feat(app): profit estimator per chip-hour and per gigawatt, TCO Calculator and Fleet Lifecycle to footer / 每芯片小时与每吉瓦利润估算器，TCO 计算器与集群生命周期移至页脚

[functionstackx]

Summary

Adds two tabs, both hosted by the same component with a basis prop:

• /profit-estimator (per chip-hour): one vertical stacked bar per SKU showing the US$ one chip earns in one hour of Kimi K3 agentic inference at a chosen interactivity, split into compute expense (the TCO tier's $/chip/hr), the model license fee (30% of revenue by default), and profit. Figures are dollars and cents.
• /profit-estimator-per-gigawatt (per GW-year): the same per-chip figures scaled by the GPU-hours one all-in utility gigawatt-year buys for that SKU, so chips with different power draw compare on one denominator. Figures are compact ($135.2B).

Both have /zh mirrors and /<tab>/<model slug> routes. In the nav they sit between Inference Performance and Accuracy Evals; TCO Calculator and Fleet Lifecycle move out of the tab bar into the footer's "More" column (pages, /zh mirrors and sitemap entries unchanged). Revenue and margin (profit ÷ revenue) sit above each bar, with the vendor's full-color mark above those; each segment carries its name and dollar amount when it fits. A loss is drawn below the zero line as a hatched segment in the SKU's colour.

Inputs and defaults:

• Target interactivity: 45 tok/s/user, typed number. A SKU whose measured curve does not reach the target (H200 on Kimi K3 at 45) is not drawn and not offered in the legend.
• Utilization: 60%. Scales revenue only (revenue × 0.6); TCO is unchanged. No separate haircut segment; the subtitle states the rate.
• Model License Fee: 30% of revenue (a royalty on every token sold, owed even at a loss). Code and test ids keep the older labCut name.
• Cost provider: Owning - Hyperscaler / Owning - Neocloud Giant / 3 Year Rental (same names as the /inference y-axis selector), or Custom $/GPU/hr with one input per base chip, seeded from the hyperscaler tier. The TCO $/chip/hr badges under the title list only the SKUs the legend currently shows.
• Token price: OpenRouter catalog for the model, or custom input / cached input / output $/M.
• Model selector: Kimi K3 only for now. Widening is one edit to MODEL_ROUTE_TAB_MODELS['profit-estimator']. No scenario selector (agentic traces are pinned) and no precision selector (auto).
• SKU legend toggles which bars are drawn.

Header matches /inference: title "Kimi K3 2.8T Agentic Revenue & Profit Estimates per GigaWatt Per Year at P90 45 tok/s/user Interactivity" (per chip-hour: "... per Chip per Hour at ..."), then cost tier, utilization, model license fee, updated date and source, TCO $/chip/hr badges (the TCO source line is omitted for custom $/GPU/hr), and selling prices. The section above the controls is titled "Revenue & Profit Estimator" / "Revenue & Profit Estimator per GigaWatt" with no intro paragraph. The y axis reads "Revenue per chip per hour ($ USD)" or "Revenue per all-in provisioned utility GW per year ($ USD)". Segments are labelled in place in title case (Compute Expense, Model License Fee, Profit), so there is no separate key and no hover hint under the chart. Formula notes live in a fold titled "Revenue per Chip-Hour Formula" / "Revenue per GigaWatt Formula", collapsed by default; the export button sits in the card's top-right corner and writes PNG and CSV.

X-axis labels stand upright on two lines (SKU name, then framework and precision) when every tick has room for them, and fall back to the -50° slant otherwise, with margins sized from the labels so nothing clips. The vendor's full-color mark sits only above each bar (NVIDIA green never inverted; AMD's black arrow inverts in dark mode since AMD publishes no color mark) and scales with the bar width, 30% of the band clamped to 14–48px, so desktop bars get a large logo and phones the small one. The y domain leaves exactly the pixel headroom those labels and that mark need above the tallest bar, and the top margin is 12px, so the grid starts right under the selling-price line. The chart height follows the viewport: 720px when the window has room, otherwise the viewport minus 260px for the sticky nav and card header (never below 440px), so on a MacBook the card title and the x labels fit on one screen. Below 640px the chart switches to compact margins and height, drops segment names then amounts that would not fit, keeps only the margin percentage above each bar, and uses a short y-axis title. Custom $/M token inputs take their own row under the main controls. Number inputs blur on mouse wheel so page scrolling never changes a percentage.

