Apple Will ‘Watch Everything Burn’ When the AI Bubble Bursts

当人工智能泡沫破裂时，苹果将“冷眼旁观一切化为灰烬”

Apple Will ‘Watch Everything Burn’ When AI Bubble Bursts - Ed Zitron 当人工智能泡沫破裂时，苹果将“冷眼旁观一切化为灰烬”—— Ed Zitron

Monday July 27, 2026 7:11 am PDT by Tim Hardwick 2026年7月27日，周一，太平洋夏令时间上午7:11，作者：Tim Hardwick

Memory prices have doubled, Macs and iPads have gone up, and iPhones are expected to follow. Ed Zitron – who writes the Where’s Your Ed At newsletter, hosts the Better Offline podcast, and has been described by Politico as the AI boom’s most “acerbic gadfly” – has spent years arguing the buildout driving those costs will never pay for itself. We asked him what happens to Apple if he’s right. 内存价格翻了一番，Mac 和 iPad 的价格已经上涨，预计 iPhone 也将紧随其后。Ed Zitron 撰写了《Where’s Your Ed At》时事通讯，主持了《Better Offline》播客，并被《Politico》杂志描述为人工智能热潮中最“尖刻的牛虻”。多年来，他一直认为推动这些成本增长的基础设施建设永远无法实现收支平衡。我们询问了他：如果他是对的，苹果将会面临什么？

You’ve been calling AI a bubble since before it was fashionable. For MacRumors readers who mostly know it as ChatGPT or Apple Intelligence on their iPhone, what, in plain terms, is actually broken about the economics of the LLM industry? 在人工智能成为潮流之前，你就一直称其为泡沫。对于那些主要通过 iPhone 上的 ChatGPT 或 Apple Intelligence 来了解人工智能的 MacRumors 读者来说，用通俗的话说，大语言模型（LLM）行业的经济模式究竟哪里出了问题？

At their very core, Large Language Models’ costs run contrary to basically every model of selling software. Consumers and enterprises alike have been trained to pay a monthly fee for a service, and while these services might have limits or strictures, basically nobody buying software expects to have a metered service, let alone one that’s both metered and with hard to measure costs. 从本质上讲，大语言模型的成本与几乎所有的软件销售模式背道而驰。消费者和企业都习惯了为服务支付月费，虽然这些服务可能有一定的限制或约束，但基本上没有购买软件的人会预期使用一种“按量计费”的服务，更不用说这种服务既是按量计费，其成本又难以衡量。

LLMs burn tokens at a per-million rate regardless of whether or not you get the response you want, or whether it does what you ask it to do. If you ask a coding agent to do some sort of software task and it goes off and spins its wheels in a loop, you’re paying for the tokens regardless. 无论你是否得到了想要的回复，或者它是否完成了你要求的任务，大语言模型都会以“每百万个 Token”的费率消耗资源。如果你让一个编程代理去执行某种软件任务，而它陷入了死循环，你依然要为这些消耗的 Token 付费。

AI companies knew that consumers would never pay the actual cost of their AI services, so they have, for the most part, sold them monthly subscriptions with vague rate limits that allow them to burn way more in tokens than the cost of their subscription. SemiAnalysis found that you can burn hundreds of dollars on a $20-a-month subscription and thousands of dollars on a $200-a-month subscription, and while AI boosters will claim that these companies have “70% gross margins on tokens,” there is little proof that this is the case, and my own reporting shows that OpenAI lost $20.9 billion on $13.07 billion in revenue in 2025. 人工智能公司很清楚，消费者永远不会支付其人工智能服务的实际成本。因此，在大多数情况下，他们以模糊的速率限制出售月度订阅，这使得用户消耗的 Token 价值远超订阅费。SemiAnalysis 的研究发现，你可以在每月 20 美元的订阅中消耗数百美元，在每月 200 美元的订阅中消耗数千美元。虽然人工智能的支持者声称这些公司在 Token 上拥有“70% 的毛利率”，但几乎没有证据支持这一点。我自己的报道显示，OpenAI 在 2025 年营收 130.7 亿美元的情况下，亏损了 209 亿美元。

(Image credit: SemiAnalysis) （图片来源：SemiAnalysis）

This means the very basic economics are broken. If Anthropic and OpenAI believed customers would actually pay the real cost of AI tokens, they wouldn’t have to give away 20 to 40 times the amount of tokens to subscribers. Meanwhile, back in March of this year, both moved their enterprise customers over to token-based billing. Within a few weeks, it came out that Uber had spent its entire annual token budget in the space of a quarter, and its COO said that it was getting “harder to justify” the cost of AI because it was hard to track the cost of AI to any actual useful features shipping. Sam Altman would eventually say it was a “huge issue” but declined to say how it might be fixed. 这意味着最基本的经济逻辑已经崩溃。如果 Anthropic 和 OpenAI 相信客户真的愿意支付人工智能 Token 的实际成本，他们就不必向订阅用户赠送相当于订阅费 20 到 40 倍的 Token。与此同时，今年 3 月，这两家公司都将企业客户转为基于 Token 的计费模式。几周内就有消息传出，Uber 在一个季度内就花光了全年的 Token 预算。其首席运营官表示，人工智能的成本变得“越来越难以证明其合理性”，因为很难将人工智能的成本与任何实际交付的有用功能挂钩。Sam Altman 最终承认这是一个“巨大的问题”，但拒绝透露如何解决。

