The Latticework Monday Morning Briefing is a research-based, data-driven slide presentation sent on a separate mailing list (complimentary to members). If you do not wish to receive it, opt out here.
On Friday evening Washington reached directly into a frontier AI model and switched it off. A few hours earlier the same market that is supposed to be pricing AI risk handed a rocket-wrapped compute business the largest IPO valuation in history. Underneath both news items sits a single key question: Is the AI build-out a super cycle that compounds, or a capital cycle that mean-reverts? Chris Bloomstran of Semper Augustus and Luke Bridgeman of Hosking Partners tackle the question from different angles. They both arrive at roughly the same place.
Washington pulled Anthropic’s most powerful models off the market on Friday night. At 5:21pm Eastern on June 12, the Commerce Department, in a letter from Secretary Howard Lutnick drafted with the Bureau of Industry and Security, issued an export-control directive ordering Anthropic to suspend access to Fable 5 and the non-public Mythos 5 for any foreign national, whether outside the US or inside it, and explicitly including Anthropic’s own non-citizen employees. Because the company cannot verify citizenship at the model layer, it had no practical option but to disable both models for every customer. Its other models, including Claude Opus 4.8, are unaffected. This is the first time the US has applied export controls to an AI model itself rather than to the chips that train it. The trigger, reportedly surfaced by Amazon, was a single narrow jailbreak. Anthropic says it is narrow rather than universal, confers no meaningful incremental capability, and can be reproduced on rival models such as OpenAI’s GPT-5.5 that face no comparable order.
We share the view, argued at length by Zvi Mowshowitz, that whatever the merits of the underlying concern, the implementation was clumsy and reveals either real animus toward Anthropic or a basic misunderstanding of how jailbreaks and defense in depth actually work. The incoherence is hard to miss. The administration is simultaneously relaxing controls on selling advanced chips to China while barring, say, a British employee of a New York bank from touching the best American model. If the standard applied to Fable were applied across the industry, it would halt essentially every frontier deployment, which is why the more honest framing is not “Anthropic broke a rule” but “who is next?” The own-goal risk is the one worth holding in mind: a large share of the technical staff at every American lab are foreign nationals, and a policy that locks them out of their own tools is the rare measure that could slow the US and speed its competitors at the same time. It is, as Zvi puts it, close to the opposite of coordinating for safety. His full piece is here.
The case is not closed, and that is the point. Axios reports the pause may last only weeks, and David Sacks, speaking for the administration, frames it as a contained request: fix the specific jailbreak and the control lifts. If that is all it is, the episode resolves quickly and cheaply. If instead the demand is to guarantee that no jailbreak at this level ever recurs, that is not a thing any model provider can deliver, and the standoff hardens. What the action proves matters to anyone underwriting AI cash flows. The state has now demonstrated, with no warning and on a Friday night, that it may reach into a deployed model and turn it off. That is a new and non-trivial input to the discount rate for the entire AI complex, and it lands while three of the largest equity offerings in history are lining up to sell that complex to the public.
The first of those offerings priced and traded on Friday. SpaceX listed on the Nasdaq under the ticker SPCX, pricing more than 555 million shares at $135 and raising roughly $75 billion, the largest IPO ever completed. The stock opened at $150, traded up more than 30 percent at the intraday high (briefly worth more than $2.25 trillion), and closed at $161, up 19 percent on the day, for a first-session market capitalization above $2.1 trillion. Up to 30 percent of the deal went to retail, volume exceeded 500 million shares, and Elon Musk became, on paper, the world’s first trillionaire.
A week ago we flagged what was actually being sold, “a narrative priced as though the option has already paid off.” On Friday the market did not discount that narrative. It underwrote it in full, and at a $2.1 trillion close the implied multiple on trailing sales is richer still than the figure that already looked heroic seven days ago. We are not being offered cash flows but the assumption that the optionality has already been exercised. Last week’s note, with the full reframe, is here.
Our first featured piece this week is the full recap, audio, and 54-page deck from Chris Bloomstran’s session at The Zurich Project, held June 2 to 4. Chris is President and Chief Investment Officer of Semper Augustus and the author of what many regard as the most rigorous outside analysis of Berkshire Hathaway anywhere. His title was self-deprecating. Having called a secular peak in his 2021 letter, only to watch the recovery carry valuations above those highs, he now borrows Irving Fisher’s “permanently high plateau” and calls it a secular plateau. The numbers are the spine of the argument. The S&P 500 ended 2025 at 26 times operating earnings, matching 1929 and approaching the 29 times of early 2000, with a CAPE of 40, price to sales near 3.8, price to book of 6.2, and a dividend yield of roughly 1 percent, an all-time low. His verdict is that there is no moment in the history of the American cap-weighted market when valuations have been higher than today. Offered the choice between that index and a Treasury bill, he would own the bill, though he is careful to add that cash has a clock: his opportunity-cost table shows the market must fall about 28 percent within five years, or 48 percent within ten, merely to justify holding cash against equities compounding at 10 percent.
