Concise Summary简洁概述
GitHub Copilot's shift to usage-based pricing exposes a structural mismatch that has existed since generative AI subscriptions began: flat monthly fees cannot cover wildly variable per-user LLM inference costs.
AI companies have deliberately obscured token costs behind subscriptions, letting users incur $8-13.50 in compute for every $1 paid — a subsidy the author says cannot continue as OpenAI and Anthropic face their own revenue-growth deadlines.
GitHub Copilot 转向按量计费,暴露的是生成式 AI 订阅制自诞生起就存在的结构性错配:固定月费根本无法覆盖因人而异、剧烈波动的大模型推理成本。
AI 公司一直在用订阅制刻意掩盖 token 成本,让用户每付 1 美元就能烧掉 8 到 13.5 美元的算力——作者认为,随着 OpenAI 和 Anthropic 自身面临营收增长的死线,这种补贴已经难以为继。
Infographic信息图
The broken Uber analogy
被滥用的 Uber 类比
Uber's price hikes never changed its underlying cost structure — riders pay per ride, drivers pay their own gas. AI subscriptions are the opposite: a flat $20/month bought unlimited rides while the company secretly paid $150/gallon gas on your behalf. Usage-based billing isn't a normal price increase — it's revealing costs that were always there.
Uber 涨价从未改变其底层成本结构——乘客按次付费,司机自付油钱。AI 订阅恰恰相反:用户每月付 20 美元就能"无限打车",而公司在背后偷偷替你付着每加仑 150 美元的"油钱"。按量计费不是普通涨价,而是把一直存在、却被刻意隐藏的成本摊开给你看。
Why gyms work but AI doesn't
健身房模式为何不适用于 AI
Flat subscriptions only work when marginal cost per user is roughly stable — gyms know equipment wear, Google Workspace knows storage costs. LLM inference cost per user varies by orders of magnitude depending on context window size, model choice, and task complexity, making flat pricing structurally unsustainable rather than just underpriced.
包月订阅模式成立的前提是单个用户的边际成本大致稳定——健身房清楚器材磨损,Workspace 清楚存储开销。而大模型的推理成本会因上下文窗口大小、模型选择、任务复杂度出现数量级的差异,这不是"定价偏低"的问题,而是包月模式本身在结构上就撑不住。
Sycophancy and hidden costs breed forgiveness
隐藏成本催生的"纵容效应"
Because users don't see per-task token costs, they forgive hallucinations, stuck agents, and jagged intelligence as quirks to be patient with. The author argues that if every failed attempt carried a visible $15 bill, tolerance for LLM unreliability would collapse — the subscription veil is what sustains the industry's reputation.
因为用户看不到单次任务的真实 token 账单,他们会把幻觉、卡顿、能力参差不齐都当作"还能忍"的小毛病。作者认为,如果每次模型"卡壳"都对应一张 15 美元的账单,用户对大模型不可靠性的容忍度会瞬间崩塌——正是订阅制这层遮羞布,撑起了整个行业的口碑。
OpenAI's math as the canary
OpenAI 的数字账是警报器
The piece pivots from Copilot to OpenAI's need to roughly 10x revenue by 2030 to meet compute commitments, and its CFO's own doubts about paying for contracts without growth. When the company closest to the center of the industry shows this much fragility, the author reads it as a leading indicator, not an isolated event.
文章从 Copilot 转向 OpenAI:要履行其算力合约,营收必须在 2030 年前扩大约十倍,而其 CFO 本人都在公开质疑增长若不到位该如何支付账单。当行业最核心的公司都显露出这种脆弱性,作者认为这不是孤立事件,而是一个先行指标。
Detailed Summary详细解读
The piece opens with GitHub Copilot's June 2026 shift to usage-based pricing, framed by Microsoft as a response to Copilot becoming an 'agentic platform' with higher inference needs. The author reads through this framing: the real story is that Microsoft subsidized nearly 2 million users' compute costs for three years, echoing a 2023 WSJ report that some Copilot users cost the company up to $80/month against a $10 subscription.
