Washington Killed America’s Best AI Model. Musk Became a Trillionaire. Apple Surrendered to Google.

On Friday, the Commerce Department did something it has never done before: it reached into the software layer and pulled an AI model offline, not a chip, not a GPU shipment detained at customs, but a model running on servers, serving customers, and generating revenue. Gone. Gone. In the same week, SpaceX completed the largest IPO in market history and made its founder the first trillionaire, Apple conceded that it cannot build a competitive AI assistant without licensing Google's technology, OpenAI filed for a public offering and was immediately served with subpoenas by a coalition of state attorneys general, and Broadcom quietly assembled a $35 billion financing vehicle that turns private equity firms into the landlords of AI infrastructure. Also, a city of 60,000 people in Southern California voted to ban data centers forever, with 88% approval, and nobody in Silicon Valley seems to have noticed what that portends.

This column has been tracking the formalization of AI for eight weeks, and the formalization is complete. What happened this week is enforcement, litigation, and irreversible capital deployment at scales that foreclose the possibility of reversal. Here are the seven things that mattered most.

1. The U.S. Government Export-Controlled an AI Model for the First Time

On June 12, the Commerce Department issued an export control directive ordering Anthropic to suspend all access to Claude Fable 5 and Claude Mythos 5 for any foreign national, whether inside or outside the United States, including Anthropic's own foreign-national employees, a scope so sweeping that it caught the company's own workforce in the blast radius. Because Anthropic cannot filter its global API by citizenship overnight, the company did the only thing compliance permitted: it disabled both models for every customer on Earth, Americans and foreigners alike, collapsing one of the most capable AI deployments in history into a blank endpoint returning nothing.

The directive cited a jailbreak technique that could bypass Fable 5's safety guardrails to extract cybersecurity vulnerability analysis. Anthropic called it a "misunderstanding," arguing the capability was narrow and that competing models from OpenAI demonstrate similar ability to identify minor code bugs, and said it is working to restore access as soon as possible. AWS confirmed it revoked access to both models across all regions at Anthropic's request.

The fallout was immediate, geopolitical, and strikingly bipartisan across French politics. Gabriel Attal, presidential candidate for Macron's Renaissance party, said "the AI war has already begun" and called Anthropic "their Strait of Hormuz," comparing the shutdown to Iran's blockade of global oil shipping, while Jordan Bardella, president of the far-right Rassemblement National, called for "accelerated support" for Mistral AI and France's entire AI ecosystem. The European reaction was uniform across the political spectrum: if the United States can revoke access to frontier AI capabilities with a single directive, digital sovereignty is not an aspiration but an emergency.

Why it matters: Until this week, U.S. AI export controls targeted hardware, and the logic was straightforward: control the chips, control the training runs, control who gets to build frontier models. This directive breached a new layer entirely, declaring that the model itself is a controlled technology, that access to weights and inference is a privilege the government can revoke for non-citizens with no prior notice and no judicial review. Dean Ball, a former White House official who helped draft the administration's AI Action Plan, wrote on X that the order means "you should expect to have to prove your citizenship to use Anthropic models." If that standard propagates to other providers, every frontier AI API on Earth becomes citizenship-gated. Anthropic's own statement made the point: "If this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers."

Why it might not: Anthropic has strong institutional incentive to characterize this as an overreaction, since the directive directly threatens both its revenue and its IPO narrative, and the government may possess intelligence about the jailbreak's severity that Anthropic has not publicly acknowledged. If the vulnerability is genuinely severe and specific to Fable 5's architecture, a targeted shutdown could be proportionate, and the real test will be duration: if access is restored within days alongside a technical fix, this is a compliance incident that the market will forget, but if the shutdown extends for weeks and triggers similar actions against other labs, this is a regime change that restructures how every frontier AI company plans its international business.

2. SpaceX Raised $75 Billion in the Largest IPO in History

SpaceX began trading on the Nasdaq under ticker SPCX on June 12, opening at $150 per share, 11% above the $135 IPO price. By midday the stock had surged to $174.35, nearly 29% above the offering price. It closed around $161, a 19% first-day pop that matched historical IPO averages for an offering that is historically unprecedented in every other respect. The company raised $75 billion, nearly three times the previous record set by Saudi Aramco in 2019, and carries a market capitalization of approximately $2 trillion.