Arithmetic per SKU, written out in docs/tco-calculator.md:

gpuHours   = 1                                            # /profit-estimator
gpuHours   = (1,000,000 kW ÷ all-in kW per GPU) × 8,760 h  # /profit-estimator-per-gigawatt
revenue    = $/GPU/hr at sale price × gpuHours × utilization
tco        = cost $/GPU/hr × gpuHours
licenseFee = revenue × licenseFeePct
profit     = revenue − tco − licenseFee

The GW-year conversion is the same formula PR #973 uses for its y-metrics; the helpers here live in profit-estimator.ts and can be collapsed onto whichever lands second.

Routing

• /profit-estimator and /profit-estimator-per-gigawatt each open on Kimi K3 (defaultRouteModel). /<tab>/kimi-k3 is the same page; Kimi-K3 308s to the canonical slug; a model outside the allow-list (e.g. deepseek-v4) 404s. Same on /zh.
• calculator and fleet are navGroup: 'footer-only'; the footer links are footer-link-calculator and footer-link-fleet and call track().
• Changing the model in the selector rewrites the route.
• Sitemap lists only allow-listed model routes for this tab.

Library change

chart-update.ts applies xTickFormat to band x-axes. No existing chart combined a band x-axis with tickFormat, so nothing else changes. d3-chart-wrapper.tsx skips the instructions paragraph when instructions is an empty string; charts that pass nothing still get the default hint.

Files

• packages/app/src/components/calculator/profit-estimator.ts — pure math and types
• packages/app/src/components/calculator/ProfitEstimatorChart.tsx — D3Chart stacked bars, in-bar labels, hatch patterns, vendor marks, tooltip
• packages/app/src/components/calculator/ProfitEstimatorDisplay.tsx — controls, pricing, legend, header, export, formula fold (EN/ZH strings)
• packages/app/src/components/ui/result-context.tsx — optional utilization and licenseFee lines
• packages/app/src/app/(dashboard)/{profit-estimator,profit-estimator-per-gigawatt}/{page,[model]/page}.tsx and the .../zh/(dashboard)/... mirrors
• packages/app/src/components/footer/footer.tsx — TCO Calculator and Fleet Lifecycle links
• model-routes.ts, sitemap.ts, dashboard-routes.ts, tab-meta.ts, tab-meta-zh.ts, tab-nav.tsx
• cypress/support/profit-fixtures.ts — synthetic Kimi K3 agentic rows (the captured fixtures have no agentic_traces)
• docs/tco-calculator.md — new section

Tests

• profit-estimator.test.ts: GPU-hours, utilization scaling, revenue-based license fee, skip reasons, chip-hour basis (GPU-hours = 1, power ignored), USD formatting per basis.
• ProfitEstimatorChart.test.ts: segment stacking order, negative-profit segments, y-domain including losses, tooltip HTML, label resolution, vendor mark lookup, splitAxisLabel / xLabelLayout (two-line vs slanted), slantedMargins, pixel-based y headroom that follows the vendor-mark size, barMarkHeight bounds, width-aware segment and margin labels, chip-hour tooltip (cents, no GPU-hours footer), chartHeightForViewport bounds.
• model-routes.test.ts, sitemap.test.ts, tab-meta.test.ts, dashboard-routes.test.ts: allow-list, default model, slug rewrite, sitemap entries, nav order (both profit tabs between inference and evaluation), calculator and fleet footer-only.
• cypress/e2e/profit-estimator.cy.ts (15): per-GW defaults and title, revenue scales with utilization while TCO does not, license fee = revenue × pct, out-of-range H200 skipped, legend filter, custom cost provider, custom price triple, model route rewrite, 404 for non-allow-listed slug, /zh mirror, collapsed formula fold, two-line labels at desktop width, nav order and footer links, chart height at 720/1200/600px viewports; per chip-hour defaults, title, cent-formatted revenue labels, slug 308/404 and /zh mirror. 15/15 passing locally; navigation, throughput-calculator and csv-export specs updated for the footer links (13 + 74 + 7 passing).
• bun run typecheck, lint, fmt, check:typography clean; full unit suite passing under Node 24.

Screenshots

Light (EN) and dark (/zh) renders are in the Slack thread where this was requested; the repo does not keep doc screenshots.

