chore(pricing): update RTX PRO 6000 hyperscaler and neocloud $/GPU/hr / 更新 RTX PRO 6000 hyperscaler 与 neocloud 每 GPU 小时单价

[functionstackx]

Summary

Refreshes the RTX PRO 6000 Blackwell Server Edition rental rates in HW_REGISTRY (packages/constants/src/gpu-keys.ts) from the SemiAnalysis AI Cloud TCO model:

Tier	Field	Before	After
Hyperscaler	costh	$0.430 /GPU/hr	$0.677 /GPU/hr
Neocloud	costn	$0.676 /GPU/hr	$0.746 /GPU/hr
3-year rental	costr	$0.52 /GPU/hr	unchanged

HW_REGISTRY is the single source of truth for cost tiers, so this propagates automatically to every surface that reads it: the inference chart's y_costh / y_costn metrics, the throughput calculator's Hyperscaler / Neocloud pricing tiers, the per-dollar compare pages (/compare-per-dollar/*, which display the owning-hyperscaler $/GPU/hr directly), and the tco-feed API. No other files hardcode these numbers.

Note for reviewers: with this change the RTX PRO 6000's 3-year rental rate ($0.52) now sits below both its hyperscaler and neocloud rates, an ordering inversion relative to every other GPU in the registry (where costh ≤ costr). That's out of scope for this PR — flagging in case costr should be refreshed from the same TCO model revision.

Also: unblocks the Overview E2E suite (pre-existing master failure)

Six overview.cy.ts specs have been failing on master since #650 merged, so they fail on any PR touching packages/app — including this one. Not caused by the pricing change; fixed here so this PR can go green.

#650 made a non-disaggregated multi-node deployment its own serving series. The GB300 fixture rows for DeepSeek-V4-Pro are one single-node point (45 tok/s/user) and one multi-node aggregate point (60 tok/s/user), so they no longer blend into a single curve and the cell has no read at @50. #650 updated overview-data.test.ts for exactly this (dsGb300.candidate.read.value → null, missingReason → no_exact_at_tier) but not the Cypress spec, which still expected the old interpolated ≈$1.19 estimate.

The spec now asserts the current behavior:

• estimate / evidence-date / mobile / Chinese-sibling assertions point at cells that still carry a read — B200 ≈$0.89, including the zh sr-only text — so ≈-estimate rendering stays covered;
• the cost-delta grading case moves to GB200 +71%, read from the computed background color because the browser re-serializes the saturated 0.40 alpha as 0.4;
• GB300 gets explicit assertions for its new outcome: a missing cell titled no exact @50 result, with no evidence date and no delta badge.

No product code changed for this — only test expectations.

Testing

• bun run typecheck — pass
• bun run test:unit — pass (436 tests)
• bun run lint / bun run fmt — pass
• overview.cy.ts against a local E2E_FIXTURES=1 production build — 14/14 passing in Chrome (Firefox isn't installed locally; CI covers it)

The pricing change is a data-only constant edit; no UI, i18n, or overlay code paths touched, so no new /zh page is required.

中文说明

依据 SemiAnalysis AI Cloud TCO 模型，更新 HW_REGISTRY（packages/constants/src/gpu-keys.ts）中 RTX PRO 6000 Blackwell Server Edition 的租用单价：

层级	字段	变更前	变更后
Hyperscaler	costh	$0.430 /GPU/hr	$0.677 /GPU/hr
Neocloud	costn	$0.676 /GPU/hr	$0.746 /GPU/hr
3 年租赁	costr	$0.52 /GPU/hr	保持不变

HW_REGISTRY 是各成本层级的唯一数据来源，因此该修改会自动同步到所有读取它的界面：推理图表的 y_costh / y_costn 指标、吞吐量计算器的 Hyperscaler / Neocloud 定价层级、每美元性能对比页面（/compare-per-dollar/*，会直接展示 owning-hyperscaler 的 $/GPU/hr），以及 tco-feed API。仓库中没有其他位置硬编码这两个数值。

给评审的提示：此次改动后，RTX PRO 6000 的 3 年租赁价（$0.52）低于其 hyperscaler 与 neocloud 价格，与注册表中其他 GPU 的排序惯例（costh ≤ costr）相反。这不在本 PR 范围内，仅作提示——如需按同一版 TCO 模型一并更新 costr，请告知。

附带修复：解除 Overview E2E 阻塞（master 上已存在的失败）

自 #650 合并以来，overview.cy.ts 中有 6 个用例在 master 上持续失败，因此任何改动 packages/app 的 PR 都会失败，本 PR 也不例外。该失败与定价改动无关，在此一并修复以使本 PR 通过 CI。

#650 将非分离式多节点部署独立为单独的服务序列。DeepSeek-V4-Pro 的 GB300 夹具数据包含一个单节点点位（45 tok/s/用户）和一个多节点聚合点位（60 tok/s/用户），二者不再合并为同一条曲线，因此该单元格在 @50 档位下没有结果。#650 已相应更新 overview-data.test.ts（dsGb300.candidate.read.value → null，missingReason → no_exact_at_tier），但未同步 Cypress 用例，后者仍期望旧的插值估算值 ≈$1.19。

现将用例改为断言当前行为：

• 估算值、证据日期、移动端与中文页面的断言改为指向仍有结果的单元格——B200 ≈$0.89（含中文 sr-only 文案），从而保留 ≈ 估算值的渲染覆盖；
• 成本差异色阶的断言改用 GB200 +71%，并读取计算后的背景色，因为浏览器会将达到上限的 alpha 值 0.40 重新序列化为 0.4；
• 为 GB300 显式补充新表现的断言：标题为 no exact @50 result 的缺失单元格，且没有证据日期与差异标签。

此项仅修改测试预期，未改动任何产品代码。

测试：bun run typecheck、bun run test:unit（436 项全部通过）、bun run lint、bun run fmt 均通过；在本地 E2E_FIXTURES=1 生产构建下运行 overview.cy.ts，Chrome 环境 14/14 通过（本地未安装 Firefox，由 CI 覆盖）。定价改动为纯数据常量修改，未涉及 UI、i18n 或 overlay 代码路径，因此无需新增 /zh 页面。

🤖 Generated with Claude Code

---

Note

Low Risk
Registry constant refresh plus test-only overview assertions; no auth or runtime logic changes, though RTX 3-year costr now sits below hyperscaler/neocloud (noted in PR description).

Overview
Updates RTX PRO 6000 hyperscaler and neocloud $/GPU/hr in HW_REGISTRY (costh 0.43→0.677, costn 0.676→0.746; costr unchanged at 0.52), so every consumer of HW_REGISTRY picks up the new SemiAnalysis TCO rates automatically.

Realigns Overview Cypress specs with current matrix behavior for DeepSeek-V4-Pro: estimated cost and evidence on B200 (≈$0.89) instead of GB300; GB300 cells assert missing @50 (∞, no blended estimate) because single- vs multi-node series do not interpolate to the tier; GB200 cost-delta expectations move to +71% with a computed-style background-color assertion; mobile and /zh flows mirror the same semantics.

^{Reviewed by Cursor Bugbot for commit 0421d75. 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 (UTC)
inferencemax-app	Ready	Preview	Jul 30, 2026 6:10am