Summary

Apple’s U.S. App Store tollbooth has been judicially dismantled for now; Apple is going to appeal. Judge Yvonne Gonzalez Rogers found Apple willfully violated her 2021 injunction by charging 27% on external purchases, tracking users who bought up to a week after clicking out, and deploying “scare screens” to suppress conversion. She referred Apple and finance VP Alex Roman for potential criminal prosecution, concluding: “As always, the cover-up made it worse. For this Court, there is no second bite at the apple.”
The immediate beneficiaries are subscription, media, and creator businesses that can now communicate directly with iPhone customers and retain more revenue. Kindle has added a Get Book button, while Spotify and Patreon are implementing external-payment links; Roose’s hoped-for endgame is price competition that forces Apple to improve its own system and reduce fees. The remaining uncertainty is whether fraud rises without Apple’s controls, though Casey Newton called its “pristine, vigilant control” mostly fantasy, citing fake ChatGPT apps and a $10 Blue Prince copy.
Apple’s strategic vulnerability is growing even as its financial results remain formidable. Newton connected extractive developer economics to the abandoned car project, generative-AI stumbles, and weak Vision Pro ecosystem; Roose countered that Apple still reported $95.4 billion of quarterly revenue, up 5% year over year. The investable tension is between emerging “cracks in its armor” and a cash engine that has not yet paid a visible price for missing new platforms.
OpenAI’s announced restructuring preserves nonprofit control while removing the limits that made its economics genuinely unusual. Its LLC is slated to become a public benefit corporation controlled by the nonprofit, but investor and employee profit caps will disappear and the nonprofit will receive an equity stake. Alongside Fidji Simo’s appointment as CEO of applications, the restructuring supports Karen Hao’s contention that the labs are increasingly building products “that they can charge lots of money for,” even as Sam Altman says OpenAI “is not a normal company and never will be.”
Hao’s “empire of AI” thesis treats scale as a resource-extraction model, not a neutral technical inevitability. She traces frontier models to scraped work, low-paid annotation, labor automation, and permanent data centers consuming enormous power and water in communities promised mostly temporary construction jobs. In drought-stricken Uruguay and parts of Chile, she argues, the “benefit all of humanity” mission collides with people bearing extraction’s costs while receiving few of its purported gains.
The sharpest disagreement concerned whether general-purpose scaling is worth its externalities or even necessary for useful AI. Roose cited GPT-4 tutoring gains in Nigeria and AlphaFold; Hao rejected the premise that communities must surrender lithium or water to obtain those benefits, pointed—cautiously—to DeepSeek as evidence that constrained development can yield generality, and favored task-specific systems like AlphaFold. She could believe OpenAI might build systems appearing capable of automating most labor within two years, but called sweeping promises to cure cancer or climate change an elastic justification for capital and regulatory deference.
Italian brain rot could be an AI-native, crowdsourced entertainment ecosystem, with open remix economics and obvious commercialization risk. A 24-year-old Romanian creator’s Ballerina Cappuccina drew more than 45 million TikTok views and 3.8 million likes, while creators recombined absurd characters into affairs, pregnancies, airstrikes, and Ice Bucket Challenges. Newton saw the possible birth of crowdsourced entertainment—“if they can think it, they can visualize it”—while Roose immediately imagined Disney licensing the IP and turning it into seven movies.

Deep dive

1. Apple’s up-to-30% toll finally lost its shield against direct links

Newton’s consumer case begins with Kindle’s missing buy button: Apple’s commission made in-app book sales uneconomic, forcing readers into a browser, an Amazon login, and a purchase flow that should have been trivial.
Beyond charging up to 30%, Apple prohibited developers from telling users that Spotify, Netflix, games, or other services might cost less on the web because the seller could avoid Apple’s fee.
In 2021, Judge Yvonne Gonzalez Rogers found Apple had violated California unfair-competition law and ordered it to permit links to direct payment. Apple changed its rules, but Newton characterized the result as “malicious compliance”—the absolute minimum, delivered while “dragging and kicking and screaming.”

2. Apple’s compliance design became evidence against it

Apple imposed a 27% commission on externally completed transactions, only three points below its standard cut; once payment-processing costs were added, leaving Apple’s system could cost more than staying inside it.
The reach extended beyond an immediate checkout: if someone clicked from an iOS app and purchased on the web a week later, Apple could attribute the transaction and charge the developer its commission.
Apple also inserted a warning screen whose message Newton translated as: “Hey, loser, looks like you’re trying to do something stupid. You’re probably gonna die.” Disclosed emails showed senior executives, including Tim Cook, debating language designed to make the exit feel riskier and suppress clicks.
Rogers’s conclusion was categorical: “Apple willfully chose not to comply with this Court’s injunction,” intending to preserve an anticompetitive revenue stream. Believing the court would tolerate that “was a gross miscalculation,” she wrote; “the cover-up made it worse.”

