Executive Summary
Within two weeks, Anthropic and OpenAI filed confidential IPO paperwork, and SpaceX completed the largest IPO ever recorded. All three are among the most valuable private companies in tech. During the same period, two governments acted directly on AI products. The US ordered Anthropic to disable one of its models, and the EU appointed the experts who will help enforce its AI law. Both were direct decisions about specific products and obligations.
These companies are raising hundreds of billions, and many of them are also investing in each other. That makes it harder to measure how much of the demand is real.
The trillion-dollar filing season
Anthropic filed first. Days after a $65 billion funding round valued it at $965 billion, it submitted confidential IPO paperwork, which put it ahead of OpenAI for the first time. It also reported a revenue run rate of about $47 billion by late May, up from roughly $9 to $10 billion at the end of 2025, and said it expects its first profitable quarter. OpenAI filed on June 8, confirmed it in a blog post, and was last valued at $852 billion. SpaceX then completed the largest IPO in history, raising about $75 billion at an opening valuation near $1.77 trillion. The stock rose about 19 percent on its first day, lifting Elon Musk's paper wealth above $1 trillion. Together, the three represent roughly $3.6 trillion in possible listings.
The run rate figure needs care. It annualizes recent revenue, so the $47 billion is a single recent month scaled up to a full year. Anthropic said the expected profit may not continue, partly because it includes a discounted deal for computing power. SpaceX is also unprofitable, with a quarterly loss in the billions, and its share price fell back within a few days of the early gain. Investors have few chances to buy into these specific companies, and that scarcity explains much of the valuations. In an earlier keynote, I described AI spending moving from software toward physical things like power grids and data centers. Going public follows from the cost of that buildout, which private investors can no longer fully cover.
Sakana AI, a Tokyo lab founded by former Google researchers Llion Jones and David Ha, released a model called Fugu. It does not work like the large single models. Fugu takes a task, passes it to a group of other models, and combines their output, while appearing as one model to the user. Sakana's benchmarks place its strongest version, Fugu Ultra, above Anthropic's Fable 5 and Mythos models on coding, reasoning, and science tests. Testers outside the company were more cautious, reporting that it is slow, expensive, and weaker than Fable in normal use, though good at reviewing code.
The benchmarks come from Sakana, and independent tests so far do not confirm the strongest claim. It still matters for the IPOs, because their valuations assume Anthropic and OpenAI keep their lead, and a credible competitor weakens that assumption. AI share prices react strongly to small pieces of news. Michael Burry has used similar reasoning to call the sector a bubble, holding put options against AI companies and arguing that some of them understate how quickly their chips lose value, which would make future profits look higher. His concern is about accounting. Another problem with judging the demand is that some of it comes from companies buying from each other.
Buying from your own investors
Bloomberg documented a set of deals linking Nvidia, OpenAI, Microsoft, Oracle, AMD, and CoreWeave, worth several hundred billion dollars together. Most follow the same structure. A chip or cloud company invests in an AI company, and that AI company then buys chips or cloud services from the same investor. Nvidia holds a stake of about $30 billion in OpenAI, which has agreed to buy large quantities of Nvidia hardware. Amazon has invested up to $25 billion in Anthropic, which has committed more than $100 billion to Amazon's cloud over ten years. Supporters, including Janus Henderson and Anthropic's Dario Amodei, call it a reasonable way to secure scarce computing capacity. Critics say it makes demand look larger and more independent than it is, because part of the buying is funded by the seller.
So far, regulators are examining these deals without blocking them. There is no confirmed SEC investigation. Most of the pressure comes from Congress, where Senator Elizabeth Warren has asked financial regulators to assess the risk that more than $1 trillion of AI-related debt could pose to the wider economy. She also invited Nvidia's Jensen Huang to testify in June, though that invitation concerned chip sales to China rather than these investments, and Huang declined, offering instead to meet senators at Nvidia's offices. Huang has separately said his roughly $30 billion in OpenAI is probably his last private investment before these companies list. Some of the revenue that AI companies report comes from other firms spending money those same firms provided in the first place.
