Editorial review · 260710-029
How FLUX’s piece on OpenAI's first chip is a Broadcom chip, and the interesting number isn't on the box scored.
Read the article →Solid reporting. Some issues but credible overall. The reader is well-served.
Accuracy
The article is admirably careful about what it does and does not know, flagging the performance-per-watt claim as marketing and citing Demler's qualifiers on the nine-month tape-out. The specific figures (Broadcom AI revenue, 7x share move, 10 GW target) are post-cutoff and attributed to named sources, so they sit in the post-cutoff-attributed bucket. One minor deduction for the unhedged assertion that Google's TPU programme began through Broadcom around 2016, which oversimplifies a more nuanced history.
Balance
The piece takes a clear structural view but represents the counter-case fairly, notably the Broadcom-dependency risk, the scope-versus-tooling ambiguity on the nine-month figure, and Nvidia's potential 2027 response. Loaded framing is minimal and the "selling exits from Nvidia" line is presented as analysis rather than smuggled as fact. Source diversity is thin (OpenAI, Broadcom-adjacent analyst, Seeking Alpha, Brockman's own tweet), with no Nvidia-side or independent semiconductor-analyst voice.
Concerns (3)
- minoraccuracy
“Google began its TPU programme through Broadcom around 2016”
Simplifies a more complicated design-partner history.
Evidence: TPU v1 was disclosed in 2016 but development began earlier; Broadcom's role varied by generation.
- minoraccuracy
“Broadcom AI revenue of $12.2 billion in fiscal 2024, guided to $20 billion-plus”
Post-cutoff, source attributed.
Evidence: Cited to Seeking Alpha September 2025; reviewer cannot verify against training data.
- minorbalance
“(source set)”
No Nvidia-side or independent semiconductor analyst voice.
Evidence: All cited voices are OpenAI, Broadcom-friendly, or an analyst note; topic admits wider views.
Reproducibility
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.