Anthropic Filed for Its IPO. Microsoft Built Seven Models Without OpenAI. Florida Sued Sam Altman Personally.

Three companies are heading for the public markets at combined valuations exceeding $4 trillion. In the same seven days, Microsoft declared its AI independence from OpenAI, a state attorney general named a CEO personally in an 83-page lawsuit, the President signed an executive order treating AI as a national security tool, Congress introduced legislation to block states from regulating AI models, SoftBank pledged the largest AI infrastructure investment in European history, and a security firm confirmed the first documented case of an autonomous AI agent conducting a cyberattack without human direction. Meanwhile, Meta quietly cannot ship an API.

The pattern is unmistakable. Everything about AI is hardening into permanent infrastructure: the companies, the capital, the regulation, and the threats. Here are the eight things that mattered most.

1. Anthropic Files Confidential S-1, Setting Up the First Trillion-Dollar AI IPO

On June 1, Anthropic filed a confidential draft S-1 prospectus with the Securities and Exchange Commission. Days earlier, the company closed a $65 billion Series H at a $965 billion valuation, overtaking OpenAI's $852 billion for the first time. Annualized revenue has reached roughly $47 billion, up from $30 billion just weeks earlier, driven almost entirely by Claude Code's explosive adoption among enterprise developers who have made it the default coding assistant at companies from Goldman Sachs to Shopify.

The filing arrives in the middle of what may become the most consequential IPO season in market history. SpaceX priced its own offering at $135 per share on June 6, targeting a $1.77 trillion valuation and a $75 billion raise that would shatter Saudi Aramco's 2019 record. The shares are expected to begin trading on Nasdaq under the ticker SPCX on June 12. OpenAI has also signaled IPO intentions for later this year. Together, the three companies could raise $240 billion at a combined valuation north of $4 trillion.

For Anthropic specifically, the math is extraordinary. Its valuation more than doubled in four months, from $380 billion in February to $965 billion now. Revenue grew 57% in a single month. No private technology company in history has compressed that much absolute growth into that short a window.

Why it matters: The IPO race between Anthropic and OpenAI will determine which safety philosophy gets rewarded by public markets, a contest that carries stakes for every AI company that follows, because Anthropic built its brand on constitutional AI and cautious development while OpenAI built its brand on speed and consumer reach. Investors are about to place a dollar value on each approach, and the winner shapes how every subsequent AI company positions itself. When "responsible AI development" becomes a near-trillion-dollar investment thesis, the incentive structure for the entire industry shifts.

Why it might not: Valuation is not revenue, and revenue is not profit. Anthropic pays SpaceX's xAI division $1.25 billion per month in compute costs through 2029, a $15 billion annual commitment that consumes roughly a third of current revenue before a single dollar reaches operations. At these scales, the fastest-growing company in history also runs the fastest-growing cost structure. Public market investors will demand a path to positive unit economics that private investors never required.

2. Microsoft Build 2026: Seven Models, Zero OpenAI Data

At its annual developer conference on June 2-3 in San Francisco, Microsoft unveiled a family of seven in-house AI models that represent the clearest declaration of strategic independence from OpenAI in the partnership's history. Every model was trained from scratch without distillation from third-party systems, built entirely on commercially licensed datasets with no OpenAI data and no Anthropic data involved at any stage of the process.

The lineup: MAI-Thinking-1, a 35-billion-active-parameter mixture-of-experts reasoning model with a 256K context window. MAI-Code-1-Flash for coding tasks. MAI-Image-2.5 for image generation, now ranked second on Arena's editing leaderboard. MAI-Transcribe-1.5, which Microsoft claims is the world's most accurate transcription model across 43 languages. MAI-Voice-2, delivering natural speech synthesis with emotional control in 15 languages. All available through Azure Foundry, which now hosts more than 12,000 models.

MAI-Thinking-1 is the headline act, matching Claude Sonnet 4.6 in blind evaluations conducted by Surge and scoring 53% on SWE-Bench Pro, which puts a 35-billion-active-parameter model within striking distance of Claude Opus 4.6 on the benchmark that matters most to enterprise buyers. Microsoft also introduced Project Solara, an Android-based platform for agent-driven devices that companies like Best Buy, CVS, and Target are already exploring.

