Status: built 2026-09-19, all three phases. Code: index.html (markup, charts) and site/market-ui.js (rail, tables, detail, pins, URL state).
Objective: restructure the market tab around the inference supply chain (apps, routers, models, providers, hardware, plus public and private) so a reader can find a layer, read who holds the flow there, and compare its members, without reading prose.
Three reviews: nine crypto market-structure sites, ten AI and internet measurement sites, and the UX literature. Claims marked (2nd) come from search results or docs because the site blocked fetching.
| # | Pattern | Seen on | Use here |
|---|---|---|---|
| 1 | Flow overview as the spine, then one view per layer | orderflow.art, Cloudflare Radar (2nd) | Overview, then sections 1 to 6 |
| 2 | One template reused for every entity type | Artificial Analysis, L2BEAT, Radar (2nd) | Identical section anatomy |
| 3 | One wide sortable table per layer, grouped column headers, fixed column grammar | L2BEAT, Artificial Analysis, DefiLlama (2nd) | The side-by-side table |
| 4 | Data quality is a column and a grouping, never a footnote | L2BEAT stages, Epoch "administered by us" | "Publishes" column, source badge |
| 5 | Unit restated in every subtitle; shares-only sources say so | OpenRouter rankings, Vercel leaderboards | Unit in every header |
| 6 | Source line under every figure: source, unit, window, updated | Our World in Data, mevboost.pics, relayscan | Section footer line |
| 7 | Unknown flow stays visible | orderflow.art "Unknown", relayscan caveats | "Not in the top 20 apps" node |
| 8 | Methodology in tiers: strip, per-metric note, one full page | growthepie, Artificial Analysis, orderflow.art | Badge, Method toggle, docs page |
| 9 | Scatter with a Pareto line for price against speed | Artificial Analysis | Model detail (the price board) |
| 10 | Compare picker capped small, state in the URL | L2BEAT compare, Rated (2nd) | Pin up to 4 rows (phase 2) |
What to avoid: the 13-section single scroll (Artificial Analysis, OpenRouter rankings), rankings with no entity comparison (OpenRouter), credits above the hero (orderflow.art), sidebar sprawl (DefiLlama, 2nd).
| Rule | Source |
|---|---|
| Overview first, zoom and filter, details on demand | Shneiderman 1996 |
| At most two disclosure levels | NN/g, Progressive Disclosure |
| No tabs between things readers compare; one page with in-page navigation | NN/g, Tabs Used Right; In-Page Links |
| 57% of viewing time is above the fold | NN/g, Scrolling and Attention |
| Same anatomy everywhere ("meaningless variety" is pitfall 6); a number needs "compared to what" | Few, Common Pitfalls in Dashboard Design |
| Side-by-side compare holds 5 items at most; drop attributes most items lack | NN/g, Comparison Tables |
| Frozen first column, right-aligned figures, unit in the header, sorted by value | NN/g Data Tables; UK Analysis Function |
| Bars in one column only; sparklines on a common scale when rows are compared | Datawrapper; Few |
| Never a blank cell or a false zero: [x] not published, [z] not applicable, [e] estimated | UK Analysis Function, symbols in tables |
| Sankey: order nodes to cut crossings, bucket weak links into Other | data-to-viz; EU Data Visualisation Guide |
| Ranks overstate differences; show n or an interval | Goldstein and Spiegelhalter 1996 |
| Mobile: sticky header, frozen first column, hide secondary columns | NN/g Mobile Tables |
Not settled by any authority: Sankey node limits, how many key numbers above the fold, global against local controls. Those are our calls.
The market | Our probe | Compare | Taxonomy and method
┌ rail ────────────┬───────────────────────────────────────────────────────┐
│ THE FLOW │ 0 Overview Where does inference flow? [reported] │
│ 0 Overview │ 101T tokens/wk · 88 providers · 1.4% attested │
│ THE STACK │ ┌ Sankey: app → model → provider → attested? ──────┐ │
│ 1 Apps │ │ OpenRouter, one day. Top 8 per stage + Other. │ │
│ 2 Routers │ │ "Not in the top 20 apps" stays visible (61%). │ │
│ 3 Models │ └──────────────────────────────────────────────────┘ │
│ 4 Providers │ The stack leader coverage │
│ 5 Hardware │ 1 Apps 20 named Hermes 12% 39% of tokens │
│ CROSS-CUT │ 2 Routers 6 OpenRouter 2 of 6 tokens │
│ 6 Public and │ 3 Models 20 + 59 GPT-5.6 Luna full │
│ private │ 4 Providers 88 … full │
│ │ 5 Hardware [x] [x] 1 of 88 │
│ │ 6 Private 4 attested 1.4% tok / 1.7% usd │
│ ├───────────────────────────────────────────────────────┤
│ │ 1 Apps Which apps send the flow? [reported] [Method]│
│ │ … │
└──────────────────┴───────────────────────────────────────────────────────┘
The stack strip is both the summary and the navigation: each row jumps to its section.