The analytical core is his five-factor decomposition of total return into dollar sales growth, margin change, multiple change, share-count change, and dividend yield. He illustrates it with Coca-Cola, a great business that nonetheless compounded at just 4.6 percent over the 27 years since mid-1998, when it traded at 58 times earnings and was half the Berkshire portfolio, as that multiple unwound by 60 percent to roughly 23. Apple, bought cheap at 13.9 times earnings (10 times net of cash) and compounded near 30 percent, now sits at 34.5 times, and on his three scenarios is priced to deliver somewhere between zero and 7.5 percent, which is why Berkshire has been a seller. Run the same arithmetic on the index and the Ibbotson 10.5 percent looks implausible: to earn it over the next decade, net margins must reach about 20.7 percent at a constant 26 multiple, or the multiple must reach 43 times at constant margins. He expects neither, and expects the index to trade substantially below current levels at some point within ten years.
The sharpest section applies capital-cycle history, the canals, railroads, autos, electrification, and fiber, to the AI build-out. Hyperscaler capex ran near $400 billion last year and is headed toward $750 billion or more, with cumulative 2023 to 2030 estimates now as high as $7 trillion. The depreciation math is unforgiving: $400 billion of capex on a ten-year straight line creates $40 billion of first-year depreciation against AI revenues of $40 to $50 billion, and a 20 percent return on $4 trillion of cumulative capex would require $800 billion of profit, nearly a third of all S&P 500 earnings. He layers on roughly $650 billion of off-balance-sheet financing, points to structures such as Meta’s five-gigawatt Hyperion data center carried through a Blue Owl joint venture on just $500 million of Meta equity and a Meta debt guarantee, and hears echoes of Enron and Lucent. Berkshire is his discipline case study throughout: running hard away from softening reinsurance, sitting on roughly $120 billion of holding-company cash, and making the surprising $10 billion Google purchase at 10 times sales that he reads as a signpost on Greg Abel’s capital allocation. The full session, with his data and the deck, is worth your time and is available here.
Our second featured piece, from Luke Bridgeman of Hosking Partners, asks the capital-cycle question from the supply side of the same trade: memory. The three companies that dominate global DRAM, Micron in the US and SK Hynix and Samsung in South Korea, have run up between five and eight times over twelve months and have each breached a trillion dollars of market value. Hosking has owned all three since inception, and the essay is an honest interrogation of its own thesis: did the position work for the reason they bought it? The original case was pure capital-cycle theory. As Moore’s law broke down around 2014 and transistor shrinkage stalled near 10 nanometres, capex slowed, a field of more than thirty DRAM makers consolidated to three holding 97 percent of the market, and returns recovered into higher lows and higher highs that rewarded counter-cyclical trimming and adding around 1 times book value. What has happened since ChatGPT, Hosking argues, is different in kind: a demand-led super cycle, driven by high-bandwidth memory for AI that consumes four times the wafer capacity of conventional DRAM, squeezing the conventional product and lifting valuations far above their historical anchor at book.
Their caution is the value investor’s caution, and it rhymes exactly with Bloomstran’s. They are wary of the fallacy of composition: many enterprises can pay full price for AI tokens at high returns, but not all use cases will clear that bar, so the true addressable market may be smaller than it looks. High prices always call forth new supply, and China’s CXMT is likely to IPO this year to fund exactly that, even with its technology still lagging. The survivors now demand binding long-term volume commitments and up-front payment precisely because they remember what the cycle does to them. So Hosking has been taking profits into strength. Set against the heroic assumptions embedded in the coming AI-lab listings of OpenAI, Anthropic, and SpaceX, the firm notes that DRAM still trades on 6 to 10 times forward earnings, which is the market’s way of saying it does not believe this time is different. Their closing line is the one to keep: “chips are currently not ‘cheap as chips.’” The full note is here.
In summary, the market is pricing AI optionality as already won, the government has proved it can withdraw the underlying tech overnight, and two disciplined voices are independently warning that the arithmetic does not close. Super profits invite supply. Capex demands a return. The fallacy of composition caps the addressable market. None of that says the ride ends this year, and Bloomstran is explicit that it may not. It says only what it has always said. Do not buy high.
Feedback on the Monday Morning Briefing
“Most of what I monitor, all in one place. Great value add.” —Brad Lummis
“Loving these Monday briefings!” —Jon Bartel
“Tightly presented and easy to digest. I just spent 20 minutes going through it, and it’s helped to level set me for the week ahead.” —Michael Loftis
“A great piece and thoughtfully assembled.” —Brian Wolf
“I don’t think I have ever seen more valuable content in one place.” —Bill Coleman
“Worth its weight in gold.” —Shree Viswanathan
The new issue follows.