The core argument is an extended, inverted Uber analogy. Real Uber pricing keeps its cost structure stable — riders pay per ride, drivers absorb their own gas costs. AI subscriptions are structurally opposite: users pay a flat monthly fee for effectively unlimited 'rides' while the company secretly absorbs runaway 'gas' (compute) costs behind the scenes, making a later shift to usage-based billing feel like a bait-and-switch rather than a normal price hike.
The author explains why flat subscriptions work for gyms or Google Workspace but not LLM services: those businesses have relatively stable, predictable marginal costs per user. LLM inference cost varies enormously depending on context window size, model tier, and task type — and providers have no real lever to control usage except degrading the product, so they instead hide token counts and impose opaque rate limits.
A key psychological claim: hidden per-task costs make users forgive LLM failures — hallucinations, stuck agentic sessions, 'jagged intelligence' — because a bad output feels like wasted time, not wasted money. The author argues that if every failure carried a visible dollar cost, tolerance for unreliability would collapse quickly, which is why the industry has strong incentives to keep costs opaque.
The piece then scales up to enterprise exposure: Uber's CTO reportedly burned through a full year's AI budget in months; Goldman Sachs found some companies' token spend approaching 10% of labor costs, projected to reach 100% within quarters. Team-tier subscriptions marketed at $125/seat routinely translate into $5,000-10,000/month in actual API spend for a 10-person team — a gap the author says makes ROI claims for enterprise AI largely unverifiable.
The piece closes by connecting Copilot's pricing shift to OpenAI's existential math: the company reportedly needs roughly 10x revenue growth by 2030 to meet its compute commitments, and its own CFO has voiced doubt about paying for contracts absent that growth, plus concerns about IPO-readiness. The author treats Microsoft — the best-capitalized AI subsidizer — capitulating on subsidies as a leading indicator that OpenAI and Anthropic's eventual full shift to usage-based billing (or worse) is the real signal to watch for.
文章开篇讲述 GitHub Copilot 在 2026 年 6 月转向按量计费,微软官方说法是 Copilot 已演变为需要更多推理算力的"智能体式平台"。作者拆穿这套话术:真正的故事是微软已经连续三年为近两百万用户补贴算力成本,呼应了《华尔街日报》2023 年的报道——部分 Copilot 用户每月给公司造成的成本高达 80 美元,而订阅费只有 10 美元。
核心论证是一个被反转的 Uber 类比。真实的 Uber 涨价不会改变底层成本结构——乘客按次付费,司机自担油钱。AI 订阅在结构上恰恰相反:用户每月付固定费用换取近乎无限的"乘车次数",而公司在背后默默吸收了失控的"油费"(算力成本),这使得后续转向按量计费更像是一场"钓鱼式"背叛,而非正常涨价。
作者解释了为何包月制适用于健身房或 Google Workspace,却不适用于大模型服务:前者的单用户边际成本相对稳定、可预测。而大模型的推理成本会因上下文窗口大小、模型档位、任务类型出现巨大差异,服务商除了让产品变差之外几乎无法真正控制用户的使用方式,于是转而隐藏 token 数量、设置不透明的速率限制。
一个关键的心理学论点是:隐藏的单次任务成本让用户更容易原谅大模型的失败——幻觉、卡死的智能体会话、参差不齐的能力表现——因为糟糕的输出只让人觉得浪费了时间,而非浪费了钱。作者认为,如果每次失败都对应一笔清晰可见的账单,用户对不可靠性的容忍度会迅速崩溃,这正是行业极力维持成本不透明的动机所在。
接着文章把视角拉到企业层面:据报道,Uber 的 CTO 几个月内就烧光了全年 AI 预算;高盛发现部分公司的 token 支出已逼近人力成本的 10%,并可能在几个季度内升至 100%。市面上宣传每席位 125 美元的团队版订阅,落到一个 10 人团队身上往往对应每月 5000 到 10000 美元的真实 API 开销——作者认为,这个巨大落差使得企业 AI 的投资回报率几乎无从验证。
结尾把 Copilot 的定价调整与 OpenAI 的生存算账联系起来:据报道该公司需要在 2030 年前把营收扩大约十倍才能履行算力合约,其 CFO 本人也公开表达过增长不到位就难以支付合约款项的担忧,以及公司尚未具备上市所需的严格报表规范。作者认为,连资本最雄厚、最有能力持续补贴算力的微软都撑不住补贴,这本身就是一个先行信号——真正值得警惕的信号,将是 OpenAI 或 Anthropic 把所有订阅用户都转向按量计费。
FAQ常见问答
Why can't AI companies just keep subsidizing usage the way they have been?AI 公司为什么不能继续像以前那样补贴算力?