3. Alleged courtroom deception raised the stakes beyond antitrust

Finance vice president Alex Roman testified that Apple did not know before January 16, 2024 what external-purchase fee it would charge and decided on 27% that day. Business records showed the plan’s main components had been determined in July 2023.
The judge referred Roman and Apple for potential criminal prosecution for perjury; the episode does not treat that referral as a conviction, but as the most extraordinary consequence of the company’s conduct under oath.
Roose was struck by internal project names such as Project Michigan and Project Wisconsin, which executives used while designing the fee so they could discuss the plan without making obvious that they were doing some sort of price fixing. His broader reaction: dominant platforms “did not end up that way by accident”; they worked deliberately to prevent competition and preserve market power.

4. Developers can now test whether direct commerce expands the market

Effective immediately under the order, Apple had to stop charging commissions on external links and updated its U.S. guidelines. Kindle added Get Book, while Spotify and Patreon began implementing the links they had long wanted.
The Kindle flow still opens a browser rather than completing the purchase natively, but Newton emphasized the larger change: businesses can again tell customers about their sites, prices, and offers without Apple restricting that communication.
Roose kept Apple’s strongest defense alive as an open question: removing controls might expose users to scams, malware, or users being taken advantage of. The alternative outcome is simpler—more choice and lower prices because developers no longer surrender 30% of revenue.
Casey doubted Apple’s protection story, recalling more than a dozen misleading ChatGPT apps when no official app yet existed and a $10 copy of Blue Prince uploaded on launch day. “They are not paying the attention to the App Store that they are telling you.”

5. Apple’s cash machine masks a weakening innovation bargain

Newton argued Apple treated developers like grateful tenants even though apps are a core reason people buy iPhones. Remove Amazon, Spotify, Patreon, and similar services, and customers would begin considering alternative devices.
He connected that imbalance to Apple’s broader misses: billions spent on an abandoned car, embarrassing AI walk-backs, and Vision Pro’s weak adoption, partly because developers see little opportunity to get rich building for it. These were “little cracks in its armor,” not a prediction of imminent collapse.
Roose supplied the counterweight: Apple’s latest quarter produced $95.4 billion in revenue, up 5% year over year. Newton replied that the ruling was never meant to impoverish Apple, only to shift a small portion of wealth toward developers and prove the disputed tactics were unnecessary.
Their preferred outcome is competition, not zero commission: external sellers charge materially less than 27%, Apple improves its payments, and developers regain leverage. Newton contrasted Microsoft, which never demanded 30% of every Windows software sale and thereby left room for its ecosystem to thrive.

6. OpenAI kept nonprofit control but normalized the upside

OpenAI abandoned a plan that would have separated its nonprofit from control of the commercial enterprise after objections from former employees, Elon Musk, others in AI, and attorneys general. Under the revision, the nonprofit remains in control.
The operating LLC is slated to become a public benefit corporation, formally responsible to the public as well as shareholders including Microsoft and SoftBank. The nonprofit receives a stake, while previous profit caps disappear and investors and employees gain unlimited upside.
Altman opened his employee letter with “OpenAI is not a normal company and never will be.” Yet the economics become more conventional, while board member Fidji Simo is leaving Instacart to serve as OpenAI’s CEO of applications over product and business divisions.
Hao said constant news can distract from enduring fundamentals. During her 2019 three-day embed—the first journalist profile after OpenAI became a company—she already saw idealism, openness, and freedom from commercialization yielding to secrecy and the competitive drive to reach AGI first.

7. Hao reframes AI ethics as accountability for concentrated power

OpenAI expected extensive access to produce a sympathetic narrative, Hao said, but it could not clearly articulate its vision, plan, or definition of AGI. Her critical profile was followed by three years in which the company stopped speaking to her.
The pivotal exchange concerned scaling’s environmental cost. Ilya acknowledged the concern but told her, “when we get to AGI, climate change will be solved”—a claim she saw as a “cop-out card” in which an undefined future achievement erases harms created along the way.
Hao prefers “AI accountability” to “AI ethics” because it names the power relationship: developers accumulate extraordinary resources by claiming they need them for AGI. Addressing present harms, she argues, is also how society prevents future ones rather than treating the two agendas as rivals.

8. The “empire” operates through data, labor, resources, and mission

Hao’s analogy to older empires is structural, while explicitly conceding that today’s companies are not comparably overt in their violence: both claim resources, design rules legitimizing the claim, exploit labor, and concentrate the resulting benefits at the center.
The claimed resources include artists’ and writers’ work, scraped online data, and contract labor used to annotate data and clean models. OpenAI’s stated goal of outperforming humans at most economically valuable work also makes AGI, in her formulation, “a labor automation machine.”
The legitimizing story is a modern civilizing mission: development supposedly benefits all humanity. Hao’s reporting instead found rural, poor, marginalized, and Global South communities absorbing extraction’s costs without receiving the promised share of “progress.”

