Silicon Analysts

GPU Rental Price per Effective FP8 PFLOP — Historical Time Series Data

Derived price-per-performance series for rented AI GPUs: public on-demand rental prices divided by NVIDIA peak FP8 (sparse) throughput — H100 SXM5 and H200 at 3,958 TFLOPS, B200 at 9,000 TFLOPS. Tracks whether new silicon generations are priced at a per-performance discount (the neocloud pattern) or at per-PFLOP parity/premium (the hyperscaler scarcity pattern). Companion dataset to the GPU rental premium decomposition analysis.

GPU Rental Price per Effective FP8 PFLOP

Derived price-per-performance series for rented AI GPUs: public on-demand rental prices divided by NVIDIA peak FP8 (sparse) throughput — H100 SXM5 and H200 at 3,958 TFLOPS, B200 at 9,000 TFLOPS. Tracks whether new silicon generations are priced at a per-performance discount (the neocloud pattern) or at per-PFLOP parity/premium (the hyperscaler scarcity pattern). Companion dataset to the GPU rental premium decomposition analysis.

9 data points·Unit: $/hr per PFLOP·Last updated: Mar 2026

Historical data blurred — unlock Pro for exact values across 2 data points

2 data point values hidden. Unlock exact values, sources, and exports with Pro.

Unlock Pro

Data Table

PeriodSeriesValue
Q4 2023
H100 (on-demand median)
•••
Q2 2024
H100 (on-demand median)
•••
Q4 2024
H100 (on-demand median)
1.2
Q1 2025
B200 (on-demand median)
0.72
Q1 2026
H100 (on-demand median)
1.56
Q1 2026
B200 (on-demand median)
0.96
Q2 2026
H200 (on-demand median)
1.3
Q2 2026
B200 (on-demand median)
0.78
Q2 2026
H100 (on-demand median)
1.01

Methodology & Sources

Derived series — provider public list prices ÷ NVIDIA peak FP8 sparse throughput (H100 SXM5 and H200: 3,958 TFLOPS; B200: 9,000 TFLOPS, per lib/constants/chipSpecs.ts). All values are Silicon Analysts estimates; providers do not publish per-performance pricing. Each point's source note names the underlying price observation.

Citations:

  • Silicon Analysts (derived) (Dec 2023)
  • Silicon Analysts (derived) (Jun 2024)
  • Silicon Analysts (derived) (Dec 2024)
  • Silicon Analysts (derived) (Mar 2025)
  • Silicon Analysts (derived) (Jan–Mar 2026)
  • Silicon Analysts (derived) (Apr 2026)

Related Tools

Related Analysis

Frequently Asked Questions

How much does it cost to rent a B200 per hour?
As of the April 2026 snapshot, Lambda lists the B200 on-demand at $5.50 per GPU-hour ($44.00 for an 8-GPU instance), while AWS capacity blocks price the B200 at $9.36 per GPU-hour ($74.88 for the 8-GPU p6.48xlarge). Per unit of FP8 (sparse) compute, Lambda’s B200 works out to roughly $0.61 per effective PFLOP-hour — about 39% below its own H100 — while AWS prices Blackwell at per-PFLOP parity with the H100.
What does "$/hr per PFLOP" mean for GPU rental pricing?
It is a derived price-per-performance metric: the public on-demand rental price per GPU-hour divided by the chip’s peak FP8 sparse throughput in PFLOPS (H100 SXM5 and H200: 3.958; B200: 9.0, per NVIDIA spec sheets — dense throughput is roughly half). Providers do not publish per-performance pricing, so all values are Silicon Analysts estimates; each data point cites the underlying price observation.
Is the B200 cheaper than the H100 per unit of compute?
On neoclouds, yes: Lambda rents the B200 at 1.38x its H100 price for 2.27x the FP8 (sparse) throughput — about 39% cheaper per effective PFLOP ($0.61 vs $1.01). On AWS capacity blocks the B200 is priced at per-PFLOP parity with the two-year-old H100 (~$1.04 vs ~$0.99), meaning the newest silicon carries zero performance discount — hyperscalers price scarcity, neoclouds price compute.