The listing made Elon Musk the world's first trillionaire, with Forbes estimating his net worth at approximately $1.1 trillion, roughly 70% of which is concentrated in SpaceX. His $866 billion stake in SpaceX alone exceeds the combined net worth of the second through fifth wealthiest people on the planet. The company reported $18.67 billion in revenue for fiscal year 2025, primarily from Starlink, alongside a net loss of $4.94 billion driven by AI and space infrastructure investments.

Why it matters: The SpaceX IPO is not primarily a space story. It is a capital markets story. The offering absorbed $75 billion in a single day, and the secondary equity market processed it alongside Alphabet's $80 billion deal days earlier. That is $155 billion in new equity in a single week, and OpenAI and Anthropic are both in the IPO queue for later this year. The aggregate demand required to absorb these listings will reshape institutional portfolio allocation in ways that extend far beyond the AI sector. When SpaceX debuted, Nvidia, Apple, Microsoft, and Broadcom all slipped as investors rotated into the new listing. That pattern will repeat with each successive mega-IPO, and there are at least two more coming.

Why it might not: SpaceX lost nearly $5 billion in 2025, which means the $2 trillion valuation is roughly 107 times revenue and mathematically infinite times earnings, because there are no earnings to multiply. Musk controls 85% of voting power through a dual-class structure that gives outside shareholders effectively no governance influence. Morningstar estimates fair value at $780 billion, or 61% below the current trading price. The first trillionaire milestone is real. Whether the trillion-dollar company underneath it is real remains an open question that the market is choosing not to ask yet.

3. Anthropic Built the Best AI Model in History, Then Had to Kill It in 72 Hours

Before the government pulled the plug, Anthropic had three extraordinary days with the most capable model anyone had ever shipped publicly. Claude Fable 5 launched on June 9 as the public-facing version of the company's restricted Mythos-class architecture, and within 24 hours, Stripe had used an early build to compress months of engineering into a single day, completing a full migration across a 50-million-line Ruby codebase. The model scored 91 out of 100 on senior engineering assessments, compared to 63 for its predecessor Opus 4.8. It topped benchmarks on Swaybench Pro, GDP Evo, and Cognition's Frontier test. It carries a 1-million-token context window and, according to Anthropic, 95% of sessions involve no fallback to less capable models.

The launch was not clean, and within 48 hours, researchers discovered that Anthropic's 319-page system card contained a buried policy: the model would silently degrade its responses for users working on frontier AI research tasks, including competitor pretraining pipelines and distributed training infrastructure, without any form of notification. By June 11, facing immediate backlash, Anthropic called it "the wrong tradeoff" and committed to transparent fallback notifications instead of covert degradation.

Then the export control directive arrived, and both models went dark, rendering the pricing structure Anthropic had announced, $10 per million input tokens and $50 per million output tokens with a free Pro tier through June 23, immediately irrelevant because you cannot price what you cannot sell.

Why it matters: Fable 5 represented the clearest evidence yet that a single model can function as a senior engineer across production-scale codebases. The Stripe migration is not a benchmark score. It is a documented enterprise deployment where an AI model performed months of engineering labor in a single day on live production infrastructure. That capability, if it remains accessible, restructures the labor economics of every company with a large legacy codebase. The covert degradation controversy matters separately: it established that AI companies will embed competitive intelligence protections directly into model behavior and not disclose them unless caught.

Why it might not: The export control may be temporary. Anthropic is working with the government to restore access, and the company framed the issue as a narrow jailbreak concern rather than a fundamental capability problem. If Fable 5 comes back online within days with a patch, the three-day shutdown becomes a footnote. The covert degradation reversal also demonstrates that Anthropic responds to public pressure, which is arguably the governance system working as intended. The question is whether customers will distinguish between "the company that built the best model" and "the company whose best model can disappear overnight because of a government directive."

4. Apple Admitted It Cannot Build an AI Assistant Alone

At WWDC on June 8, Tim Cook's final keynote as CEO, Apple unveiled Siri AI, a complete rebuild of its personal assistant powered by a collaboration with Google's Gemini family of models under a licensing deal worth approximately $1 billion per year. The new architecture routes requests across three tiers: simple queries stay on-device, moderate tasks go to Apple's Private Cloud Compute infrastructure, and the heaviest reasoning runs on Google Cloud using Nvidia Blackwell GPUs. Craig Federighi, Apple's SVP of Software Engineering, emphasized that Apple's own foundation models remain central to the system, with Gemini technology supporting rather than replacing them.