中文说明

新增两个页签，共用同一组件并以 basis 属性区分：/profit-estimator 按每芯片每小时计价，收入为目标交互性下的 $/GPU/hr 乘以利用率，算力支出为所选 TCO 层级的 $/chip/hr，模型许可费默认为收入的 30%，金额显示到美分；/profit-estimator-per-gigawatt 将同一组每芯片数据乘以一个全电源配置吉瓦年可购得的 GPU 小时数，使功耗不同的芯片在同一分母下比较。两者均含 /zh 镜像与 /<页签>/<模型 slug> 路由。导航中两个页签位于推理性能与准确率评估之间；TCO 计算器与集群生命周期从页签栏移至页脚「更多」栏（页面、/zh 镜像与 sitemap 不变）。每款芯片一根垂直堆叠柱，拆分为算力支出（TCO）、模型许可费与利润；柱顶标注收入、利润率与厂商全彩标志，各段内标注名称与金额；亏损以该芯片颜色的斜纹段绘制在零线以下。

输入与默认值：目标交互性 45 tok/s/user（数字输入，实测曲线未覆盖目标的芯片不绘制且不出现在图例中）；利用率 60%（仅缩放收入，TCO 不变）；模型许可费为收入的 30%；成本供应商选项与 /inference 纵轴选择器同名（自有 - 超大规模、自有 - Neocloud Giant、3 年租赁）或按芯片自定义 $/GPU/hr；标题下的 TCO $/chip/hr 徽标仅列出图例当前显示的芯片；Token 售价取自 OpenRouter 或自定义输入/缓存输入/输出价。模型选择器目前仅开放 Kimi K3，切换模型会改写路由；不在白名单内的 slug 返回 404。移除场景与精度选择器。标题与副标题格式对齐 /inference（每吉瓦每年收入与利润估算），模块标题为「收入与利润估算器」且无说明段，自定义 $/GPU/hr 时不显示 TCO 来源；x 轴标签在空间足够时两行竖排（SKU 名称 / 框架），否则斜排并按标签长度扩展边距；厂商全彩标志仅显示在柱顶并随柱宽缩放（14–48px），顶部留白随之调整；各段标签首字母大写；每吉瓦视图的悬浮框不再显示 GPU 小时数；副标题含利用率与模型许可费；各段就地标注，无单独图例，图表下方无悬停提示；公式折叠默认收起并命名为「每芯片小时收入公式」/「每吉瓦收入公式」；图表高度随视口调整（有余量时 720px，否则为视口高度减 260px，最低 440px），使标题与 x 轴标签在笔记本屏幕上同屏可见；手机宽度下使用紧凑边距并按宽度省略柱内标签；自定义 $/M token 输入独占一行；导出按钮位于卡片右上角，支持 PNG 与 CSV；数字输入框在滚轮时失焦。

计算公式见 docs/tco-calculator.md：GPU 小时数 = 1（每芯片小时）或 (1,000,000 kW ÷ 每 GPU 全电源配置 kW) × 8,760（每吉瓦年）；收入 = 售价折算的 $/GPU/hr × GPU 小时数 × 利用率；TCO = 成本 $/GPU/hr × GPU 小时数；许可费 = 收入 × 比例；利润 = 收入 − TCO − 许可费。Cypress 使用 cypress/support/profit-fixtures.ts 中合成的 Kimi K3 智能体数据（现有 fixture 无 agentic_traces）。

---

Note

Medium Risk
New revenue/profit calculations and a major nav reshuffle change how users reach key tools; risk is mitigated by unit tests, Cypress specs, and reuse of existing TCO/token-revenue logic rather than new auth or data stores.

Overview
Adds two primary nav tabs—/profit-estimator (per chip-hour) and /profit-estimator-per-gigawatt (per GW-year)—that share ProfitEstimatorDisplay and plot stacked revenue, TCO, model license fee, and profit per SKU at a fixed agentic-traces workload (Kimi K3 only for now), with OpenRouter or custom token pricing, TCO tiers or custom $/GPU/hr, utilization, and typed interactivity.

TCO Calculator and Fleet Lifecycle leave the tab bar for footer links; pages and sitemap entries stay the same. Routing adds EN//zh pages plus per-model slugs (308 aliases, 404 outside allow-list); sitemap uses per-tab modelRoutesForTab and defaultRouteModel (Kimi K3 on profit tabs).

Supporting changes: profit-estimator.ts economics, D3 stacked chart with responsive height and in-bar labels, ResultContext utilization/license fee lines, band-axis xTickFormat in chart updates, optional empty chart instructions, docs in tco-calculator.md, Cypress profit fixtures, and broad unit/e2e coverage.

^{Reviewed by Cursor Bugbot for commit 879bb96. Bugbot is set up for automated code reviews on this repo. Configure here.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
inferencemax-app	Ready	Preview	Sep 4, 2026 2:52am UTC

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 4cdabfd. Configure here.}