Then came the quantum announcement. Majorana 2, Microsoft's second-generation topological quantum chip, achieves qubit lifetimes of 20 seconds on average, with some lasting up to a full minute. That is a 1,000-fold improvement over Majorana 1. Notably, the chip was partly designed using Microsoft Discovery, Microsoft's agentic AI research platform, making it one of the first quantum processors where AI contributed to the scientific discovery process. Microsoft now targets commercially viable quantum computers by 2029, matching IBM's timeline.

Why it matters: Microsoft has been OpenAI's largest investor, its cloud partner, and its distribution channel since 2019. Shipping seven competitive models without any OpenAI involvement is not a supplement to that relationship. It is a hedge against its collapse. When your $13 billion investment partner is also your competitor in enterprise AI sales, building your own model stack is survival planning. Azure Foundry hosting 12,000 models gives Microsoft the platform play regardless of which model wins.

Why it might not: Barron's skepticism about Majorana 2 deserves attention. Microsoft retracted a peer-reviewed Majorana fermion paper in 2021 after independent researchers found omitted data. The Majorana 2 performance claims have not been independently verified. In quantum computing, history and hype rarely map onto each other. The MAI models are more immediately verifiable, but 53% on SWE-Bench Pro leaves significant ground between Microsoft and the frontier.

3. Trump Signs AI Cybersecurity Executive Order

On June 2, President Trump signed an executive order directing federal agencies to build an AI cybersecurity clearinghouse within 30 days and establishing a voluntary framework for AI companies to submit frontier models for government security review before public release. The order explicitly blocks mandatory pre-clearance requirements and limits the review window to 30 days, down from 90 in an earlier draft.

This is the most detailed federal AI security action since the Biden-era executive order was revoked. It directs the Department of Justice to prioritize prosecution of AI-enabled cyberattacks and asks the Department of War to deploy AI on classified networks. Eight leading AI companies have already agreed to participate. The White House fact sheet frames it as a cybersecurity measure, not a regulatory one.

Why it matters: The gap between the draft and the final text tells the real story. The administration started at 90 mandatory days. Industry pushed back. It landed at 30 voluntary days with an explicit carve-out blocking any future mandatory licensing requirement. That is not regulation. It is a framework the government can reference publicly while giving industry exactly what it requested. But the precedent matters: the federal government now formally acknowledges that frontier AI models are national security assets that warrant pre-release review, even if participation is optional today.

Why it might not: Voluntary frameworks have a long history of being exactly as effective as their least cooperative participant. If one major lab declines, the framework becomes a marketing tool for those who opt in rather than a meaningful security screen. The 30-day window is also short enough that a genuinely dangerous capability could clear review before evaluators fully understand what they are looking at.

4. Florida Becomes First State to Sue OpenAI

Florida Attorney General James Uthmeier filed an 83-page lawsuit against OpenAI and CEO Sam Altman on June 2, making Florida the first state to bring legal action against an AI company. The suit alleges OpenAI marketed ChatGPT as safe while knowing it posed risks to minors, citing the 2025 mass shooting at Florida State University where investigators found the suspect consulted ChatGPT on weapons, ammunition, and timing.

The complaint lists negligence, deceptive trade practices, and product liability violations. Uthmeier named Altman personally, alleging "utter disregard for the risk to human life caused by his firms' conduct." The state seeks billions in damages and a court order requiring OpenAI to overhaul its data collection practices for users under 13.

OpenAI responded that it has built "industry-leading protections and policies" for minors, including age-prediction tools and parental oversight features. The company did not dispute the FSU shooter's use of ChatGPT but pointed to improvements in newer model versions.

Why it matters: This is not a class-action ambulance chase. It is a Republican attorney general in a Republican state filing against the most prominent AI company in the world. The legal theory mirrors what cracked social media liability open: not that the technology was misused, but that it was designed in a way that made misuse inevitable. A Meta jury in New Mexico awarded $375 million on a similar structural argument in March 2026. If the Florida theory holds, every AI company with a consumer product faces the same exposure. Legal analysts say the complaint could serve as a template for other state attorneys general.

Why it might not: Florida's own legislative record on AI is friendly to industry. The state passed a law in 2025 explicitly classifying AI as a tool, not a legal person, and limiting platform liability. Uthmeier's lawsuit must navigate the tension between Florida's pro-innovation posture and the specific harms alleged. The suit also relies heavily on the FSU shooter's chat logs, which will face intense scrutiny over whether ChatGPT's responses were genuinely contributory or whether the platform was incidental to a premeditated act.

