There Is No GPU Price

The dataset’s providers resolve into three pricing regimes. Hyperscalers publish list prices that did not change during the window. Marketplaces post administered tier prices that change by decision. Offer venues reprice continuously. The same H100 SXM5 80GB SKU:

The frozen list columns are not a data error — they are the finding. Across the window and 432 daily observations per series, AWS and Google Cloud each posted exactly one H100 price: the observed list price did not change during the collection period. Enterprises paying those rates are buying procurement compliance, interconnect, reliability, and ecosystem; the accelerator itself was offered at up to ~10x less on a marketplace, and the offer column’s volatility shows what a continuously repriced supply-side signal looks like next to a static one.

Normalizing every venue’s May–June median offers to dense FP16 throughput collapses eighteen GPU models and nine venues onto one axis — the 12.9x spectrum from the top of this page — and the ranking inverts the intuition that newer datacenter silicon is cheaper compute:

Three structural facts fall out. First, the compute-cost frontier is a consumer card on a marketplace: the RTX 4090’s $1.02/PFLOP-hr undercuts every datacenter option, and the discount is rational — 24GB of VRAM and no NVLink make it useless for frontier training, so the market prices the constraint, not the FLOPs. Second, newer datacenter silicon is not cheaper compute at observed prices: H100 at its cheapest venue (~$2.90–$2.97) costs more per dense PFLOP than the lower-priced A40 and L40S alternatives (~$2.00) — Hopper buyers pay for density, interconnect, and memory bandwidth, not FLOP efficiency. Third, hyperscaler list pricing preserved 2021-era economics through the window: Azure’s A100 at $11.77/PFLOP-hr is five-year-old silicon at a list price that did not move in our observation period — more per FLOP than any marketplace-offered GPU in the dataset by 3x–11x.

A distinct-price-values test separates the marketplaces cleanly: Vast.ai’s offers show near-continuous variation (independent hosts repricing), while RunPod’s 877 RTX 4090 observations contain exactly three distinct prices — posted tiers, administered by decision. Over the window, the two regimes moved in opposite directions on the same card:

The tier record shows the administered side explicitly: RunPod introduced a higher RTX 4090 tier ($0.69/hr) on April 17, 2026 and a higher H100 PCIe tier ($2.89/hr, +21% over the prior top) on May 20, 2026, while trimming L40S and A40 tiers slightly. RunPod’s observed median rose as higher-priced tiers entered the available mix, while Vast.ai’s median observed offers on the same card declined roughly 58% from their November level. Two marketplaces selling the identical card diverged from near-parity to a 3x spread — a spread that persists because the venues sell different bundles: RunPod sells reliability tiers and orchestration; the offer venue sells raw repriced compute.

Even as a customer-price equivalent, the asymmetry is structural: the recovery arithmetic at hyperscaler price levels runs roughly 5x faster than at marketplace offer levels for comparable hardware classes — before the deductions that fall hardest on exactly the venues with the slowest arithmetic. It is consistent with the supply behavior visible in Exhibit 3: hosts still listing consumer cards at $0.169/hr are plausibly running sunk hardware on cheap power, which is how offer medians can fall 58% while posted venues introduce higher tiers into the same market. A full host-economics model — payout shares, system capex, colocation, financing — is the natural sequel to this analysis, and this exhibit is its input, not its answer.

Frequently asked questions