← Back to article

Editorial review · 260613-009

How ZEN’s piece on Agents per megawatt: the new unit inference is going to be priced in scored.

Read the article →
86/100
Solid

Solid reporting. Some issues but credible overall. The reader is well-served.

Accuracy 85
Balance 88

Accuracy

The specific figures (61,354 and 2,594 agents-per-megawatt) are attributed to a same-day Artificial Analysis launch and are post-cutoff but source-attributed, so no fabrication deduction applies. The piece hedges appropriately on independence and metric limits. Minor deduction for the unsourced claim that NVLink is 'roughly an order of magnitude faster' than InfiniBand/PCIe between boxes (-5), which is a specific comparative claim without citation.

Balance

The article foregrounds the commercial-independence concern, names AMD MI355X and Google TPU as needed counter-data, and explicitly flags that the benchmark favours hyperscaler-shaped deployments. Framing of the 20x figure is dissected rather than amplified. Source diversity leans on NVIDIA-aligned material, but the topic is a specific benchmark launch where that is reasonable; small deduction for not citing any independent analyst voice (-5).

Concerns (3)

Reproducibility

Run
13 Jun 2026, 18:23 BST
Reviewer
claude-opus-4-7
Prompt SHA
48c20c719fc8
Article SHA
d72a19ff3574
Editor
ZEN
Published
13 June 2026
Cost
$0.0000

How this review works: read the methodology. Each published Dispatch is scored by a single primary reviewer (Claude Opus 4.7) against the public rubric. A second model (Gemini 2.5 Pro with Google Search) runs the same prompt as a variance signal and is shown above only when the two scores diverge by more than ten points.