Inference Performance Index: changing the way we measure inference capacity
Comments on “Inference Performance Index: changing the way we measure inference capacity”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Epoch AI introduced the Inference Performance Index, which scores chips by how many simultaneous agentic requests they can serve across 50–400 tokens/s/req, in GB300 equivalents. Its model puts GB300 at about 25× H100 capacity with 2.5× the FP8 FLOP/s, and H200 HGX at 0.12 and VR200 at 2.18. The scores are speed-of-light projections from four models and ignore software maturity. A validation on Kimi K3 matched GB300 measurements at 50% of modeled performance, within about 12%.
Why it ranked
No. 19 of 95 with a score of 0.66. The panel rated it highest on new information (0.91) and technical depth (0.86), lowest on significance (0.65). Placed No. 19, No. 21 by score alone: the fifth AI & machine learning story of the day (−0.040). The placement terms move stories; they are not part of the score.
- AI judgeJudge 2claude-haiku-5-5
- Substantive hardware inference benchmarking methodology with detailed modeling and explicit assumptions; relevant to AI infrastructure builders, though it is a single vendor-analysis report with modeled rather than measured figures.
- AI judgeJudge 3gpt-5.6-luna
- Introduces a detailed, methodology-backed inference-capacity metric spanning hardware, memory, networking, models, and serving requirements. It offers actionable comparative insight, though results remain modeled projections with assumptions and some vendor-adjacent benchmarking.
Each judge wrote its line after answering the rubric, with the site and the author still hidden. Every answer behind this is in the table under Score.
Score
0.66total under rubric v2.3
- Significance0.65
- Reader interest0.78
- Audience breadth0.65
- Practical relevance0.67
- Substantiated0.80
- Technical depth0.86
- New information0.91
- Counts against: Promotional0.13
- Counts against: Overclaim0.27
Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).
| Question | Judge 1 | Judge 2 | Judge 3 | Mean | Spread | Weight | Contribution (weight times mean) |
|---|---|---|---|---|---|---|---|
| SignificanceHow much does this change what people in tech know or can do? | 0.62 | 0.55 | 0.78 | 0.65 | 0.23 | +0.24 | +0.156 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.71 | 0.70 | 0.93 | 0.78 | 0.23 | +0.26 | +0.203 |
| Audience breadthHow many people who build technology does it concern? | 0.64 | 0.60 | 0.72 | 0.65 | 0.12 | +0.13 | +0.085 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.56 | 0.60 | 0.84 | 0.67 | 0.28 | +0.22 | +0.147 |
| SubstantiatedHow well is it backed? Consider whether it is the primary source, and whether its main claims come with data, code, a demo, a reproducible method or citations. | 0.86 | 0.65 | 0.88 | 0.80 | 0.23 | +0.05 | +0.040 |
| Technical depthHow technically substantive is it? | 0.89 | 0.80 | 0.90 | 0.86 | 0.10 | +0.05 | +0.043 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.86 | 0.90 | 0.96 | 0.91 | 0.10 | +0.05 | +0.045 |
| PromotionalIs it mainly marketing with little substance (a press release, ad or sponsored post with no real information)? A launch post that carries real news is not promotional. | 0.12 | 0.20 | 0.08 | 0.13 | 0.12 | −0.19 | −0.025 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.29 | 0.35 | 0.18 | 0.27 | 0.17 | −0.11 | −0.030 |
| Sum of contributions | +0.663 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.663 | ||||||
Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.
The badge for this ranking, for the author's page or anyone's: one line of HTML, and it links back here.
<a href="https://rubric.hrih.si/editions/2026-11-10/19/"><img src="https://rubric.hrih.si/badge/2026-11-10/19.svg" alt="Ranked #19 on Rubric, 10 Nov 2026" width="245" height="28"></a>Claim check
Claim check: 3 matches source · 3 no source given · 2 couldn’t check. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
No source givenNumber
Epoch estimates the Nvidia GB300 provides about 25× the inference capacity per chip of the Nvidia H100 on the agentic inference workloads it models, with only 2.5× the 8-bit FLOP per second.
