MARKET REPORT

The State of AI Inference — August 2026

Who is winning, where the market is moving, and how large the infrastructure gap has become.

Key findings
  • Nebius leads the global provider ranking at 92.7 under the General profile.
  • The tool-reliability spread on a single model (Qwen3-235B) reaches 4.5 points across 12 providers.
  • Only 3 of 13 ranked providers sit on the price–performance frontier — every other provider is dominated on both axes.

The global ranking

Aggregated across benchmarked models with coverage and freshness shading, Nebius holds #1 with 92.7, followed by Fireworks (91.4) and Groq (90.1). The top of the table is tight; the middle is not — the gap between #1 and the median ranked provider is 6.3 points.

Nebius92.7
Fireworks91.4
Groq90.1
Cerebras88.2
Together87.6
OpenRouter86.4
DeepInfra86.4
Scaleway84.4
Global provider ranking, General profile · DEV DATA

The frontier is narrow

Plotting score against output price, 3 providers are Pareto-dominant. The rest are dominated: for each of them there exists a provider that is simultaneously cheaper and better under this workload profile.

78838894$0.31$0.62$0.92$1.23Output price $/M →Score ↑Nebius
Score vs output price — Pareto set highlighted · DEV DATA

Same model, different result

The thesis holds at market scale: identical models diverge by whole reliability points depending on who serves them. Qwen3-235B's tool-call success ranges 94.7%–99.2% across 12 infrastructures.

Fireworks99.2OpenRouter99.0Cerebras98.8Nebius98.1Groq97.5Together96.9Nextbit96.7Scaleway96.6DeepInfra95.7Mistral95.7SiliconFlow95.2OVHcloud94.7
Tool-call success for Qwen3-235B by provider · DEV DATA
What this means for engineers
  • Model selection is only half the decision — provider selection moves reliability and latency by margins that dwarf most model-version differences.
  • Frontier position is workload-dependent: re-run the ranking under your own workload profile before committing.
Dataset

Published figures are pinned to this snapshot and are not recalculated retrospectively. View the current benchmark →

Cite this researchInferenceBench (2026). The State of AI Inference — August 2026. InferenceBench Research.
← All research
Related research