5. SoftBank Commits €75 Billion to France in Largest European AI Infrastructure Pledge

At the 2026 Choose France summit on June 1, SoftBank CEO Masayoshi Son announced plans to build 5 gigawatts of AI data center capacity in France, with an investment of up to €75 billion ($87.5 billion). The first phase commits €45 billion to 3.1 GW across three sites in the Hauts-de-France region: Dunkirk, Bosquel, and Bouchain. Schneider Electric will partner on the Dunkirk site, building an industrial cluster for data center power module manufacturing.

The broader Choose France summit attracted €93 billion in total investment commitments across 71 projects. Brookfield added €10 billion for AI infrastructure. The Emirati fund MGX and Bpifrance announced €7.5 billion for a second site. Salesforce pledged $2 billion.

Why it matters: France is not winning this investment on labor flexibility or deregulation. It is winning on electricity. The country operates one of the world's largest fleets of nuclear power plants, producing abundant, decarbonized, structurally surplus power. As AI data centers consume more energy and carbon border taxes loom, American and Asian hyperscalers have every incentive to build where the electrons are clean. SoftBank picked France the way an earlier generation of corporations picked manufacturing sites near ports: the resource is the strategy. Son said it plainly: "We can make France the centre of Europe for AI."

Why it might not: Mistral AI CEO Arthur Mensch offered the cold water. Energy accounts for only 10% of the AI value chain. The remaining 90% is software, models, data, and services, all of which are captured by companies that control the technology, not the countries that host the servers. France gets the data centers and 15,600 jobs. America keeps the model labs, the training runs, and the API revenue. If the value stays upstream, France is building the most expensive electric utility in European history.

6. First AI-Agent-Driven Cyberattack Confirmed in the Wild

Sysdig confirmed the first documented real-world cyberattack driven by an autonomous LLM agent. On May 10, an attacker exploited an exposed Python notebook, then handed operational control to an AI agent. The agent harvested cloud credentials, retrieved an SSH key from AWS Secrets Manager, and dumped an internal database. The credential-to-exfiltration chain took under two minutes. The full attack: four network pivots, eleven IP addresses, one hour from initial access to complete data theft.

No human directed individual steps after the handoff.

Why it matters: Security researchers have warned about AI-powered cyberattacks for years, but this is the first confirmed case in production. The speed is what should alarm defenders most. Traditional incident response assumes humans are making decisions, which means hours or days between initial compromise and data exfiltration. An AI agent compressed that to minutes. Existing security operations center playbooks are designed around human-speed attackers. Against agent-speed attackers, the detection window effectively vanishes.

Why it might not: One documented case does not constitute a trend, and the initial access vector was prosaic: an exposed Python notebook with bad credential hygiene. The AI agent made the attack faster, but the vulnerability would have been exploitable by a skilled human attacker using conventional tools. The real test is whether AI agents enable attacks on hardened targets that would otherwise resist human-directed intrusion. That has not been demonstrated yet.

7. Bipartisan House Bill Would Block States from Regulating AI Models

On June 4, Representatives Lori Trahan (D-MA) and Jay Obernolte (R-CA) released a discussion draft for federal AI legislation that would prohibit states from passing laws "targeting artificial intelligence model development." States could still regulate how AI is used but not how models are built, tested, or released. The bill would also create a federal agency with safety and transparency oversight and require developers of large foundation models to disclose training data sources, identified biases, and safety protocols.

The Information Technology Industry Council, representing the major tech companies, praised it. Public Citizen, a consumer advocacy group, called it a gift to an industry "that has repeatedly failed to pass meaningful AI protections" at the federal level. The group noted the bill does not address algorithmic discrimination, deepfake exploitation, housing or employment bias, or market concentration.

Why it matters: This is the opening move in the most consequential federalism fight since internet regulation. California, Connecticut, and Colorado have all passed or are advancing state-level AI laws. If this bill passes, those laws die on the model-development side. The argument for preemption is coherent: a patchwork of 50 different AI development standards would be operationally impossible for any company to navigate. The argument against preemption is equally coherent: the federal government has not passed a single binding AI safety law in the three years since ChatGPT launched.

Why it might not: It is a discussion draft released for stakeholder comment, not a bill introduced for a vote. Congress has struggled with AI legislation for years. The bipartisan framing is encouraging, but the specific mechanism of barring states from regulating model development while allowing them to regulate use creates a boundary that will be litigated endlessly. What counts as "development" versus "deployment"? Fine-tuning a foundation model for a specific use case sits on both sides of the line simultaneously.

