Apple poised to profit as AI bubble bursts, says critic Ed Zitron

Ed Zitron, who publishes the Where's Your Ed At newsletter and hosts the Better Offline podcast, has spent years warning that the AI buildout underlying recent tech valuations is economically unsustainable. In an extended interview, he laid out the core dysfunction: large language models burn tokens at a per-million rate regardless of whether the output is useful or whether the system accomplishes its assigned task. If a coding agent spins in a loop, the user still pays for every token consumed.

AI companies have responded by selling subscriptions priced far below the actual cost of tokens. According to SemiAnalysis, users can burn hundreds of dollars on a $20-a-month subscription and thousands on a $200-a-month tier, yet AI companies claim 70% gross margins on tokens—a claim Zitron disputes. The scale of the loss is evident in OpenAI's financials: the company lost $20.9 billion（約3.3兆円） on $13.07 billion（約2.1兆円） in revenue in 2025. Zitron argues this is the inevitable result of a business model fundamentally at odds with software economics, where consumers expect to pay a fixed fee for unlimited access.

When Anthropic and OpenAI shifted enterprise customers to token-based billing in March, the illusion collapsed. Uber disclosed that it had spent its entire annual AI token budget within a single quarter, with executives unable to justify the spend against tangible features delivered. Uber's COO stated it was "harder to justify" the cost of AI. Sam Altman acknowledged the issue as "huge" but declined to offer a solution. The problem extends across every AI-dependent startup: Perplexity, Cursor, and GitHub Copilot (which moved to token-based billing in June) are all unprofitable because customers will not pay the real cost of inference. Zitron notes that 89% of all AI revenues accrue to Anthropic and OpenAI, while other AI startups report "annualized revenue" (monthly revenue × 12) rather than actual trailing revenue—"because actual revenues are very depressing," even then totaling barely $100 million（約160億円） on an annualized basis for most.

The larger contagion risk stems from data center financing. An AI data center costs billions of dollars and takes 18 to 36 months to build. Hyperscalers have collectively spent over $1 trillion（約160兆円） in capex since 2022, with over $650 billion（約100兆円） designated for AI infrastructure in 2024 alone. These facilities are financed via project-based structures—special purpose vehicles (SPVs) backed by private credit funds, which themselves are funded by pension systems including CalPERS and the SF teachers fund. Zitron argues that when the bubble deflates, the money will be "basically gone." A bailout would require hundreds of billions in government intervention and would constitute "political cancer." He points to Oracle as a canary: with revenues stagnant for 20 years and $85 billion（約14兆円） in acquisitions merely keeping the company flat, Oracle has bet $340 billion（約54兆円）-plus (with hundreds of billions in debt) on the premise that OpenAI becomes the world's most profitable company by 2030. "Good luck on that one Larry," Zitron remarks.

The international fallout will be severe. Taiwan's TWSE and Korea's KOSPI are heavily reliant on ODMs (original design manufacturers) like Quanta and Hon Hai (Foxconn), which have seen revenues boosted by selling AI servers. A collapse in demand will depress their stock prices and harm pension investors holding these equities. Meanwhile, ordinary consumers are already paying the price. Hyperscalers have absorbed so much of the global memory supply that DRAM prices have roughly doubled in 2024. Tim Cook has characterized Apple's recent price increases as "unavoidable," with Macs and iPads already raised and iPhones expected to follow. Zitron concludes: "We are all paying more for stuff for effectively no reason other than that everybody in big tech has gone insane and wants to build as many data centers as possible."

Apple's position stands in sharp contrast. The company has spent only about $14 billion（約2.2兆円） on AI and outsourced compute-intensive workloads to Google, paying roughly a billion a year for Gemini to power Siri. It has not committed to massive data center buildout. Zitron argues that Apple Intelligence—the company's flagship AI feature rollout—was a "mass-radicalization of its users against AI," with barely functional add-ons, mocked summaries, and a Siri that was "somehow, even worse than the old Siri." This reception appears to have signaled to Apple's leadership that consumers do not want or need AI features, prompting the company to "pump the brakes" and treat AI as a commodity play rather than a differentiation vector. Despite media narratives claiming Apple is "falling behind," Zitron notes nobody can articulate what it is falling behind on or why it matters. If the bubble deflates as Zitron predicts, Apple will "sit on the sidelines and watch everything burn," while remaining in a position to acquire assets at distressed valuations. Zitron also highlights Apple's Vision Pro as a genuinely forward-looking bet—the most "interesting and future-forward thing" he has seen in years—though he laments its premature release by a departing CEO, noting that its current software update requires users to wear the device continuously, and that one slight movement throws the display out of focus. If Apple could make the Vision Pro weightless and invisible while solving focus tracking, Zitron argues, it would represent a genuine new interface—something the entire tech industry has lacked, driving the desperation behind the AI bubble itself.

Zitron's critique centers on a structural mismatch between how AI services are consumed and how they are priced. Large language models burn tokens at a per-million rate regardless of output quality or utility, yet consumers and enterprises have been conditioned to expect flat monthly subscriptions. SemiAnalysis found that users can burn hundreds of dollars on a $20-a-month subscription without seeing corresponding value. When Anthropic and OpenAI moved enterprise customers to token-based billing in March, the result was swift and damning: Uber exhausted its entire annual token budget in a quarter, and executives could not justify the cost against measurable features shipped. This pattern repeats across the AI startup ecosystem—89% of all AI revenues flow to Anthropic and OpenAI, while every other AI-powered company reports losses.

The cascade of losses is being masked by the sheer scale of hyperscaler capital deployment. Over $1 trillion（約160兆円） in capex has flowed since 2022, with over $650 billion（約100兆円） allocated to AI infrastructure in a single year. These data centers, funded by private credit vehicles backed by pension funds like CalPERS and the SF teachers fund, will need to generate roughly $1.5 trillion（約240兆円） in new profit—not revenue—to justify the spend. Zitron argues none of them will turn a profit, creating potential systemic contagion through the pension ecosystem and cascading losses in semiconductor suppliers (particularly Taiwanese companies like Quanta and Foxconn), Korean investors, and American hyperscaler stocks.

Apple's position is notably asymmetrical. It has spent roughly $14 billion（約2.2兆円） on AI and outsourced the compute-intensive work to Google, paying roughly a billion a year for Gemini integration. By contrast, Apple Intelligence—the company's headline AI feature—was poorly received and appears to have signaled to management that mass-market AI demand does not exist. This caution, combined with minimal data center exposure, positions Apple to emerge from a collapse relatively unharmed, potentially acquiring valuable assets at depressed valuations. The irony is that ordinary consumers are already subsidizing the buildout through higher hardware prices driven by memory scarcity caused by hyperscaler overbuilding.

No comments yet. Be the first to share your thoughts!