// methodology · verified 27 Sep 2026

How we verify

Every figure on the Builder's Radar traces back to its primary source. No source, no number — no estimated numbers, no filler, no theater.

The rule

The Radar is a command board, not a feed. Before publishing a number we find the primary source (public API, official report, earnings call), store its URL and the date we manually verified it. No source, no number — even if the number is viral. On 27-Sep-2026 we discarded two adoption figures (75.6% and 93.6%) for exactly that reason.

live vs curated

The system is hybrid: ~70% refreshes itself from public APIs, ~30% needs human judgment. Badges say which world each number comes from:

  • ● liverefreshed automatically — by the source page itself or by our hourly cron — and stamped with its last fetch time.
  • ○ curatedpoint-in-time document (survey/report), manually verified on 27 Sep 2026.

They are never mixed without warning. If an API fails we fall back to the curated figure — never to fabricated values.

Every source

The exact list the radar's “Data & Sources” panel opens, rendered from code (single source of truth, always in sync):

Every figure on the board

The 7 comparative metrics, with their labels exactly as shown on the radar and the sources backing each:

Global inference share · 28d tokens

USA · 35.5%

US models · OpenRouter 28d

CHN · 63.5%

CN models · MiMo-V2.5 #1 (10.5T/wk)

Frontier training run · public cost

USA · > $100M

US runs · Grok 4 ≈ $388M (Epoch)

CHN · $3.47M

MiMo-V2.6 · RL phase (live log)

Flagship API cost · $/1M input tokens

USA · $10.00

GPT-6 Astra

CHN · $1.32

DeepSeek V4 Pro

Value-tier API cost · $/1M input tokens

USA · $0.75

Gemini 3.8 Flash

CHN · $0.14

MiMo-V2.6-Flash (Xiaomi)

Open-weight flagships · tracked

USA · 0 of 3

GPT-6 · Claude · Gemini closed

CHN · 3 of 3

DeepSeek · Qwen · GLM open

Max flagship context window

USA · 1.05M

GPT-6 Astra

CHN · 1.0M

DeepSeek V4 Pro

Consumer AI adoption · national stats

USA · 57.9%

adults 18–64 · US RPS survey (Feb 2026)

CHN · 42.8%

share of population · CNNIC (Dec 2025)

Adoption: explicit denominators

The two figures are NOT directly comparable and we say so on purpose: the US surveys working-age adults, China counts the whole country.

regionfiguredenominatorsource (date)
CHINA42.8%total population · 602M users, +141.7% YoY↗ CNNIC 57th Statistical Report · Dec 2025

Discarded figures: 75.6% / 93.6% — went viral with no primary source findable (searched 27-Sep-2026). No source, no number.

Verified users (globe rings)

A hub only draws its MAU ring when an officially published figure exists. No official figure → no ring. Never estimated:

San Francisco — OpenAI · AnthropicChatGPT · 1B weekly actives↗ OpenAI — 1B weekly actives · Aug 2026
Mountain View — Google DeepMindGemini app · 950M MAU↗ Alphabet — Q2 2026 earnings · Jul 2026
Hangzhou — Alibaba Qwen · DeepSeekQwen 167M · DeepSeek 130M MAU↗ QuestMobile H1 2026 · Jul 2026

What stays human

The curator's value lives in the data companies don't publish and in the strategic reading. This is NEVER automated: training costs (almost never in APIs), open vs closed philosophy, “usefulness” context of a model, and the insights that give the numbers meaning. If a figure has no public source, it's written by hand or it doesn't appear.

Region classification

Region follows the real organization, not the repo's language: meta-llama = USA, mistralai = EU (neither USA nor CHINA → OTHER; sides are never forced) and Chinese labs are verified orgs (deepseek-ai, QwenLM, XiaomiMiMo, minimaxai…).