Market Signals

CTI components · updated with each pricing ingest · benchmark v5.0

The sections below are the underlying inputs to the 1337 Compute Tightness Index. Each signal is shown independently so the methodology is transparent before any composite score is published. Price signals measure market conditions in real time. The efficiency frontier shows where a buyer can get the most useful work per dollar across the full hardware and provider landscape. Supply breadth tracks how many providers are competing in each GPU family. The composite index will be introduced once these signals have accumulated sufficient history to weight with confidence.

2%
Tightest price spread · L40S
11,262,304
Best LLM tok/$ · B200
53
GPU x provider combinations tracked
·
Floor prices · 0 falling · 0 rising

Price Signals

Spread and velocity of on-demand floor pricing across providers. Narrow spread and rising floors signal a tightening market. Wide spread and falling floors signal supply at the margin.

Spread · floor to ceiling by GPU family

GPU Floor Ceiling Spread $ Spread % Providers
NVIDIA H100-SXM-80GB$2.15$12.29$10.14472%13
NVIDIA H200-SXM-141GB$2.29$10.84$8.55374%11
NVIDIA A100-SXM4-80GB$1.36$5.12$3.76277%11
NVIDIA B200-SXM-180GB$3.95$14.00$10.05254%7
AMD Instinct MI300X$1.99$6.00$4.01202%5
NVIDIA H200-NVL-141GB$4.29$10.60$6.31147%2
NVIDIA A100-SXM4-40GB$1.99$4.10$2.11106%4
NVIDIA H100-PCIe-80GB$1.68$2.00$0.3219%2
NVIDIA L40S$1.54$1.57$0.032%2

Velocity · 7-day floor change

GPU Current floor 7-day floor Reference date Change
AMD Instinct MI300X$1.99$1.992026-07-10· +0.0%
NVIDIA A100-SXM4-40GB$1.99$1.992026-07-10· +0.0%
NVIDIA A100-SXM4-80GB$1.36$1.362026-07-10· +0.0%
NVIDIA B200-SXM-180GB$3.95$3.952026-07-10· +0.0%
NVIDIA H100-PCIe-80GB$1.68$1.682026-07-10· +0.0%
NVIDIA H100-SXM-80GB$2.15$2.152026-07-10· +0.0%
NVIDIA H200-NVL-141GB$4.29$4.292026-07-10· +0.0%
NVIDIA H200-SXM-141GB$2.29$2.292026-07-10· +0.0%
NVIDIA L40S$1.54$1.542026-07-10· +0.0%
NVIDIA RTX PRO 6000$3.00$3.002026-07-10· +0.0%

Workload Efficiency Frontier

Best and median throughput-per-dollar per model across all GPU and provider combinations, at current on-demand floor pricing. Each row is one benchmark model so the comparison is apples-to-apples within a workload type. The ratio shows how close the median option is to the efficiency frontier. A low ratio means most buyers are significantly overpaying relative to the best available option for that specific task. Benchmark: MLPerf v5.0 · Offline scenario.

Workload Model Best Median Median / Best Options Frontier leader
LLMMixtral 8x7B11,598,269 tok/$4,317,935 tok/$37%24H200 · Lambda Labs
LLMLlama 2 70B11,262,304 tok/$3,277,170 tok/$29%66B200 · Nebius
VisionResNet-50150,963,210 samp/$68,183,044 samp/$45%26H200 · Lambda Labs
VisionRetinaNet3,042,858 samp/$1,332,660 samp/$44%26H100 · Nebius
VisionStable Diffusion XL3,481 samp/$1,293 samp/$37%24H100 · Nebius
NLPBERT15,253,116 samp/$6,549,161 samp/$43%26H100 · Nebius
Medical3D U-Net10,946 samp/$4,856 samp/$44%26H100 · Nebius
Rec.DLRM-v2126,022,605 samp/$46,626,466 samp/$37%24H100 · Nebius

Supply Breadth · Providers per GPU Family

Number of providers actively offering each GPU family on-demand, updated each ingest cycle. Providers missing for 3 or more consecutive days are excluded. Trend shows change vs. 30 days ago where data exists. More providers signals supply expansion and pricing competition. A declining count signals consolidation or inventory tightening.

GPU Active providers
NVIDIA H100-SXM-80GB
12flat vs 30d
NVIDIA H200-SXM-141GB
10+1 vs 30d
NVIDIA A100-SXM4-80GB
10flat vs 30d
NVIDIA B200-SXM-180GB
6flat vs 30d
NVIDIA A100-SXM4-40GB
4flat vs 30d
AMD Instinct MI300X
4flat vs 30d
NVIDIA L40S
2flat vs 30d
NVIDIA H200-NVL-141GB
2flat vs 30d
NVIDIA H100-PCIe-80GB
2flat vs 30d
NVIDIA RTX PRO 6000
1flat vs 30d

Tracking 13 providers with pricing data in the past 3 days: AWS, Azure, CoreWeave, Crusoe, DigitalOcean, GCP, Hyperstack, Lambda Labs, Latitude.sh, Nebius, OCI, RunPod, Vultr. Providers are added as they are discovered. Providers that stop returning pricing data for 3 or more days are excluded from the count and flagged in the ingest log for EOL review.