Reasoning models consume more tokens over time, not fewer, so inference costs haven't fallen as expected. Combined with rising enterprise usage and looming debt obligations (e.g. OpenAI's compute contracts), the subsidy math no longer closes for even the best-capitalized players.
推理模型消耗的 token 不减反增,算力成本并未像预期那样下降。再加上企业使用量攀升、算力合约债务临近(如 OpenAI 的算力承诺),即便是资本最雄厚的公司,补贴的账也算不下去了。
Is the Uber comparison actually fair, or is it a rhetorical trick?Uber 类比真的成立吗,还是只是一种修辞手法?
The author explicitly inverts it to show the mismatch: real Uber pricing never required restructuring the platform's economics. The comparison is a deliberate contrast, not a direct analogy — its purpose is to highlight how differently AI pricing behaves.
作者明确是在"反转"这个类比来凸显错配:真实的 Uber 涨价从不需要重构平台底层经济结构。这个比较本身是刻意制造的对比,而非直接类比,目的是凸显 AI 定价行为的异常之处。
Does the piece offer hard data on AI companies' actual margins?文章有给出 AI 公司真实利润率的确凿数据吗?
Only fragments — a 2023 WSJ report on Copilot losses, an $8-13.50-per-dollar burn ratio attributed to Anthropic, and Anthropic's own now-revised Claude Code cost documentation. Company-wide unit economics are not disclosed and the author relies on inference and leaked/reported figures.
只有零散数据点——2023 年《华尔街日报》关于 Copilot 亏损的报道、归于 Anthropic 的"每 1 美元烧 8 到 13.5 美元"比例,以及 Anthropic 自家后来被修改过的 Claude Code 成本文档。公司层面的完整单位经济数据并未披露,作者更多依赖推断与二手报道数字。
What would actually falsify the author's thesis?什么情况会真正推翻作者的论点?
If inference costs per query fell faster than usage grew — through model efficiency gains or hardware cost drops — flat subscriptions could become viable again. The piece doesn't seriously engage with this possibility, treating rising reasoning-model token use as a fixed trend rather than a current phase.
如果单次查询的推理成本下降速度超过使用量增长——比如通过模型效率提升或硬件成本下降——包月订阅仍有可能重新变得可行。文章并未认真讨论这种可能性,而是把推理模型 token 消耗上升当作一种固定趋势,而非当前阶段的特征。
Is Ed Zitron a neutral source on this topic?Ed Zitron 在这个话题上是中立信源吗?
No — he's a long-time, vocal AI-bubble skeptic (see his 'Subprime AI Crisis' essay) writing on his own blog, not a financial analyst. His track record on this specific prediction lends credibility, but the piece is opinion journalism with a consistent, declared point of view, not neutral reporting.
不是——他是长期公开唱衰 AI 泡沫的博主(参见其《次贷式 AI 危机》一文),这篇文章发在自己的博客上,而非出自金融分析师之手。他此前对这一具体预测的命中记录增加了可信度,但本质上仍是立场鲜明的评论文章,而非中立报道。
In-depth Analysis · Pros & Cons深入解读 · 优缺点
This piece traces GitHub Copilot's shift to usage-based pricing back to a structural flaw the author has warned about for years: subscription pricing cannot cover LLM inference costs that swing wildly per user. It uses that single pricing change as a lens to argue the entire generative-AI industry is economically unsound.