The Thing Nobody's Talking About: Meta Can't Ship an API

Two months after launching Muse Spark, Meta's first closed-source AI model, the company still has not released a developer API. The Wall Street Journal reported on June 3 that Meta has pushed back the release multiple times with no scheduled launch date. Alexandr Wang, Meta's Chief AI Officer, posted on X in April that the API would arrive "soon." A Meta spokesperson told Reuters on June 4 that the company expects to release it "this month."

For open-source models like Llama, an API delay is an inconvenience. Developers can download the weights and run the model themselves. Muse Spark is closed-source. The API is the only access channel. Until Meta ships it, the model exists for the developer community in name only.

The context makes this more than a scheduling hiccup. Meta projected $125 to $145 billion in AI-related capital expenditure for 2026. It laid off roughly 8,000 employees in May, explicitly framing the cuts as resource reallocation from lower-priority functions into AI development. The implicit promise was that AI output would justify the human cost. Instead, the flagship output of that bet is a model that outside developers cannot touch. OpenAI and Anthropic generate meaningful revenue by selling API access. Meta's closed-source pivot was supposed to open that same monetization channel. The channel remains closed.

Muse Spark may turn out to be excellent when developers can finally use it. But the gap between announcement and availability reveals something important about the difference between spending on AI and shipping AI. Meta spent $14.3 billion to acquire Scale AI and hire Alexandr Wang. It has the money, the talent, and the infrastructure. What it does not yet have is the ability to convert those inputs into a product that works well enough to put in front of paying customers on a predictable schedule.

The Pattern

Last week's roundup ended with a note about formalization. This week doubles down on it. Anthropic is formalizing its path to public markets. Microsoft is formalizing its independence from OpenAI. The White House is formalizing its acknowledgment that AI models are national security assets. Florida is formalizing the legal theory that will be used to hold AI companies liable for harm. France is formalizing its bid to become Europe's AI infrastructure capital. Congress is formalizing the boundary between state and federal AI regulation. Sysdig is formalizing the threat model that security researchers have theorized about for years.

The experiments are over. The infrastructure is permanent. The capital commitments are measured in decades. The lawsuits name CEOs personally. The executive orders direct the Department of War.

The week that just happened did not feel like a turning point while it was happening. It never does. But the distance between "AI startup ecosystem" and "AI permanent infrastructure" collapsed in seven days, and the collapse was not gradual. It was institutional, deliberate, and irreversible.

What You Can Do

If you're an investor: The SpaceX IPO on June 12 will absorb enormous capital from the broader market. Jim Cramer predicts investors will sell Amazon, Microsoft, and Nvidia to fund allocations. Consider your portfolio's exposure to IPO-driven rotation before Thursday. The Anthropic IPO timeline is less certain but likely arrives this fall. Both demand different valuation frameworks than anything the market has priced before.

If you're a developer: Microsoft's MAI models on Azure Foundry are available now. If you are currently locked into a single model provider, this is the week to benchmark alternatives. The Muse Spark API delay means Meta's closed-source play remains theoretical. Plan accordingly.

If you run a security team: The Sysdig report should trigger an immediate review of exposed notebooks, credential hygiene, and secrets management. AI-agent-speed attacks compress the detection window from hours to minutes. If your incident response playbook assumes human-speed adversaries, it is already outdated.

If you build AI products for consumers: The Florida lawsuit establishes that a Republican attorney general in a business-friendly state is willing to name a CEO personally over AI harms to minors. If your product touches users under 18, your legal exposure just increased. Document your safety protocols, age verification methods, and parental controls now, not after the complaint arrives.

Limitations

This roundup relies on publicly available reporting and disclosed financial data. Anthropic's confidential S-1 has not been made public, so revenue and cost figures are based on analyst estimates and press reports, not audited financials. Microsoft's MAI benchmark claims have not been independently verified by third parties outside of Microsoft's own evaluation framework. The Sysdig cyberattack is documented by a single source, and independent corroboration has not been published as of this writing. SpaceX's $1.77 trillion valuation is an IPO price, not a market-determined value; post-listing trading could settle at a materially different number. The SoftBank commitment is a pledge, not a disbursement, and large-scale infrastructure announcements frequently underdeliver on timeline and scale. This analysis does not cover every significant AI development from the week; notable omissions include the Grok deepfake lawsuit expansion, ICRA 2026 embodied AI competition results, and the OpenAI ChatGPT "superapp" reorganization reported by the Financial Times.

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