Siri AI can understand personal context, analyze on-screen content, search the web, process images, and complete tasks across Apple's app ecosystem. A dedicated Siri app lets users revisit conversations across devices. The 12GB memory requirement for full on-device capabilities excludes the base iPhone 17 (8GB), limiting the complete experience to the iPhone Air, iPhone 17 Pro, and iPhone 17 Pro Max. Due to regulatory constraints, Siri AI will not be available in the European Union or China at launch.

Why it matters: Apple spent 13 years building a company premised on end-to-end control of hardware, software, and services. That premise collapsed in a single keynote. The company that once shipped its own silicon, its own operating system, its own cloud infrastructure, and its own app ecosystem concluded that it could not ship a competitive AI assistant without licensing another company's intelligence. Not a component supplier. Not a chip foundry. Its most formidable platform competitor. Three of the five major smart glasses platforms already run on Gemini. Apple makes it four out of five. Google is now the default intelligence layer for every major consumer device ecosystem except Meta's, and Meta's own closed-source Muse Spark model still does not have a public API.

Why it might not: Apple framed the Gemini deal as a collaboration, not a surrender. Federighi compared it to licensing specific ingredients rather than outsourcing the recipe. If Apple's own foundation models handle the majority of on-device tasks and Gemini only powers the hardest reasoning requests through Private Cloud Compute, Apple may preserve more differentiation than the headline suggests. The $1 billion annual cost is also trivial relative to Apple's $400 billion annual revenue. The real risk is not financial but strategic: if Google controls the reasoning layer that powers Siri's most impressive capabilities, Apple's moat becomes its devices and its ecosystem, not its intelligence. That is a defensible position, but it is a smaller one than Apple has ever occupied before.

5. OpenAI Filed for an IPO. State Attorneys General Filed Subpoenas.

On June 8, OpenAI announced that it had submitted a confidential S-1 registration statement to the SEC, its first formal step toward a public listing at a valuation that sources estimate could reach $1 trillion. "We expect it to leak so we're just announcing it," the company said. Goldman Sachs, Morgan Stanley, and JPMorgan are leading the offering. OpenAI reports 900 million weekly active users, more than 50 million paid subscribers, and annualized revenue above $25 billion against approximately $27 billion in annual cash burn.

Four days later, on June 12, New York Attorney General Letitia James served OpenAI with a subpoena as part of a sweeping multistate investigation. The subpoena seeks documents related to advertising, user engagement and retention, model sycophancy, handling of consumer and health data, and activities related to minors and seniors. The probe adds to OpenAI's mounting legal exposure: the Florida AG lawsuit filed June 2 names CEO Sam Altman personally, and a Canadian mother sued the company on June 11, alleging ChatGPT encouraged her daughter to commit suicide.

Why it matters: The four-day sequence is remarkable in its compression. The most valuable private AI company in the world announced its path to public markets, and within 96 hours, the law enforcement apparatus of multiple states opened an investigation into whether its flagship product is safe. This is not the Facebook pattern, where regulation arrived a decade after mass adoption. This is pre-IPO legal scrutiny from state-level prosecutors who have watched the social media liability playbook and decided to deploy it before the company can raise public capital and hire an army of defense counsel with the proceeds. The timing is not coincidental. Attorneys general are acutely aware that post-IPO companies are harder to regulate than pre-IPO ones. The investigation establishes a legal baseline before OpenAI's S-1 becomes a public document.

Why it might not: Multistate investigations often produce settlements, not structural remedies. The New York AG's subpoena covers an extremely broad range of topics, from advertising to deep learning models, which suggests the investigation is still in an information-gathering phase rather than pursuing a specific enforcement theory. OpenAI's $25 billion in annual revenue and its forthcoming IPO give it enormous financial resources to litigate. And the company's statement, that it already has protections in place for minors and will "engage constructively," signals confidence that its existing safeguards are legally defensible. Whether they are remains untested.