这篇文章从 GitHub Copilot 转向按量计费这一具体事件切入,追溯到作者多年来反复警告的一个结构性缺陷:订阅制定价根本无法覆盖因人而异、剧烈波动的大模型推理成本。作者借这一次定价调整,把矛头指向整个生成式 AI 行业的经济根基。
- Predictive track record预测记录经过验证The author's 2024 'Subprime AI Crisis' essay predicted this exact pricing shift years in advance, giving the piece unusual credibility as a follow-through rather than post-hoc commentary.作者 2024 年发表的《次贷式 AI 危机》一文提前数年精准预言了这次定价转变,使本文作为"预言兑现"而非事后诸葛亮的评论,具备了不寻常的可信度。
- Clear structural mechanism结构性机制阐述清晰The gym/Workspace vs. LLM cost-variance comparison gives a concrete, testable economic reason why flat subscriptions fail for AI specifically, rather than a vague 'AI is expensive' complaint.健身房/Workspace 与大模型成本波动性的对比,为"为何包月制在 AI 领域行不通"给出了具体、可检验的经济学解释,而非泛泛的"AI 很贵"式抱怨。
- Grounded in a real, dated event锚定在真实的具体事件上Rather than abstract speculation, the argument is anchored to a verifiable, dated policy change (Copilot's June 1, 2026 pricing shift), making the claims falsifiable against real subsequent developments.论证并非抽象空谈,而是锚定在一个可核实的、有明确日期的政策变化上(Copilot 2026 年 6 月 1 日的定价调整),使这些论断可以被后续真实进展验证或证伪。
- Connects micro to macro打通了微观与宏观论证链The piece links a single product's pricing memo to systemic risk (Oracle's stock, Larry Ellison's margin exposure, OpenAI's 2030 obligations), showing how a small policy change reflects industry-wide fragility.文章把一份单一产品的定价备忘录,与系统性风险(Oracle 股价、Larry Ellison 的保证金敞口、OpenAI 到 2030 年的义务)串联起来,展示了一次微小的政策调整如何折射出整个行业的脆弱性。
- Cherry-picked cost figures成本数字存在选择性引用The $8-13.50-per-dollar burn ratio and $80/month Copilot cost figures come from a single 2023 WSJ report and unverified Reddit estimates, not audited financial disclosures, so the magnitude of the mismatch may be overstated.文中反复引用的"每 1 美元烧 8 到 13.5 美元"比例,以及 Copilot 每月 80 美元成本,均来自 2023 年一篇 WSJ 报道及未经核实的 Reddit 估算,而非经审计的财务披露,错配程度可能被夸大。
- Assumes reasoning costs won't fall假设推理成本不会下降The piece treats rising token consumption per reasoning-model query as a permanent trend, but model efficiency, hardware costs, and inference optimization have historically driven per-query costs down even as capability rose — a possibility the essay dismisses rather than argues against.文章把推理模型单次查询 token 消耗的上升视为一种永久趋势,但从历史看,模型效率提升、硬件成本下降与推理优化往往能压低单次查询成本,即便能力在同步提升——这一可能性文章只是搁置,并未认真反驳。
- Heavy reliance on tone over evidence论证依赖情绪化语言多于证据Large portions of the argument rely on charged language ('scam,' 'abusive,' 'insulting') rather than additional data, which may alienate readers seeking a more dispassionate cost-benefit analysis of enterprise AI spend.论证中大量段落依赖情绪化措辞(如"骗局""虐待式""侮辱"),而非补充数据支撑,可能让追求更冷静的企业 AI 成本收益分析的读者感到不适。
- No counter-scenario for genuine ROI cases未讨论真实存在正向 ROI 的场景The piece asserts no company has shown measurable AI ROI but doesn't engage with published case studies or benchmarks showing productivity gains in narrower, well-scoped use cases, weakening the universality of its claim.文章断言没有公司展示出可衡量的 AI 投资回报,但并未讨论已发表的、在范围明确的具体场景中显示生产力提升的案例研究或基准测试,削弱了其论断的普适性。
Worth reading for anyone evaluating enterprise AI spend or subscription-based AI tools, especially those who haven't scrutinized their own token consumption. The piece is polemical and light on audited financials, so treat its dollar figures as directionally suggestive rather than precise — but its core structural argument about subscription pricing vs. variable inference cost holds up independent of the author's rhetoric.
