Context: I run an LLM aggregation platform offering top-tier AI models. Through self-taught SEO/GEO in collaboration with agents, I now have steady overseas orders. 前情提要：我自己在做大模型聚合平台，提供顶尖 AI 模型服务。通过自学跟 Agent 协作 SEO/GEO，现在已经顺利有了海外订单。
About two weeks ago I onboarded an Indian client. His first question was "what models do you have." His second: "How is the token quota for Sonnet 4.6 calculated? I switch by workload — Sonnet for daily work, Opus only for critical tasks." 大概两周前我接待了一个印度客户，他第一句话问“你们有什么模型”，第二句话问：“Sonnet 4.6 的 token 额度怎么算？我按量切换，日常用 Sonnet，只有关键任务才上 Opus。”
The same day, an Indonesian developer asked almost the same questions. He had even pre-calculated his monthly token budget — $12. 同一天，一个印尼的开发者问了几乎一样的问题。他甚至自己算好了一个月的 token 预算——$12。
$12. An entire month's AI budget. $12。一个月全部的 AI 预算。
Run the numbers and you'll see what this figure means. Take a $100 model budget: for an American programmer that's about 5% of a monthly salary. For an Indian developer (average income $800–1,000/month), $100 is 15% of their pay. My European clients, meanwhile, are mostly just curious: "How do you do this? Give me some credits to try." 算一笔账大家就知道这个数字意味着什么了：以 100 美金的模型预算为例，相当于一个美国程序员花自己月薪的 5%，但对一个印度开发者（月均收入在 800-1000 美金左右），100 美元是他们月薪的 15%。其他的欧洲客户，则更多是好奇“你们是怎样做到的？来一些额度试一试”。
Same model. Same price. Completely different weight. AI is creating a bigger monopoly — and a bigger gap — than the internet era ever did. 同一个模型，同一个价格，完全不同的重量。AI 正在制造比互联网时代更大的垄断和 gap。
01 — Worse still, this gap isn't static. It's accelerating. 01 更要命的是，这个 gap 不是静态的，它在加速拉大
One person plus Claude can do the work of half a small team. An indie developer can ship one or even several products a week with AI — 3 to 5 times the output of someone working without it, and that's a conservative estimate. 一个人加上 Claude，能顶半个小团队。独立开发者可以用 AI 一周上线一个甚至多个产品，效率是不用 AI 的人的 3 到 5 倍——这还是保守估计。
And the people in the third world who haven't even touched AI? They're not just one model generation behind — they're an entire era behind. And the distance grows every day: those who use AI get faster and faster, while those who don't are still standing still. 而第三世界那些连 AI 都没碰过的人呢？他们不止落后一个 AI 模型的版本迭代，而是落后一整个时代的差距。而且，这个差距每天都在变大——用 AI 的人越用越快，不用的人还在原地。
In the internet era, if you got online two years late, you could still catch up. In the AI era, being two years late may already put you blocks behind. AI doesn't bring linear speed-up — it brings exponential acceleration. Those who use it fly; those who don't, walk. 互联网时代，你晚两年上网，补课还来得及。AI 时代，你晚两年用上 AI，可能已经被甩开几条街了。AI 带来的不是线性提速，是指数级的迭代加速。用的人飞，不用的人走路。
02 — Third-world developers are fighting "strategy" with "tactics" 02 第三世界国家的开发者正在用“战术”对抗“战略”
My clients from Southeast and South Asia are all masters of frugal engineering: 我那些来自东南亚和南亚的客户，个个都是精打细算的高手：
- Precise model switching — Sonnet 4.6 for daily work, Opus only for critical tasks 精确切换模型——日常 Sonnet 4.6，关键任务才上 Opus
- Obsessive about prompts and caching — standardized system prompts to raise cache hit rates and burn fewer tokens 非常看重 prompt 和缓存——标准化 system prompt，提高缓存命中率，少花 token
- Relay stations instead of official APIs — because $13 with us buys $100 of official quota 使用中转站而不是官方 API——因为我们 $13 能用官方 $100 的额度