9. Data-center promises conceal permanent demands and temporary jobs

Companies may enter communities through shell companies or sell projects as hundreds of millions of dollars of investment plus employment. Hao said the headline jobs are commonly construction roles; far fewer people remain to operate a completed data center.
The infrastructure then consumes enormous power and cooling water around the clock and is effectively permanent. In Chile, activists fought projects they feared would consume drinking water; Google chose to build in Uruguay during a drought when poorer residents faced contaminated water.
Governments accept the bargain hoping infrastructure will attract later investment—perhaps a Microsoft office and software-engineering jobs. Hao situated Chile’s response within centuries of extraction and its recent lithium role: opening natural resources to multinationals is repeatedly presented as a route to broad growth that often fails to arrive.
Roose’s pushback was concrete: a World Bank randomized trial reportedly found GPT-4-assisted tutoring significantly improved Nigerian students’ scores, with especially large gains for girls who had fallen behind. Could data-center or lithium access be exchanged for benefits such as free ChatGPT Pro?

10. Scale’s benefits do not settle whether scale is necessary

Hao rejected the premise that a country must surrender lithium or water to obtain educational gains. The current path—massive data, compute, and resources—is a choice, she argued, and constraining harms could force “true innovation” toward alternatives.
DeepSeek was her qualified specimen: she sees “a lot of” problems with it, yet says it demonstrated that resource-constrained development can still produce models with greater generality. The alternative path is not fully known because the dominant industry has barely explored it.
Personally, Hao regards today’s general-purpose models as “fruit from a poisoned tree” and rarely uses ChatGPT. She did use Google Reverse Image Search to identify OpenAI’s apparent $10,000 Brazilian designer chairs, illustrating her distinction between predictive tools and generative systems fed her unpublished research.
Roose, by contrast, uses generative AI to research topics such as AGI’s intellectual history and changing laboratory definitions—work he says fell from weeks to minutes. Hao declined to tell him to stop, calling AI itself a “perfect use case” because labs constantly stress-test their systems on that subject.

11. AGI rhetoric is broader than the products now being built

Hao believes many safety researchers are sincere; some described humanity’s possible demise with voices quivering. Her critique is that Silicon Valley’s wealth enables a narrow focus on distant catastrophe while a disproportionate share of resources bypasses harms already affecting people.
She could believe OpenAI might, within two years, produce systems that appear able to automate most labor, prompting executives to hire AI instead of workers. Other AGI forecasts require case-by-case scrutiny because the definition changes with the claim.
Newton’s practical AGI is a subscription replacing his assistant’s customer service, scheduling, and sales work. Hao accepted that definition but distinguished it from the sweeping system invoked to raise capital, rally support, resist regulation, “solve climate change,” and “cure cancer.”
Roose offered AlphaFold as evidence that AI investment can unlock science; Hao agreed it was remarkable precisely because it was task-specific, trained on curated data for a scoped problem. She called that path “perpendicular” to general scaling and argued engagement optimization, flattering chatbots, and the need to justify a $40 billion raise are pulling labs toward ordinary products.

12. Democratic AI governance has to cover the whole supply chain

Hao divided governance into data, compute, models, and applications, arguing that affected people—not only companies—need agency at every stage. Individuals should be able to opt in or out of datasets, with public consortiums debating what accessible data belongs in them.
Content moderation cannot remain an improvised internal choice, as when OpenAI employees debated whether pornographic images should stay in a dataset. Hao wants those value judgments subjected to open public discussion.
Communities should know data centers are coming, receive credible long-term water, power, and employment projections, and contest projects through local government. Contract workers should receive conditions consistent with international human-rights norms; simply being an American company does not make OpenAI’s development “democratic.”

13. Italian brain rot previews an open, AI-native entertainment market

Emerging in January, the format combines AI-generated creatures, faux-Italian names, and theatrical narration. Ballerina Cappuccina—created by a 24-year-old Romanian, not an Italian—surpassed 45 million TikTok views and 3.8 million likes.
Its compressed storytelling is the feature: one roughly 15-second clip moves from honeymoon to affair, pregnancy, discovery, pursuit, and airstrike. Another has Chimpanzini Bananini, a banana-chimp hybrid, nominate fellow characters for an Ice Bucket Challenge.
Roose saw a shift from universally visible memes toward siloed feeds where millions can follow a phenomenon he never encounters. Newton’s explanation joined novelty—“Fuck that. Give me Ballerina Cappuccina”—with young people’s perennial desire for a language older audiences cannot understand.
Newton’s bullish case is permissionless remixing: creators without animation skills can visualize an idea, reuse apparently untrademarked characters, and avoid the Bored Apes model of a “homeowner’s association for creating entertainment.” Roose’s bear case was immediate institutional capture: Disney licensing the characters and stretching Chimpanzini Bananini across seven movies.

Benchmark vs a16z: Why Stage Specific Firms Win Interactive Brokers: Margin Masters - [Business Breakdowns, EP.216]