适合正在评估企业 AI 支出、或依赖订阅制 AI 工具却从未细算自己 token 消耗的读者。文章论战色彩浓厚、财务数据未经审计,其中具体金额宜作方向性参考而非精确数字看待——但剥离掉修辞外壳,其关于"订阅定价 vs 波动推理成本"的核心结构性论证依然站得住脚。
Excerpt原文节选
This is a short excerpt, not the full piece — the complete essay belongs to its original author; please read it in full at the link above.
以下仅为节选，并非全文——完整文章版权归原作者所有，请点击上方链接阅读全文。
If you liked this piece, please subscribe to my premium newsletter. It’s $70 a year, or $7 a month, and in return you get a weekly newsletter that’s usually anywhere from 5,000 to 18,000 words, including vast, detailed analyses of NVIDIA , Anthropic and OpenAI’s finances , and the AI bubble writ large . I recently put out the timely and important Hater’s Guide To The SaaSpocalypse , another on How AI Isn't Too Big To Fail , a deep (17,500 word) Hater’s Guide To OpenAI , and just last week put out the massive Hater’s Guide To Private Credit .
I also just did a piece about how OpenAI will kill Oracle , and I’ve used some of the materials in today’s piece. It's one of my best pieces I've ever done and I'm extremely proud of it.
Subscribing to premium is both great value and makes it possible to write these large, deeply-researched free pieces every week.
[…the source continues — read the rest at the link above]
[……原文更长，完整内容请点击上方链接阅读]
Good luck, Larry! You’re going to need it.
AI 的经济账根本算不通
作者：Ed Zitron 原文： AI's Economics Don't Make Sense
昨天早上，GitHub Copilot 用户终于得到了一个确认： 我一周前报道过的那件事 成真了—— 从 2026 年 6 月 1 日起，所有 GitHub Copilot 计划都将改为按用量计费（usage-based pricing） 。
以前，微软会给用户一定数量的“ 请求（requests） ”。现在，它要根据用户实际使用模型的成本来收费。微软把这称为“……朝着一个可持续、可靠、面向所有用户的 Copilot 业务和体验迈出的重要一步”。换句话说，用户每月订阅 GitHub Copilot 花多少钱，就得到等值的 token（词元，token）额度，比如每月 19 美元的套餐，就给你 19 美元的 token。
翻译一下 ：“ 我们不能再继续补贴 GitHub Copilot 用户的算力了，否则 Amy Hood 会拿棒球棍开始揍人。 ”
不管怎样，这份公告本身很有意思。它提前展示了这些涨价将会被包装成什么样：
Copilot 已经不是一年前的那个产品了。
它已经从编辑器里的助手，演变成了一个智能体式平台（agentic platform）。它能运行长时间、多步骤的编程会话，使用最新模型，并在整个代码库中反复迭代。智能体式使用正在成为默认方式，而这会带来明显更高的计算和推理（inference）需求。
今天，一个快速的聊天问题，和一次持续数小时的自主编程会话，可能让用户付出同样的价格。GitHub 一直承担了这类使用背后不断攀升的推理成本，但目前的高级请求模式已经不可持续。
按用量计费可以解决这个问题。它能让定价更好地对应实际使用情况，帮助我们维持长期服务可靠性，也减少我们限制重度用户的必要。
你看，问题并不是“ 微软一直在补贴将近 200 万人的计算成本 ”，而是“ AI 已经变得太强、太 powerful、太复杂了，所以它基本上已经是另一个产品了！ ”
也许 Copilot 的确已经不是“……一年前的那个产品”，但底层的经济错配并没有发生太大变化：微软连续…
[…the source continues — read the rest at the link above]
[……原文更长，完整内容请点击上方链接阅读]
祝你好运，Larry！你真的会需要它。