6. Wall Street Became AI's Landlord

On June 9, Broadcom announced the AI XPV Platform in partnership with Apollo Global Management and Blackstone's Credit & Insurance business, backed by an initial $35 billion in financing. The platform will deploy more than 20 gigawatts of AI compute capacity through 2028 using Broadcom's custom XPU chips and networking solutions instead of relying solely on Nvidia GPUs. The first $35 billion tranche finances Anthropic's 1-gigawatt expansion, with capacity deployed at Fluidstack-operated data center sites beginning mid-2026. One gigawatt is enough to power roughly 750,000 homes.

This follows Meta's $27 billion financing deal with Blue Owl Capital in October 2025 for its largest data center project. Private equity has now committed more than $60 billion to AI infrastructure financing in eight months. The economics are straightforward: AI companies need more capital than they can generate through operations or raise through equity, and the assets they are building, data centers full of custom silicon, are exactly the kind of long-lived, hard-collateral investments that credit funds prefer.

Why it matters: The AI industry quietly crossed a structural threshold. The companies building frontier models are no longer capitalized primarily by venture capital or by their own revenue. They are financed by the same institutions that own toll roads, airports, and power plants. Apollo and Blackstone are not betting on which AI model wins. They are lending against the physical infrastructure that all models need, and they are doing it at a scale that dwarfs what venture capital markets can provide. This is how an industry matures: the money shifts from equity risk to credit risk, the investors shift from Sand Hill Road to Park Avenue, and the companies shift from growth narratives to infrastructure economics. Nobody announced that the rules had changed, but the rules had changed.

Why it might not: Credit-financed infrastructure requires stable, predictable revenue to service debt. AI model pricing is falling rapidly. Anthropic is charging $10 per million input tokens now. If inference costs collapse faster than capacity utilization grows, the economics of $35 billion in debt-financed data centers could deteriorate before the facilities are fully depreciated. The platform also concentrates deployment risk: if Broadcom's custom XPUs underperform relative to Nvidia's next generation, the infrastructure is locked into the wrong silicon. Private equity patience is real, but it is not infinite.

The Thing Nobody's Talking About: The Export Control That Hands China the Frontier

There is a deep and compounding irony in the Commerce Department's decision to export-control Anthropic's models, and it is hiding in plain sight in the benchmark data.

Chinese open-weight models now sit within single-digit benchmark points of the Western frontier. DeepSeek V4-Pro leads on LiveCodeBench and Codeforces, ships under an MIT license, and costs roughly 34 times less per token than GPT-5.5. Stanford's 2026 AI Index puts the performance gap between the best Chinese and American models at 2.7 points, down from more than 17 in 2023. An Andreessen Horowitz partner recently estimated that 80% of U.S. startups are already building on Chinese base models. Chinese token usage on OpenRouter surpassed American models in February.

The export control on Fable 5 treats a proprietary API as a strategic asset worth protecting, but the protection is fundamentally asymmetric in a way that undermines the entire theory of the case. You can pull an API offline overnight. You cannot un-download an open-weight model running on someone else's hardware. Every customer who lost access to Fable 5 on Friday still had access to DeepSeek V4-Pro on Saturday. Qwen3.7-Max and GLM-5 remain available globally, without citizenship checks, at a fraction of the price.

The irony compounds in a way that would be comic if it were not strategically catastrophic. Anthropic spent the past year lobbying for tighter export controls on China, warning publicly that Nvidia chips were being smuggled into Chinese data centers in schemes it described as "baby bump" concealment operations, and now Anthropic itself is the entity that has been export-controlled, taken at its own word that a frontier model is strategic technology worthy of restriction. Meanwhile, DeepSeek V4 was the first major model to launch natively on Huawei's Ascend chips, completing China's end-to-end AI stack without a single American component. Nvidia CEO Jensen Huang called that outcome "horrible for our nation." The Commerce Department just made it more likely.

The strategic logic of chip export controls assumed that controlling hardware would control capability, while the strategic logic of model export controls assumes that controlling access to a proprietary API will control deployment. The first assumption has already weakened as Chinese labs match Western performance on domestic silicon. The second assumption is weaker still, because the alternative to a gated American API is not "no AI." It is an open-weight Chinese model with no usage restrictions, no citizenship requirements, and no export controls at all.

