The Week Washington Bought a Stake in Your Chatbot

The Government Stopped Pretending

For three years the working fiction was that the U.S. government was a neutral observer of the AI industry, occasionally issuing strongly worded letters and convening listening sessions. That fiction died this week, quietly, in a series of stories that should be read together rather than apart.

Anthropic is negotiating directly with the Trump administration to remove restrictions on its top models. The NSA, it turns out, had lost access to Claude during the dispute, which is the sort of detail you only learn about after the fact, when someone in a windowless building decides the leak is useful. A startup that builds on Claude is now suing the federal government over an order blocking foreign nationals from accessing advanced models. The administration also asked OpenAI to stagger the release of its next model, and OpenAI complied, shipping GPT-5.6 only as a limited preview to vetted government users.

Read that again. The most capable publicly known model in the United States debuted this week not on a product page but inside a federal vetting program. Whatever you thought the relationship between Washington and the frontier labs was, it is now closer than that. The labs are not being regulated. They are being onboarded.

Anthropic, for its part, accused Alibaba of stealing Claude through fake accounts, which is the kind of accusation that, true or not, conveniently reframes the company as a national asset under foreign attack. That framing is now load-bearing for the entire deregulation pitch.

The Bubble Has a Central Banker Now

While Washington was integrating the labs into the national security apparatus, the Bank for International Settlements, which exists specifically to worry about things, published a warning that the AI bubble will pop and take the global economy with it. The BIS does not write these things casually. When the bank of central banks tells you the returns are weak and a funding pullback could threaten worldwide financial stability, the correct response is not to argue. It is to look at where your money actually is.

If you are an Australian retiree, the answer this week was that up to 12 percent of your balanced superannuation fund was sitting in tech and AI stocks, a fact most retirees did not know and did not consent to in any meaningful sense. The exposure is structural now. The S&P is the AI trade. Your pension is the AI trade. The sovereign wealth fund of whatever country you live in is, with high probability, the AI trade.

SpaceX provided a small preview of what fragility looks like. Days after a $25 billion debt offering, its bonds traded at junk-rated yields and the company shed roughly $400 billion in market value. SpaceX is not an AI company in the strict sense, but it is a Musk-adjacent megacap funded on the same vibes, and the vibes wobbled. Bonds price reality faster than equities. Worth remembering.

Meanwhile the IMF warned that an AI wealth boom could accelerate global inflation, on the theory that consumers feeling richer spend more. So the menu of outcomes, according to the official institutions paid to model these things, is roughly: the bubble pops and drags down the world, or it doesn't pop and the resulting wealth effect makes everything more expensive. Pick a hand.

The Guardian put it less diplomatically, describing investors and tech firms as collaborating to delay an inevitable reckoning. OpenAI slowing model releases while the industry pumps profit narratives is not a coincidence. It is choreography.

Energy, Chips, and the Physical Footprint

OpenAI's custom chip program made the news this week with a number attached. The design would consume 10 gigawatts. That is the steady-state electrical demand of millions of households, or a small nation, depending on which small nation you pick. This is for one company's next-generation training and inference stack. Multiply by the number of frontier labs and you start to understand why every utility executive in America is suddenly an AI analyst.

The physical AI economy is now competing with residential power, with industrial power, with decarbonization timelines, and with the basic premise that the grid can be planned at all. None of this shows up in the chatbot UI. You type a prompt, a small nation's worth of electricity rearranges itself, you get back a paragraph about dog breeds. The abstraction is the product.

This is also the context in which to read the executive order accelerating the deadline for quantum-safe cryptography. The government is moving on multiple compute frontiers at once because it has concluded, correctly, that the next decade's strategic advantage is a function of who controls silicon, electrons, and math. The civilian internet is along for the ride.

Surveillance Creep, Domestic Edition

The domestic surveillance file got thicker this week without anyone particularly noticing. The Met Police in London extended its Palantir trial by another twelve months while it searches for a permanent supplier, which is the British way of saying the pilot is the policy now. Google Home expanded facial recognition so it can identify you when you are facing away from the camera, a capability that exists for no consumer reason and several non-consumer ones. Google was also reported to be quietly harvesting images uploaded during search interactions to train its models, opt-out rather than opt-in, as is tradition.

Meta provided the dark comedy. Its keystroke-monitoring program, which collected employee keystrokes ostensibly to train internal AI, suffered an internal leak that exposed the sensitive data it had collected. Meta then paused the program. The lesson the company appears to have drawn is that surveilling employees is fine as long as the surveillance data does not leak. The lesson available to everyone else is that any dataset assembled will eventually be exposed, and that the question is never whether but when.

YouTube settled a teen addiction lawsuit with a 15-year-old plaintiff this week, with three similar cases against Facebook, TikTok, and Snapchat heading to trial next month. The settlement is the platforms' tell. They would not pay if they thought they would win, and they would not pay quietly if they thought the next three cases were defensible. The recommender systems are about to have a very expensive summer.

Lobbying as a Product Category

AI companies are now spending heavily on congressional races through super PACs, with a Manhattan seat receiving particular attention. The strategy is legible. You concentrate resources in a small number of competitive races, you signal to every other member of Congress what happens if they cross you, and you do this with money raised from investors who, two years ago, were funding a research lab. The pipeline runs from venture capital to model weights to electoral politics in one continuous flow.

The New York Times amended its lawsuit against OpenAI to add allegations that Microsoft pushed OpenAI to train on copyrighted material. Whether or not that survives discovery, it reframes the copyright fight from a question of negligent scraping to a question of corporate strategy. Microsoft is a more attractive defendant than a research lab. It has lawyers, but it also has a market cap that makes settlement attractive. The Financial Times reported a broader wave of AI ethics litigation, which is the polite term for the avalanche of cases now being assembled by firms that have figured out the addressable market.

And in Hollywood, Google DeepMind dropped $75 million into A24, the indie studio whose entire brand is being the alternative to algorithmic content. The fans revolted, which was the predictable reaction, but the deal will close anyway, which is the actual story. Cultural production is being absorbed into the same capital stack as model training. The pipeline from chips to scripts is shorter than it used to be.

The Humans Are Still Around, For Now

Dave Eggers used the week to argue, in The Guardian, that outsourcing thinking and writing to machines is a marker of species-level decline. He is, in the strict sense, correct, though species-level decline is a long arc and the immediate concerns are smaller and more embarrassing. An AI rental platform in New York was caught advertising apartments that do not exist, a category of fraud that did not require generative AI but is now scalable in ways that will make every renter's search measurably worse. The Vision Pro chief at Apple defected to OpenAI, continuing a talent flow so consistent it should be drawn on a map.

A fatal Tesla crash in Texas drew a federal investigation, which is a sentence that has been written so many times the editor probably has a keyboard macro for it. The driver-assistance system has now killed enough people to support a small subgenre of regulatory journalism. The cars keep shipping. The investigations keep opening. The graphs keep going up and to the right.

What ties this week together is not any single story but the gradient. The labs are inside the government. The government is inside the labs. The bubble is too large to deflate cleanly and too central to the index to be allowed to. The surveillance infrastructure is being built out under cover of consumer convenience. The lawsuits are arriving in batches. The energy footprint is approaching the scale of nation-states. And the people running the show are spending their political donations to make sure none of this changes.

The correct posture, if you are still trying to think clearly about any of it, is to stop expecting a discrete crisis. The crisis is the distribution. Every week looks roughly like this one now. Some weeks the doom average is 3.8, some weeks it is 4.2. The slope is what matters, and the slope is up.

See you next week. Bring water.