─────────────────────────────────────────────────────────────────────────────
4 Providers Who serves it, at what price, speed and uptime?
[reported + measured by others] [Method]
88 providers top 3 hold 28% tok / 41% usd HHI 497 tok / 796 usd
┌ share by provider, 90 days, top 8 + everyone else ─── [tokens | dollars] ┐
└──────────────────────────────────────────────────────────────────────────┘
Side by side sort any column · attested only ☐ · show all 88
── volume ──── ── price ─── ── service ──────── ── trust ──
provider tokens/day % vs builder cache tok/s TTFT up% tier TEE
DeepInfra ████ 1.9T 10 −38% 61% 94 1.2s 99.9 no-store ·
Chutes ██ 0.9T 5 −52% [x] 71 1.9s 98.1 confid. ●
…
[x] not published · [e] estimated · TTFT and up% measured by TrustedRouter, n per row
▸ Concentration over time (HHI)
openrouter.ai providers · tokens · 90 days · updated 2026-09-19 · TrustedRouter leaderboard
─────────────────────────────────────────────────────────────────────────────
| # | Question | Key numbers | Figure | Side-by-side columns | Folds |
|---|---|---|---|---|---|
| 0 Overview | Where does inference flow? | tokens/wk, providers, attested share | Sankey | the stack strip | none |
| 1 Apps | Which apps send the flow? | named apps, top app share, tokens with an app named | app categories, one bar | app, category, tokens/wk, requests, tokens per request | task mix in other populations (ChatGPT, Claude.ai); Claude share by agent (Requesty) |
| 2 Routers | Which routers carry it, and how do their mixes differ? | routers, how many publish tokens, open-weights range | one 100% bar per router by lab, same lab order and colours | router, publishes (tokens, shares, spend), volume, window, open-weights %, top lab, top model, updated | Kilo 63 weeks by lab; Vercel labs over time; OpenCode by country |
| 3 Models | Which models, from which labs? | models, open-weights share, top lab | lab share | model, lab, open weights, tokens/wk, % on OpenRouter, % on Vercel, % on OpenCode, providers, price min to max, attested providers. Row opens the provider board | pulls outside routers (Ollama, Hugging Face); AI Horde |
| 4 Providers | Who serves it, at what price, speed and uptime? | providers, top-3 share, HHI | share by provider over 90 days | as drawn above | concentration over time |
| 5 Hardware | What silicon serves it? | providers that publish silicon (coverage) | Darkbloom chips | quantization by endpoint count, TEE-capable endpoints; silicon column [x] | none |
| 6 Public and private | How much is attested, and by whom? | attested share tok / usd, Chutes TEE share, confidential routes (TrustedRouter) | attested share over time, tokens and dollars | provider, attestation, verified by registry, verified by TrustedRouter, volume and its unit, TTFT, availability | Chutes by its own meter; Darkbloom network; registry coverage; privacy tiers of 1,763 routes |
Sites: orderflow.art, l2beat.com, relayscan.io, mevboost.pics, growthepie.com, ultrasound.money, artificialanalysis.ai, openrouter.ai/rankings, vercel.com/ai-gateway/leaderboards, epoch.ai, arena.ai/leaderboard, gs.statcounter.com, inferencex.semianalysis.com, ourworldindata.org. Second-hand only: defillama.com, rated.network, radar.cloudflare.com, tokenterminal.com.
Literature: Shneiderman, "The Eyes Have It" (1996). Few, "Common Pitfalls in Dashboard Design" and "Best Practices for Scaling Sparklines". NN/g: Progressive Disclosure; Tabs, Used Right; In-Page Links; Scrolling and Attention; Comparison Tables; Data Tables; Mobile Tables. UK Government Analysis Function: tables guidance; symbols in tables. ONS service manual: chart text; annotations. Datawrapper: tables guide. data-to-viz: Sankey. EU Data Visualisation Guide: Sankey and alluvial. Goldstein and Spiegelhalter, "League Tables and Their Limitations" (1996).