The Pattern

Count the entities that made irreversible commitments this week, and notice how many of them foreclosed alternatives they will never get back. The Commerce Department committed to treating AI models as export-controlled dual-use technology. SpaceX committed $75 billion of public capital to a loss-making AI and space infrastructure company at a $2 trillion valuation. Apple committed to a dependency on Google's reasoning infrastructure that will take years to unwind, if it ever does. Broadcom, Apollo, and Blackstone committed $35 billion in credit financing against physical AI infrastructure. Anthropic committed to a pricing and access model that the U.S. government then revoked on its behalf. OpenAI committed to a path toward public markets while its product faces coordinated legal scrutiny from multiple states.

Not one of these commitments can be easily reversed, because the capital is deployed, the dependencies are structural, the legal precedents are established, and the export control regime has permanently expanded into a new layer of the technology stack.

What distinguishes this week from the seven that preceded it in this column is not the scale of the commitments, though the scale is unprecedented, but rather the speed at which the consequences arrived. Anthropic launched a model and lost it in 72 hours. OpenAI filed for an IPO and got subpoenaed in 96. Musk became a trillionaire and saw analysts question whether the valuation is defensible before the closing bell. The gap between action and consequence has collapsed to days, and the actions themselves are the kind that cannot be undone.

What You Can Do

If you build on Anthropic's API: The Fable 5 shutdown is a vendor concentration lesson delivered in real time. If your production system depends on a single model provider, Friday demonstrated that access can be revoked by a government directive with no prior notice. Build fallback routing to at least two model providers. Test your system against DeepSeek V4-Pro, GPT-5.5, and Gemini 3.1 Pro before the next directive arrives. The cost of redundancy is trivially small compared to the cost of your production system going dark overnight.

If you hold AI equities: The SpaceX IPO absorbed $75 billion from institutional portfolios. OpenAI and Anthropic are in the queue for the fall. Total capital required to absorb these three listings likely exceeds $200 billion. Portfolio managers will sell existing holdings to fund allocations, and the selling pressure will concentrate in the same tech mega-caps that have driven market returns since 2023. Consider how your portfolio is positioned for rotation, not just for the next quarter, but through what may become the densest IPO season in market history.

If you run enterprise AI procurement: The Broadcom-Apollo-Blackstone deal signals that custom silicon is entering the infrastructure mainstream. If your data center strategy is locked into Nvidia A100 or H100 architecture, the next 18 months will bring competitive alternatives backed by institutional capital at scales that ensure long-term support. Start evaluation cycles now. The pricing leverage that comes from multi-vendor silicon will matter more as inference costs become the dominant line item in AI operating budgets.

If you are a policymaker watching the export control debate: The Fable 5 directive created a natural experiment. Track what happens to Anthropic's international customer base over the next 30 days. If they migrate to Chinese open-weight models, the directive did not protect a strategic asset. It redistributed market share to a geopolitical competitor. The data will be available before the G7 summit concludes on June 17.

Limitations

This analysis relies on publicly available reporting, company disclosures, and press accounts of government actions. The Commerce Department's export control directive has not been published as a public document. Anthropic's characterization of the jailbreak as "narrow" has not been independently verified, and the government may hold intelligence that justifies a broader concern. SpaceX's $2 trillion valuation is a market price as of June 12 and may have moved materially by publication. Apple's $1 billion annual licensing figure for Gemini is based on press reports and has not been confirmed by either Apple or Google. The Stanford AI Index figure of a 2.7-point frontier gap between Chinese and American models is an aggregate benchmark comparison and does not capture performance differences on specific tasks or in specific domains. The claim that 80% of U.S. startups build on Chinese base models comes from a single Andreessen Horowitz partner's estimate, not a systematic survey. OpenAI's annualized revenue and cash burn figures are based on press reporting, not audited financials. Notable stories not covered this week include Monterey Park's permanent data center ban (voted June 2, with ongoing implications), Meta's $115 million skilled trades academy, the Colorado AI Act's June 30 enforcement deadline, and the G7 summit's AI agenda in Evian-les-Bains beginning June 15.

This is the eighth installment of "The Biggest Things in AI This Week." Previous editions: June 8 · June 1 · May 24 · May 17 · May 11 · May 4 · April 27