verified contract obligation. Calendar billing, payment schedule, eligibility, cancellation, taxes, storage, networking and other charges must be confirmed separately. The cost guide states its full-window payment assumption explicitly.

Collection and updates

An automated job reads the public pricing page once a day, extracts the single-GPU instance table and the commitment discount schedule, and publishes a new snapshot only when the extracted values pass validation (all tracked configurations present, prices inside sanity bounds, discounts increasing with term, no implausible day-over-day jump). The site fetches that snapshot and caches it for up to six hours. If the source page changes shape or a fetch fails, the previous snapshot stays in place with its original check date; no timestamp is advanced to imply a successful refresh. The machine-readable base price is used for calculations; the rounded figure shown in the supplier’s table is recorded alongside it. The source URL, check date, original base, discount and calculation remain inspectable, and every daily observation is kept in a history file in the public repository.

Comparison boundaries

Other supplier prices appear in the buyer guide with their own configuration notes. They are not pooled into this curve. No availability check, booking or supplier inquiry is performed by the site.