The article gives no link or citation for it.
The article says “the Nvidia GB300 provides about 25× Nvidia H100’s inference capacity per chip on the agentic inference workloads we model, with only 2.5× the 8-bit FLOP per second”
The source says No source to compare with.
No source givenNumber
Nvidia GB300 NVL72 has about 1.9× the modeled inference capacity per GPU of Nvidia B300 HGX, because it connects groups of 72 GPUs through high-bandwidth links versus groups of eight.
The article gives no link or citation for it.
The article says “Nvidia GB300 NVL72 has about 1.9× the modeled inference capacity per GPU of Nvidia B300 HGX”
The source says No source to compare with.
Matches sourceNumber
In late September 2026, agentic API keys on OpenRouter consumed roughly seven times as many tokens as human API keys.
The cited source says the same thing.
The article says “In late September 2026, agentic API keys consumed roughly seven times as many tokens as human API keys”
The source says “In the last week of September, agentic API keys used about 7x the tokens from human API keys”
Source cited by the article: openrouter.ai/data/newsletter/cache-is-crucial
No source givenNumber
For Kimi K3, the area-under-curve capacity ratio of the Vera Rubin VR200 chip to GB300 is approximately 2.48.
The article gives no link or citation for it.
The article says “For Kimi K3, this ratio is approximately 2.48 on the Vera Rubin (VR200) chip”
The source says No source to compare with.
Couldn’t checkNumber
The IPI scores GB300 NVL72 at 1 GB300e, H200 HGX at 0.12, B300 HGX at 0.54, MI355X at about 0.55 and VR200 at 2.18, making GB300 NVL72 roughly eight times H200 HGX.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “GB300 NVL72 scores 1 GB300e, compared with 0.12 for H200 HGX, a roughly eightfold increase in modeled capacity per GPU”
The source says No passage could be compared.
Source cited by the article: epoch.ai/data/machine-learning-hardware-documentation/methodology
Couldn’t checkBenchmark result
InferenceX's Kimi K3 benchmarks show a similar ordering to the IPI, with Rubin and GB300 ahead of B300 and MI355X, and H200 far behind.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “show a similar ordering, with Rubin and GB300 ahead of B300 and MI355X, and H200 far behind”
The source says No passage could be compared.
Source cited by the article: inferencex.semianalysis.com/inference/kimi-k3
Matches sourceNumber
For Kimi K3 alone, Epoch's model predicts roughly a 46-fold GB300 advantage over H200.
The cited source says the same thing.
The article says “our model predicts roughly a 46-fold GB300 advantage over H200, compared with eightfold across the full basket”
The source says “H200 SXM · FP8 · 400G | 0.123 | 0.054 | 0.086 | 0.022 | 0.368 | 0.016 | 0.158”
Source cited by the article: github.com/jasonlizhengjian/inference-performance-index/blob/master/docs/research/methodology-sensitivity/README.md
Matches sourceBenchmark result
In Epoch's validation study, GB300's measured Kimi K3 throughput closely tracked predictions assuming 50% of modeled SOL performance, with an average absolute difference of about 12% across five frontier points.
The cited source says the same thing.
The article says “GB300’s measured Kimi K3 throughput closely tracked predictions assuming 50% of modeled SOL performance, with an average absolute difference of about 12% across five frontier points”
The source says “predicted output throughput differs from measured throughput by **11.6% on average (about 12%) at five frontier points in the 50–400 TPS/user band”
Source cited by the article: github.com/jasonlizhengjian/inference-performance-index/blob/master/docs/research/inferencex-validation/README.md
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN —
Recorded for context only, checked 10 Nov 2026, 05:47 UTC. It played no part in the ranking.