DEV Community: rock zhu

The Rate Limit Cliff You Can't See

rock zhu — Sat, 11 Jul 2026 00:30:50 +0000

Most API pricing tiers don't measure what they claim to measure.

How Token Buckets Create Invisible Throughput Walls

A typical SaaS API tier promises 1,000 requests per minute at $49/month, 10,000 at $499/month. The math looks linear. But most platforms use token bucket algorithms—a system that refills a fixed bucket of allowance at regular intervals. The problem: tokens don't refill uniformly.

What this means in practice: you can stay within your tier's stated limits and still hit a hard throughput ceiling. A developer building a batch job or handling a traffic spike discovers that their "tier" actually allows far fewer concurrent requests than advertised. The only sol

The mechanism is intentional cost architecture, not a bug. Platforms benefit from the opacity: customers upgrade to tiers they don't fully need, and the vendor captures margin on what looks like incremental cost. For builders, the lesson is to model your actual burst capacity, no

Tier Jumps vs. Actual Capacity

01 Advertised burst allowance — Usually 2–5x the per-minute rate; token bucket refill rate masks the true peak.
02 Tier price ratio — Jumps 5–10x between tiers, but throughput capacity often only doubles.
03 Hidden upgrade trigger — Most users hit the ceiling 2–3 months into production, long after initial benchmarking.
04 Audit step — Model your 95th-percentile request burst, not your average; budget one tier above your peak need.

Full write-up: https://cxgo.ai/l/PLe8uzs

Your real AI cost is 3–5 higher than you calculated

rock zhu — Fri, 10 Jul 2026 00:30:36 +0000

System prompts and context windows quietly compound token spend far beyond inference alone.

Full write-up: https://cxgo.ai/l/cQoW4Ny

10亿额度制如何在高波动日蚕食散户成交

rock zhu — Wed, 08 Jul 2026 00:30:23 +0000

每日人民币额度上限在市场波动时成为隐形成本，机构T+0对冲规则套利该机制漏洞。

完整解读: https://cxgo.ai/l/SsX7lAx

Token-based pricing hides a 3–5 cost trap

rock zhu — Tue, 07 Jul 2026 00:30:27 +0000

High-frequency, low-value API calls compound hidden expenses that monthly plans would cap.

Full write-up: https://cxgo.ai/l/90zjWgs

南向资金多付的那笔税从没人算过

rock zhu — Sun, 05 Jul 2026 00:30:58 +0000

同一只港股，内地买家的实际摩擦成本在波动日比本地投资者高出约40%，差额藏在制度设计里。

完整解读: https://cxgo.ai/l/ov8KgO1

QT悄悄抬高南向资金的隐性融资成本

rock zhu — Fri, 03 Jul 2026 00:30:36 +0000

美联储缩表通过五个环节传导至港股北水的保证金利率，多数散户框架里这条链条是断的。

完整解读: https://cxgo.ai/l/dCoEd3U

黄金避险有个死角你踩过吗

rock zhu — Tue, 30 Jun 2026 00:30:27 +0000

QT缩表期间，机构用黄金补仓债券亏损，负相关性最长失效55天。

流动性虹吸窗口的四个测量值

40–55d — 单次失效窗口平均持续
6WK — 2022Q4最长连续同向下行
QT期 — 触发条件：美联储主动缩表
10Y — 参照锚：美国国债收益率

完整解读: https://cxgo.ai/l/ZcoriGk

那个'月费' 你从没付过

rock zhu — Sun, 28 Jun 2026 00:30:54 +0000

定价页显示的月费几乎都是年付折算价——真实月付价高出 30–50%，多数用户从未察觉。

完整解读: https://cxgo.ai/l/Z0EOTSG

季末最后3天南向资金悄悄撤回A股

rock zhu — Sat, 27 Jun 2026 00:30:43 +0000

过去8个季度，恒生科技在季末尾盘3日平均跑输沪深300达1.7%，机构报告几乎从不单独披露。

完整解读: https://cxgo.ai/l/8SnFQ0A

Free tier 不是为了转化你—— 是为了耗尽竞品的预算

rock zhu — Thu, 25 Jun 2026 00:31:18 +0000

占据关键词排名和社区讨论位后，竞品触达同一批用户的广告成本上涨 3–5 倍。

完整解读: https://cxgo.ai/l/dJphBri

供应商从未告诉你的那个数字

rock zhu — Wed, 24 Jun 2026 00:31:04 +0000

基准测试都是串行跑的——而你的用户，不是。

为什么 50 个并发用户是一道隐形的门槛

首 token 延迟（TTFT）是 LLM 产品中最直接影响用户感知的指标。单用户测试时，主流 API 的 TTFT 通常落在 300–800ms 之间，足以让早期演示看起来流畅。但当并发请求从 0 爬升到 50，这个数字往往膨胀至 2–5 秒，有时更高——原因不是网络，而是模型服务层的排队机制与 KV-cache 争抢。

供应商公布的 benchmark 几乎全部采用串行请求：一个请求完成后再发下一个，这在生产环境中几乎不存在。真实的 SaaS 场景里，用户同时打开页面、同时触发补全，请求在同一个 token-generation 队列里堆叠。这不是供应商隐瞒数据，而是行业惯例从未被挑战过。

对早期产品而言，这个落差的危险在于时机：它只在你开始增长时才会出现。内测阶段一切正常，公测第一天用户涌入，响应变慢，流失发生——而你的监控面板上，错误率依然是零。

你应该在 benchmark 里自己补上的测试

01 并发压力下的 TTFT 基线 — 用 50 / 100 并发分别跑 P50 与 P95，而非平均值。
02 冷池与热池的延迟差 — 部分供应商在低流量时会缩减实例，首个请求延迟可达 3–6 秒。
03 排队超时的用户可见行为 — API 超时不等于用户看到错误；更常见的是无声的白屏等待。
04 流式输出的感知优化空间 — Streaming 不能降低 TTFT，但能将「卡顿感」的感知阈值从 800ms 推到约 2s。

完整解读: https://cxgo.ai/l/a3LVBGp

每年9月离岸人民币悄悄贬值 150–300基点

rock zhu — Tue, 23 Jun 2026 00:30:33 +0000

中国企业季末集中购汇，离岸市场无央行托底，汇差静默侵蚀港股与东南亚人民币资产持仓。

完整解读: https://cxgo.ai/l/n8GpQso

DEV Community: rock zhu

The Rate Limit Cliff You Can't See

How Token Buckets Create Invisible Throughput Walls

Tier Jumps vs. Actual Capacity

Your real AI cost is 3–5 higher than you calculated

10亿额度制 如何在高波动日 蚕食散户成交

Token-based pricing hides a 3–5 cost trap

南向资金 多付的那笔税 从没人算过

QT悄悄抬高 南向资金的 隐性融资成本

黄金避险 有个死角 你踩过吗