This is a problem across basically every single AI-powered startup, which has to pay the per-million token rate. Perplexity, Cursor, GitHub Copilot (which moved to token-based billing in June) – every single AI startup is unprofitable because their users don’t want to pay the actual cost of AI. 这几乎是所有人工智能初创公司面临的共同问题，因为它们都必须支付“每百万 Token”的费用。Perplexity、Cursor、GitHub Copilot（6 月份转向了基于 Token 的计费）——每一家人工智能初创公司都在亏损，因为用户根本不想支付人工智能的实际成本。

Another issue is that AI services are just not that useful or differentiated. While people get some sort of benefit out of AI-generated code, these tools actually end up making them slower, and are filling codebases full of slop. Otherwise, an LLM is an LLM is an LLM – it can generate, it can summarize, it can search, and that’s about it, which means that every AI service is effectively the same. That’s why 89% of all AI revenues are Anthropic and OpenAI, and why every AI startup talks in terms of “annualized revenue” (monthx12) – because actual revenues are very depressing. Even then, most are barely at $100 million annualized. 另一个问题是，人工智能服务并没有那么有用，也没有什么差异化。虽然人们能从人工智能生成的代码中获得一些好处，但这些工具实际上最终让他们的工作变慢了，并且让代码库充斥着垃圾内容。除此之外，大语言模型本质上都一样——它们能生成、能总结、能搜索，仅此而已，这意味着每一项人工智能服务实际上都是相同的。这就是为什么 89% 的人工智能收入都集中在 Anthropic 和 OpenAI 手中，也是为什么每家人工智能初创公司都用“年化收入”（月收入 x 12）来谈论业绩——因为实际收入非常惨淡。即便如此，大多数公司的年化收入也勉强达到 1 亿美元。

Then there are the data centers. An AI data center is very, very expensive to build, takes 18 to 36 months, and costs billions of dollars, which means effectively anyone building one will be raising debt and only get paid once a customer moves in… except there aren’t really any customers for AI data centers outside of Anthropic and OpenAI, both of whom are so unprofitable that they’ve had to raise hundreds of billions of dollars even when Microsoft, Google and Amazon built all their infrastructure. The only reason everybody isn’t freaking out about this is because AI-related stocks have done well, even though none of the hyperscalers actually share their AI revenues. 还有数据中心的问题。建设一个人工智能数据中心非常昂贵，需要 18 到 36 个月的时间，耗资数十亿美元。这意味着任何建设数据中心的人实际上都是在举债，只有在客户入驻后才能获得回报……但问题是，除了 Anthropic 和 OpenAI 之外，人工智能数据中心几乎没有其他客户。而这两家公司本身就亏损严重，以至于即便微软、谷歌和亚马逊已经为它们建设了所有基础设施，它们仍不得不筹集数千亿美元。大家之所以没有对此感到恐慌，唯一的原因是人工智能相关股票表现良好，尽管没有任何一家超大规模云服务商真正披露过它们的具体人工智能收入。

You’ve argued that AI’s demand story is essentially a mirage – that most of the data center capacity is being absorbed by OpenAI and Anthropic themselves, which is masking the absence of real enterprise demand. If that’s right, who do you think will actually bear the cost when the whole thing unravels? 你曾指出，人工智能的需求故事本质上是一个海市蜃楼——大部分数据中心容量都被 OpenAI 和 Anthropic 自己消耗掉了，这掩盖了真实企业需求缺失的事实。如果这是对的，你认为当这一切崩盘时，谁将真正承担成本？

Honestly, it’s going to be a lot of private credit funds, because they’re the ones funding the data centers, and they’re funded by pension funds like the SF teachers fund or CalPERS, which makes me really, really worried about the systemic contagion. People will argue that this means there’s going to be a bailout, but this isn’t really a bailoutable thing. These data centers are funded by project financing, which means that the money is basically gone and the only way to “make them whole” would be to either… 老实说，这将是许多私人信贷基金，因为它们是数据中心的资金来源，而这些基金又是由旧金山教师基金或加州公务员退休基金（CalPERS）等养老基金资助的，这让我对系统性传染感到非常、非常担忧。人们会争辩说，这意味着政府会进行救助，但这并不是一个可以轻易救助的事情。这些数据中心是通过项目融资建立的，这意味着钱基本上已经花光了，而要“弥补损失”的唯一方法要么是……