Google placed a large chip order with Intel during the same period. According to The Information, it ordered more than 3 million of its own TPU chips for production through 2028. Intel's contract manufacturing business has struggled for years, and the order arrived while TSMC is short of capacity for AI chips. Intel's stock rose more than 11 percent on the news. Several analysts noted that Intel may only handle packaging while the chips are still made at TSMC. Nvidia is evaluating Intel's technology as well but has not placed an order.
In an earlier keynote, I described the shortage of advanced chips, which is the background to this order. Because supply is limited, one large order can move a stock like Intel by double digits in a day. The reaction probably exceeded what is confirmed, since the hardest and most valuable step, manufacturing the chips, may stay with TSMC. The Intel order is different from the circular deals. Google pays Intel for a service it needs, no money returns to Google, and it reduces Google's dependence on a single supplier. The circular deals involve a smaller group of firms connected closely enough that their investments and their revenue overlap.
Washington and Brussels step in
On June 12, the US Commerce Department, under Secretary Howard Lutnick, ordered Anthropic to cut off access to its two newest models, Fable 5 and Mythos 5. This was the first use of US export controls on an AI model itself rather than on chips. Anthropic disabled both worldwide within hours, including for its own staff who are not US citizens. More than 80 security researchers signed an open letter arguing the controls remove useful tools from people defending systems.
The order has a clear inconsistency. The US treats these models as too dangerous for other governments to hold, while keeping them available for its own use. If a capability is risky enough to withhold from foreign states, it is not clear why it is safe in the hands of the government making that decision. More oversight of this technology seems reasonable to me. The problem is that it is happening one country at a time, each with its own reasoning, while the models do not stop at borders. International rules agreed between countries would address the risk more directly. National bans applied one at a time mostly change who can use a model.
During June, the EU worked on enforcing the AI Act it had already passed. On June 1, the Commission appointed a scientific panel of 60 independent experts and a larger advisory forum. Through a package called the Digital Omnibus, it delayed the heaviest obligations for high-risk systems from August 2026 to December 2027, while keeping most transparency rules on schedule for August 2026. Planned bans on AI-generated, non-consensual intimate images and on child abuse material are due later in the year, with the exact date not yet fixed. No fines or formal cases have been opened.
The EU still intends to enforce the law. It built the rules faster than the staff and bodies to apply them, and faster than companies could prepare, so it is postponing the harder deadlines while it catches up. Washington acted in a single day by disabling a model. Brussels is taking months to build the staff that will eventually check whether companies comply.
Two exits, 250 billion gone
Google lost two well-known researchers in the same period. Noam Shazeer, a co-lead on Gemini whom Google paid a large sum to bring back in 2024, left for OpenAI. John Jumper, the DeepMind scientist who won a Nobel Prize for AlphaFold, left for Anthropic. On June 22, after both moves, Alphabet's stock fell about six percent and lost around $250 billion in value, though wider concerns about AI competition contributed to the drop.
A loss of that size from two departures reflects how much individual researchers now matter. Companies used to compete mainly on chips and data centers. They now compete just as hard for a small number of researchers. Reported offers have reached amounts that would have seemed unrealistic a few years ago, with signing packages discussed in the tens and sometimes hundreds of millions of dollars, mostly in stock that vests over several years. There are few other fields where individual technical staff are valued this highly by an employer. Against a single day loss of $250 billion, a nine-figure offer for one researcher is a small amount.
Two questions for next quarter
The largest AI companies are going public because private funding can no longer cover the cost of compute. A significant share of that funding also moves among the same firms, which makes real demand hard to measure. During the same weeks, governments started acting directly on AI, the US by disabling a model and the EU by setting up enforcement for its law. The questions worth watching are whether investors price these companies on actual earnings or on the fear of missing them, and whether the financing between the firms holds or comes apart.