"Juice" comes from the Experimental Gameplay Project (Gamasutra, 2005): "constant and bountiful user feedback". Jonasson and Purho ("Juice it or lose it", 2012) and Nijman ("The art of screenshake", 2013) made it a craft; Kelly ("Don't juice it or lose it", 2014) and Kao (2020: medium juice beats both none and extreme) are the cautions. For a measurement site the rule is narrower: motion may say where something went or that an input was received, and may never show a value that was not measured (Heer and Robertson 2007: "maintain valid data graphics during transitions").
| What moves | How | Why |
|---|---|---|
| Buttons, chips, tiles, rail links | colour in 100 ms; press scales to 0.97 | input received |
| Flow diagram hover and isolate | ribbon and node opacity in 120 ms | which flow is being read |
| Table sort, filter, pin, show all | rows slide to their new place in 220 ms (FLIP); over 50 rows or more than a screen of travel, a 120 ms dissolve; a newly pinned row gets one mint tint that fades in 600 ms | the same rows, in a new order |
| Row detail, Method, folds, the estimate card | fade and 4 px in 150 to 160 ms; the card grows from its mark | it belongs to what was clicked |
| Window change | sections dissolve in 140 ms; numbers and charts swap at once, nothing morphs | new data on the same axes |
| Tab change | one marker slides under the tabs in 180 ms; content swaps at once | where you are |
| Particles in the flow diagram | the only ambient motion; "pause motion" stops them and is remembered | motion that starts by itself must be stoppable (WCAG 2.2.2) |
One easing everywhere, cubic-bezier(.2,0,0,1). Opacity and transform only. Everything is off under prefers-reduced-motion.
Never: count-up numbers, charts that draw themselves in, bounce or overshoot on any mark, scroll-triggered reveals, parallax, shake, sound, pulsing badges, anything over 300 ms.
Sources: Gabler et al., "How to Prototype a Game in Under 7 Days" (2005). Swink, Game Feel (2008). Kao, "The effects of juiciness in an action RPG" (2020). Heer and Robertson, "Animated Transitions in Statistical Data Graphics" (2007). Bostock, "Object Constancy". Robertson et al., "Effectiveness of Animation in Trend Visualization" (2008). NN/g, "Animation Duration". Kowalski, "7 Practical Animation Tips". Lewis, "FLIP Your Animations". WCAG 2.2.2 and 2.3.3.
Dan's ask: a tab that shows all the data we have, the rough schemas and tables, the size, the last sync date and how far it goes back. Code: the #tab-data block in index.html and site/data-ui.js. Everything on it is read at load from api/v2/overview.json (the counts: every table and board with rows, bytes, columns, first and last period, last sync, tier, public flag) and, where served, api/v2/tables.json (the columns), api/v2/boards.json (what a board measures, its URL, metrics, citation) and api/v2/sources.json (operator, license, redistribution, citation, terms). Nothing but labels is written in the page.
Same anatomy as the market tab, no new visual language: rail, five numbered sections, header band with the question, source badge and Method, key numbers, one figure, one side-by-side table per section, a key line and a source line.
| # | Question | Key numbers | Figure or table | Row detail |
|---|---|---|---|---|
| 0 Overview | What do we hold, and how far back? | tables, boards (public and private), sources, rows served (plus rows on private boards), bytes on disk, oldest period, newest sync | the timeline strip: one row per market table and per leaderboard source, a bar from first period to last, a tick at last sync, public and private by tone, undated rows grouped at the bottom; every row is a button that opens its table or source below | none |
| 1 Tables | Which tables exist, what does each hold, how big is it? | none | table, what it holds, group, rows, size, columns, first and last period, last sync, tier, served, files (parquet, csv, guide) | every column with type, unit, tier, key and description; a copyable DuckDB one-liner over HTTP |
| 2 Boards | Which leaderboards do we read, what does each measure? | none | source, board, what it measures, ranks, period, rows, first and last period, last collected, tier, license, redistribution, served; filter "public only" | metrics with unit and direction, the board's URL, license, citation, mapped-row share |
| 3 Sources | Who publishes it, under which license? | none | source, operator, kind, tier, license, redistribution, boards (public of total), rows, first and last period, last sync, served | homepage, terms, citation line, how and how often it is read, every board of the source (each opens its row) |
| 4 How to use it | How do I read a table, what may I do with it? | none | the DuckDB and pandas one-liners, the board_entry manifest, the snapshot rule in three sentences, license and citation, the public and private rule | none |
Rules kept: dates as YYYY-MM-DD, sizes in decimal MB with one decimal (KB under 0.05 MB, so no false zero), [private] on rows that are counted and not served, [undated] on tables without a period column, † on a table the dictionary lists and the overview did not count, no bare number as navigation (the strip rows carry names), no count-ups or draw-ins, the strip rows and copy buttons are buttons and so reachable by keyboard.