It's not that they don't want the best models. They are using every possible means to squeeze the most AI capability out of a limited budget. 他们不是不想用最好的模型。他们是在用一切可能的方式，在有限的预算里，榨出最多的 AI 能力。
03 — Relay stations accidentally became the infrastructure of "AI equity" 03 中转站意外成了“AI 平权”的基础设施
I once wrote an exposé on relay stations — where mixed channels come from, real cache hit rates, the risks. That piece was written from a consumer-protection angle. 我之前写过一篇文章扒中转站的底裤——混渠怎么来的、缓存命中率多少、有什么风险。那篇是从“消费者保护”视角出发。
But looking back now, the very existence of model aggregation platforms and relay stations points to a bigger truth: when compute pricing power is monopolized by a handful of Silicon Valley companies, developers worldwide — and anyone hoping to use AI as leverage — will spontaneously build alternative channels. 但现在回头看，模型聚合平台、中转站这个东西的存在本身，说明了一个更大的问题：当算力定价权被少数几家硅谷公司垄断的时候，全球开发者或是期待借助 AI 力量撬动杠杆的人会自发地建立替代通道。
Reverse-engineered APIs, shared key pools — in Silicon Valley's eyes these may be "gray areas." But for a developer in Bangalore making $800 a month, they are the only entrance to the AI revolution. 逆向工程 API、共享 Key Pool——这些东西在硅谷眼里可能是“灰色地带”。但对一个月薪 $800 的班加罗尔开发者来说，这是他参与 AI 革命的唯一入口。
What does this resemble? Dying to Survive, and the drug-patent wars of the last century. AIDS medication sold for $10,000 a year in the US while average income in Africa was $500. Indian generic manufacturers started producing $300 versions. American pharma called it "infringement" — but without those generics, millions would have died. Note: I'm not declaring this entirely right. But swap "medicine" for "AI capability" and "generic manufacturers" for "API relay stations" — the logic is identical. 这跟什么很像？跟《我不是药神》、上世纪的药品专利之争一模一样。艾滋病药物在美国卖 $10,000 一年，非洲人均收入 $500。印度仿制药厂开始生产 $300 的仿制药。美国药企说这是“侵权”。但没有那些仿制药，几百万人会死。注意：这里我不是要定调这件事是完全正确的，但现在把“药”换成“AI 能力”，把“仿制药厂”换成“API 中转站”。逻辑一模一样。
04 — Other reflections 04 其他感受
I hadn't thought about any of this before starting the business. I just felt the API was too expensive and wanted a cheaper way to use it myself. But when you see where your clients' cities are, when you see someone's monthly AI budget is $12 — it's hard not to think about bigger things. 我做这个业务之前其实没想过这些。就是觉得 API 太贵，能不能便宜点自己用。但当你看到客户的城市来源、看到一个人月 AI 预算 12 美元——你很难不去想更大的事。
Anthropic, OpenAI, Google — three American companies set the global "price of intelligence." Ten years from now, every country's government, education, and healthcare systems will run on AI, and the inference costs will be denominated in dollars. Monopoly and war no longer need aircraft carriers or SWIFT. Just a few racks of H100s. Anthropic、OpenAI、Google——三家美国公司定了全球的“智能价格”。十年后，每个国家的政府、教育、医疗系统都会接入 AI。这些 AI 的推理成本，用美元计价。垄断和战争不需要航母，不需要 SWIFT，只需要几块 H100。
The most fundamental question remains: who has the right to set the price of "intelligence"? Can we have any say in how the world evolves and how civilized it becomes? The answer will decide whether AI makes the world more equal — or less. 最根本的问题还是：谁有权决定“智能”的价格？我们能怎样掌控世界的进化和文明程度吗？这个问题的答案，会决定 AI 到底是让世界更平等，还是更不平等。
Subagents, harnesses, swarm models… We can't wait to play with agents and find the wings that let us take off. But far beyond our field of view, there is a crowd trying to squeeze into the very edge of this map. Subagent, Harness, Swarm model……我们正在迫不及待地跟 Agent 玩耍，找到能让自己起飞的翅膀。但远在视野之外，还有那么一群人正试图挤进这块版图的边缘。